Merge pull request #2007 from debpalash/release/v0.5.2

docs(changelog): cut the 0.5.2 release section

.github/CONTRIBUTING.md

@@ -85,14 +85,20 @@ names: there is no `desktop=prod` (note the **hyphen** in `desktop-prod`).
 
 Requires [Rust](https://rustup.rs/) and platform-specific Tauri dependencies — see the [Tauri prerequisites](https://v2.tauri.app/start/prerequisites/).
 
-After installing Rust with rustup on macOS/Linux, either open a new terminal or
-load Cargo into the current one before starting the desktop app:
+After installing Rust with rustup (or `uv` with its installer), a terminal that
+was already open still has the old `PATH`. The desktop launchers (`bun desktop`,
+`bun desktop-prod`, `bun desktop-fresh`) detect this and add `~/.cargo/bin` /
+`~/.local/bin` for that run, printing a one-line note; to make it permanent,
+open a new terminal, or on macOS/Linux load Cargo into the current one:
 
 ```bash
 source "$HOME/.cargo/env"
 bun desktop
 ```
 
+If Rust is genuinely not installed, the launchers stop up front with the
+install command instead of failing later inside `cargo metadata`.
+
 On Linux, errors such as `Package gdk-3.0 was not found`, `pango.pc` missing,
 or `javascriptcoregtk-4.1` missing mean the native packages above were not
 installed; changing `PKG_CONFIG_PATH` does not fix libraries that are absent.

.github/workflows/ci.yml

@@ -11,6 +11,11 @@ on:
   push:
     branches: [main]
   workflow_dispatch:
+    inputs:
+      windows_wix_diagnostic:
+        description: Run only the tiny nonpublishing Windows MSI authoring diagnostic
+        type: boolean
+        default: false
 
 permissions:
   contents: read
@@ -21,6 +26,7 @@ env:
 
 jobs:
   test:
+    if: ${{ !inputs.windows_wix_diagnostic }}
     name: Tests (backend + frontend)
     runs-on: ubuntu-22.04
     env:
@@ -189,6 +195,7 @@ jobs:
   # and `cargo test --lib` runs the shell's unit tests natively on each OS.
   # Full bundling stays in release.yml on tag push.
   tauri-cross-platform:
+    if: ${{ !inputs.windows_wix_diagnostic }}
     name: Tauri shell check (${{ matrix.label }})
     needs: test
     strategy:
@@ -289,6 +296,7 @@ jobs:
   # job above misses. Narrow scope (tests/smoke/ only) — full pytest stays
   # on Linux until Phase 1's INST-01 lands setuptools for WhisperX.
   smoke-matrix:
+    if: ${{ !inputs.windows_wix_diagnostic }}
     name: Smoke (${{ matrix.label }})
     needs: test
     strategy:
@@ -428,17 +436,68 @@ jobs:
           PY
 
       - name: Run smoke tests
+        # Exercise credential paths on native Windows as well as POSIX hosts.
         if: matrix.backend_supported
-        run: uv run --no-sync pytest tests/smoke/ -q --tb=short
+        run: uv run --no-sync pytest tests/smoke/ tests/test_hf_token_cache_paths.py -q --tb=short
         env:
           HF_HUB_OFFLINE: "1"  # same no-silent-downloads guard as the main pytest job
           HF_HUB_CACHE: ${{ runner.temp }}/pockettts-empty-hf-cache
 
+      # The isolated backend session, on Windows. The `test` job runs it on
+      # Linux only, which is how four tests that CANNOT pass on Windows shipped
+      # unnoticed: two reach for os.WNOHANG and os.waitid (POSIX-only, an
+      # AttributeError before the first assertion), one asserts a RuntimeError
+      # that `backend_drain_fd` returns None instead of raising off POSIX, and
+      # one raced the OS reaping a crashed child — a race Linux won and Windows
+      # lost every time. All four were invisible to CI and hit every Windows
+      # contributor on their first `pytest` run. Forty seconds closes the class.
+      - name: Isolated backend session (Windows)
+        if: runner.os == 'Windows' && matrix.backend_supported
+        run: uv run --no-sync pytest backend/tests/ -q --tb=short
+        env:
+          HF_HUB_OFFLINE: "1"
+
       # Artifact commits depend on native Windows rename/replace semantics;
       # Linux emulation cannot exercise sharing rules or path parsing.
+      # test_worker_task_store and test_worker_inbound_transport joined this
+      # step after a Windows run found a real portability bug the Linux-only
+      # `test` job could not see: a staged input's artifact id was built with
+      # os.path.join, so a Windows control plane persisted and shipped
+      # `inputs\<sha>.wav` — which a Linux worker cannot resolve. These suites
+      # need no ffmpeg, so they cost seconds here.
       - name: Remote-worker artifact paths (Windows)
         if: runner.os == 'Windows' && matrix.backend_supported
-        run: uv run --no-sync pytest tests/test_worker_upload_server.py tests/test_worker_server_integrity.py -q --tb=short
+        run: >-
+          uv run --no-sync pytest
+          tests/test_worker_upload_server.py
+          tests/test_worker_server_integrity.py
+          tests/test_worker_task_store.py
+          tests/test_worker_inbound_transport.py
+          -q --tb=short
         env:
           HF_HUB_OFFLINE: "1"
           HF_HUB_CACHE: ${{ runner.temp }}/worker-artifact-empty-hf-cache
+
+  windows-wix-diagnostic:
+    name: Windows MSI authoring (no publishing)
+    needs: test
+    if: ${{ !cancelled() && (inputs.windows_wix_diagnostic || needs.test.result == 'success') }}
+    runs-on: windows-2022
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+      - uses: oven-sh/setup-bun@v1
+      - name: Bundle canonical system and per-user templates with a tiny payload
+        shell: pwsh
+        run: ./scripts/diagnose-windows-wix.ps1
+      - name: Restore hosted Installer policy after failed standard-user installation
+        shell: powershell
+        run: ./scripts/test-msi-policy-cleanup.ps1
+      - name: Preserve verbose linker output and rendered authoring
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: windows-wix-diagnostic
+          path: wix-diagnostic-artifacts/
+          if-no-files-found: warn
+          retention-days: 3

.github/workflows/release.yml

@@ -604,6 +604,21 @@ jobs:
             fi
           done < /tmp/stale.txt
 
+      # A retried job reuses its version and can collide with installers it
+      # uploaded before a later step failed. Keep other versions/arches intact;
+      # macOS versionless updater archives are scoped by release tag and arch.
+      - name: Clear this target's installer assets on retry
+        if: github.run_attempt > 1
+        shell: bash
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          RELEASE_TAG: ${{ (needs.preview-gate.outputs.is_preview == 'true') && 'preview' || github.ref_name }}
+          RELEASE_TARGET: ${{ matrix.rust_target }}
+        run: |
+          VERSION=$(python -c 'import json; print(json.load(open("frontend/package.json"))["version"])')
+          python scripts/clear-release-rerun-assets.py \
+            --tag "$RELEASE_TAG" --version "$VERSION" --target "$RELEASE_TARGET"
+
       - name: Build + release (Tauri)
         uses: tauri-apps/tauri-action@v0
         env:
@@ -780,7 +795,7 @@ jobs:
           set -euo pipefail
           MSI=$(find frontend/src-tauri/target/${{ matrix.rust_target }}/release/bundle/msi -name '*Current*User*.msi' | head -1)
           powershell.exe -NoProfile -ExecutionPolicy Bypass \
-            -File scripts/smoke-per-user-msi.ps1 -MsiPath "$(cygpath -w "$MSI")"
+            -File scripts/smoke-per-user-msi.ps1 -MsiPath "$(cygpath -w "$MSI")" -PrepareHostedRunner
 
       # linuxdeploy re-links .DirIcon as an ABSOLUTE symlink into the build
       # machine AFTER tauri's files-map has placed the real icon bytes — the

.gitignore

@@ -170,3 +170,13 @@ bin/omnivoice-tts-linux-aarch64
 # committed (they ship with the app); the per-language source WAVs are just the
 # inputs scripts/render_dub_demo_audio.py hands to scripts/build_dub_demo.sh.
 backend/assets/samples/demo/dubbing/*.src.wav
+
+# Stray sqlite session artifacts (`<db-path>.ses`). An in-memory DB yields the
+# literal name `:memory:.ses`, and a path containing `:` cannot be checked out
+# on Windows at all — committing one fails every Windows CI job at the git
+# checkout step, before a single test runs. Guarded by
+# tests/test_no_windows_hostile_paths.py.
+*.ses
+
+# Generated Windows MSI diagnostic logs and installer payloads
+/wix-diagnostic-artifacts/

.gitleaks.toml

@@ -25,6 +25,8 @@ regexes = [
   '''^hf_QWERTYUIOPasdfghjklZXCVBNM0123456789xyzAB$''',
   # NLLB generation length argument, not the value of a credential.
   '''^max_length=400$''',
+  # Dubbing pane split-position localStorage key, not a credential.
+  '''^omnivoice\.dubSplit\.v1$''',
   # cryptography's Ed25519 private-key type name, not key material.
   '''^Ed25519PrivateKey$''',
 ]

AGENTS.md

@@ -33,6 +33,11 @@ Binding for every AI agent (Claude, Codex, Cursor, review bots, …). CLAUDE.md
 - `frontend/package.json` dep changes require regenerating root `bun.lock` (Docker runs `--frozen-lockfile`).
 - Issues: absorb or decline — never defer to a future version. Check the open-PR queue before implementing community-reported fixes.
 
+## Shared select controls
+
+- Use `frontend/src/components/SearchableSelect.jsx` for all new or redesigned select boxes. Reuse `VoiceSelector` for voice choices. Do not introduce native `<select>` controls.
+- Provide a localized `ariaLabel`; use `menuPortal` inside scrolling or clipping containers. Preserve keyboard selection and disabled states.
+
 ## Agent skills
 
 Project development skills are pinned in `skills-lock.json` and installed under

CHANGELOG.md

@@ -8,25 +8,58 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ## [Unreleased]
 
+## [0.5.2] — 2026-09-10
+
 **Highlights**
 
-- Fix current-user Windows installer validation and nested resource cleanup (#730)
-
+- Supertonic-3 and PocketTTS show their license Accept button again, so they can be enabled (#2017)
+- An engine that can't run on your platform says so, instead of telling you to install it (#2018)
+- MOSS-TTS-v1.5, Confucius4-TTS, dots.tts, Supertonic-3 and PocketTTS install in one click, each in its own environment, so switching engines and back never breaks a working one (#2015, #2016)
+- A pronunciation entry that is stored but not applied yet says so, instead of looking like it did not match (#1949)
+- A bare 500 report now names the backend error class, so two unrelated faults stop filing the same issue (#1773)
+- A rejected dubbing source language now names the code it rejected (#1960)
+- The first-run install log is kept on disk instead of vanishing with the setup screen (#1847)
+- `bun run desktop` reclaims port 3900 from a backend the app itself left running, instead of refusing to start (#1974)
+- A dictation shortcut another app already owns now says so, instead of silently doing nothing (#1858)
+- Quitting on Windows is no longer reported as a crash on the next launch (#1898)
+- A Reduce motion switch in Settings, for calm without changing your whole system (#1857)
+- A light theme, and System Auto now follows a light-mode OS instead of staying dark (#1973) — thanks @CoDe-ReDz!
+- Generating from a one-character input now says the input was too short, instead of quoting a convolution error (#1826)
+- First run asks about text size before the install, not after it (#1849)
+- Cloning without a reference clip now says so, instead of naming library parameters you cannot set (#1879)
+- Upgrading torch for an RTX 50-series card no longer trades one startup crash for another, and the upgrade is documented (#1931)
+- A generation timeout now points at the compute-time budget in Settings rather than an environment variable (#1808)
+- An engine you have not installed now says so, instead of reporting a failed check (#1866)
+- The Accessibility prompt no longer floats over first-run setup and every other app until you grant it (#1845, #1886)
+- The last onboarding step offers to install a speech-to-text model instead of failing three times when none is installed (#1856)
+- A download that fails because the folder sits behind a mount point Windows will not cross now says so, and where to move it (#1957)
+- A GPU that is merely short on free memory is no longer told to reinstall its drivers (#1812) — thanks @michaelhuamanflores!
+- An error thrown by a browser extension is filtered on Safari and the macOS app too, not only on Chromium (#1901) — thanks @Chang-Jin-Lee!
+- Choosing the China mirror no longer re-races the network on every dependency step, which cost seconds per step on blocked connections (#1892) — thanks @yuezheng2006!
+- The backend log panel reports a log it cannot read instead of quietly showing less (#1847) — thanks @Chang-Jin-Lee!
+- The floating dictation bubble adds pause, resume, stop, close, and a multiline preview (#1952)
+- Transcriptions checks model readiness and offers an inline download and shortcut hints (#1952)
+- Transcriptions' missing-model prompt lists every dictation model by accuracy vs latency, languages and size, so you install the one that fits — or switch to one already on disk (#1952)
+- The Engines menu's Transcription tab picks the dictation model under Sherpa-ONNX, and that choice now also drives Sherpa transcription (#1952)
+- A failure with no stage attached no longer borrows another stage's advice, so a text-to-speech error stops telling you the video server dropped the download (#1943)
+- A generation failure that the app cannot classify now names the backend error class, so two unrelated faults stop arriving as the same untriageable report (#1800)
+- Transcriptions dictation wakes the desktop recorder, presents one contextual start action, and centers its microphone icon with the label (#1902)
+- Colab transcription and dubbing now include an explicit ASR model setup step (#1922) — thanks @nidhi-singh02!
+- Apple Silicon now shows one canonical OmniVoice choice in the engine picker while retaining its automatic crash-isolated sidecar runtime (#1913)
+- Validate current-user Windows installers under a standard account on hosted runners (#1883)
+- Model downloads survive a flaky connection instead of restarting from zero (#1940)
+- `bun run dev` recovers on Windows instead of demanding Task Manager (#1941)
+- The desktop app builds and opens from a fresh clone again (#1818) — thanks @flutterkage2k!
+- GPUs with less VRAM than the engine needs no longer get half the compute-time budget a CPU gets (#1806) — thanks @VishvakR!
+- Gallery voice previews play again — the quality guard was rejecting good renders as silent (#1819) — thanks @flutterkage2k!
+- Tilde-separated number ranges are spoken clearly without running their endpoints together (#1821) — thanks @flutterkage2k!
+- Voice modes use themed tabs, with Synthesize and Convert pinned below their scrolling forms (#1823)
+- Fix current-user Windows installer validation and nested resource cleanup (#1873)
+- Keep generated frontend assets available while building the current-user Windows installer (#1881)
 - Voice cloning now starts with a clear upload-or-record choice, reveals recording and reference details only when needed, and keeps sampling controls under Production Overrides (#1817)
-
-### Changed
-
-### Added
-
-### Docs
-
-### Fixed
-
-
-## [0.5.2] — 2026-09-02
-
-**Highlights**
-
+- The first-run welcome line uses an instruction accepted by OmniVoice and VoiceDesign engines (#1861) — thanks @psiberfunk!
+- audio.cpp joins the engine lineup as an opt-in CPU backend for Breeze-TTS-2 (English + Chinese, clone + voice design, explicit Model Catalogue install, no Python venv) (#1891)
+- audio.cpp uses installed native CUDA, HIP, Metal, and Vulkan providers and preserves device routing across remote workers (#1926)
 - Show estimated and measured model, dependency, cache, and temporary disk costs in the engine catalogue (#1718)
 - Preview builds now stay newer than Stable even when automatic post-release version bumps are disabled (#1762)
 - CosyVoice setup guidance now separates downloaded model files from the runtime that makes the engine available (#1761)
@@ -40,6 +73,27 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ### Changed
 
+- Tauri 2.11.5 with refreshed plugins (dialog, updater, log, opener, positioner, single-instance), React 19.3, TanStack Query 5.102, lucide 1.43, posthog-js 1.428, and the rest of the npm workspace on current minors; jsdom 30, jest-dom 7, concurrently 10, taze 21 (#1952)
+- eslint ignores `src-tauri/`, so a local Tauri build no longer floods `lint:hooks` with parse errors from generated assets (#1952)
+- Casting uses responsive SVG voice cards and searchable speaker menus that stay above surrounding panels (#1823)
+- Dubbing aligns output settings, brings review status forward, and simplifies transcript and glossary editing; Launchpad files and voices reflow into responsive grids (#1823)
+- Transcript segments use three readable rows for text, timing/status and voice controls, with heights that adapt to wrapping (#1823)
+- Dragging the waveform pans horizontally while a click still seeks, keeping the timed transcript aligned (#1823)
+- Bulk segment editing uses searchable voice and language menus, readable language names and a responsive selection toolbar (#1823)
+- Dubbing overlays playback controls on video, combines waveform and transcript in a compact timeline, and removes header/action background fills (#1823)
+- Dubbing uses compact casting, translation and output controls with responsive rows to leave more room for editing (#1823)
+- Export uses grouped format settings, themed track menus and switches, with a pinned filename summary and download action (#1823)
+- Dubbing output settings use icon-labelled switches, themed track and speaker menus, and clearer timing/transcript controls (#1823)
+- Casting voice menus use searchable themed options with SVG preset icons instead of native dropdowns (#1823)
+- Dubbing groups casting and translation controls with readable labels, SVG icons, searchable menus, and compact timeline spacing (#1823)
+- Production Overrides use readable icon-labelled controls and accessible Denoise/Postprocess switches (#1823)
+- Expanded navigation uses a theme-accent tint with subtle static wave gradients (#1823)
+- Convert groups source audio, target voice, and timing options into clearer controls; design choices include theme-matched SVG icons (#1823)
+- The expandable sidebar reveals workspace labels with restrained active states; language menus adapt to multiple columns on wider screens (#1823)
+- Voice design and recording use themed, keyboard-accessible selectors with clearer spacing and labels (#1823)
+- Voice tabs and upload/record controls have subtle SVG motion; Text adds clipboard paste and the upload area fills available height (#1823)
+- The title-bar label cycles through active speech, transcription, and LLM engines; bundled model labels correctly say OmniVoice (#1823)
+- The top-bar Engines panel groups Speech, Transcription, and LLM choices into tabs, with compact memory controls and no duplicate pickers (#1823)
 - Voice Design simplified: the 12-row fine-grained block collapses to one summary line with a five-field editor, English accent and Chinese dialect merge into a single field, and the starting-point chips now show 5 with an overflow toggle (#1793)
 
 ### Added
@@ -55,6 +109,11 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ### Docs
 
+- PowerShell Docker setup now generates the administrator key without requiring Python on the host (#1993) — thanks @yangfan-yf-yf!
+- The torch upgrade an RTX 50-series card needs is written down, with the second pin file the resolver checks and the command that proves the kernels are there (#1931)
+- Docker quick starts now explain the AMD64-only images and direct Apple Silicon users to the native macOS app (#1921) — thanks @yangfan-yf-yf!
+- audio.cpp (Breeze-TTS-2) is now a documented opt-in engine: prebuilt binary install, explicit GGUF download, voice modes, and the weights' research/non-commercial terms (#1891)
+- `docs/STRUCTURE.md` describes the tree as it is today, and a test now keeps its counts honest (#1981) — thanks @Dawcraft!
 - Local gigastt is now documented as a supported OpenAI-compatible ASR endpoint, with loopback privacy distinguished from remote servers (#1736) — thanks @ekhodzitsky!
 - The CosyVoice guide now states that packaged builds have no one-click runtime installer and records the exact readiness checks exposed by [Discussion 1631](https://github.com/debpalash/VoiceStudio/discussions/1631) (#1761)
 - A production private-API guide now covers pinned containers, root credentials, network isolation, streaming proxies, health checks, upgrades, and benchmark evidence (#1720)
@@ -62,6 +121,82 @@ the frozen-backend fallback mirror it for their toolchains.
 
 ### Fixed
 
+- One-click engine installs no longer inherit VoiceStudio's own PyTorch pin, which made MOSS-TTS-v1.5 and Confucius4 impossible to install (#2024)
+- Uninstalling a translation engine no longer removes a package VoiceStudio or another engine still needs (#2019)
+- Closing the dictation pill on Windows removes it from the screen: an empty dark rectangle used to stay there, always on top, until the app was quit (#2009)
+- The dictation pill on Windows no longer sits inside a bordered card wider than the pill itself (#2009)
+- Dictation uses the model you picked instead of one remembered from before the backend started, so it stops reporting no speech-to-text model while one is installed — and when none is, the main window offers the download (#2012)
+- The remote-worker loop-responsiveness tests no longer turn a build red over milliseconds of scheduling noise on shared CI hardware (#1990)
+- Remote GPU workers work when the machine running VoiceStudio is on Windows: a staged input is now identified the same way on every operating system, instead of with a path only Windows can read (#2005)
+- The pronunciation list badges an IPA or CMU entry as not applied yet, so you can see it without running a test (#1949) — thanks @utkarsha741!
+- A remote-worker test no longer fails at random on Windows CI: it waited for a background thread by spinning the event loop that thread's work needed (#1990)
+- The isolated backend test session passes on a stock Windows checkout, and CI now runs it there so it stays that way (#1990)
+- Windows contributors can run the test suite without Developer Mode: tests that create a symlink now skip instead of failing with `WinError 1314` (#1990)
+- The crash details dialog now says what the exit code means and what to try, instead of showing a raw number and a log (#1927)
+- A crash report now carries the backend's actual last words: the log tail is captured after the dying process's final output lands, not the instant it exits (#1850)
+- The first-run setup screen no longer mislabels a step when the bootstrap restarts itself: Rust now says which attempt each stage and log line belongs to, instead of the screen guessing from a once-a-second poll (#1900)
+- A port-3900 conflict now names who is actually holding it, and gives the command that ends an orphaned backend, instead of telling you to quit an app that has no window (#1933) — thanks @Chang-Jin-Lee!
+- Windows desktop launches no longer freeze at "Loading ML runtime (PyTorch)": the parent-liveness watchdog polls the stdin pipe instead of leaving a read pending, which deadlocked numpy's OpenBLAS initializer (#1952, #1955)
+- `bun desktop-prod` and `bun desktop-fresh` find Rust and uv from a terminal opened before they were installed, as `bun desktop` already did; a missing Rust toolchain fails up front with the install steps (#1952)
+- Voice synthesis progress no longer races to a fabricated 95%; it stays indeterminate until the active generation path reports real progress (#1907) — thanks @psiberfunk!
+- The Backend log tab keeps showing history across a log rollover, instead of going nearly empty until new lines arrive (#1920)
+- Clearing the logs now empties the rotated log files too, so it frees the space it appears to (#1920)
+- An error thrown by a browser extension no longer offers to file itself as a VoiceStudio bug (#1901)
+- Clearing the desktop logs no longer wipes the backend's stderr, which is the only record a native crash leaves behind and is meant to survive a respawn (#1510)
+- Long audiobook chapters now use the same device- and text-length-aware synthesis timeout as other TTS routes (#1910) — thanks @psiberfunk!
+- Interrupted audiobook renders can resume cached chapters after tab navigation, and their chapter cache is available from the recovery card (#1911) — thanks @psiberfunk!
+- System-check details and storage paths beginning with a number or a slash no longer render with their leading text moved to the end of the line (#1848) — thanks @psiberfunk!
+- An unavailable engine's row now links to that engine's guide, so the generic "check installation and configuration" message has somewhere to send you (#1866) — thanks @psiberfunk!
+- The backend log now records which engine failed a health check and whether its probe raised, instead of a line that identified neither (#1866) — thanks @psiberfunk!
+- The first-run Activity log counts every line instead of freezing at 200 while the install is still running, and Copy now hands back the whole run rather than the last 200 lines (#1847) — thanks @psiberfunk!
+- A first-run failure that happened early in a long install keeps its specific advice, instead of falling back to the generic retry hint once the log scrolled past 200 lines (#1847) — thanks @psiberfunk!
+- Opening the log panel no longer clips the Launchpad's heading and slides the feature cards up over it — the page scrolls instead of squashing itself (#1859) — thanks @psiberfunk!
+- Segmented model downloads split files into 16 MB ranges instead of one range per connection, so a dropped connection refetches one range rather than restarting the file (#1940)
+- The download accelerator is kept across retries after a transient network failure and resumes from its manifest, instead of falling back to a from-zero `snapshot_download` (#1940)
+- `dev-backend.mjs` stops the backend by process tree on Windows, so an orphaned uvicorn no longer holds port 3900 and turns a source reload into three phantom crashes (#1941)
+- `clear-dev-ports.mjs` can free a stuck development port on Windows again, bound to the inspected process instance so a recycled pid is never terminated (#1941)
+- Checkout-ownership matching no longer resolves POSIX paths with the host's separator, which made the guard's own test fail on Windows (#1941)
+- Install documentation help now prints correctly on Windows consoles using legacy encodings (#1815) — thanks @dajiaohuang!
+- Saved transcriptions with missing or invalid timestamps now remain readable (#1799) — thanks @yunaremaia and @tvbht!
+- Transcribing with an engine that reports no segment end no longer fails with a server error; the null timing is passed through the way the segment list already expects (#1904) — thanks @aeroglu!
+- Copying a saved transcription now uses the shared clipboard helper and reports failed copies accurately (#1803) — thanks @tvbht!
+- Voice reference preparation reclaims allocator memory before one bounded retry, then reports persistent GPU out-of-memory failures (#1811)
+- `bun run desktop` now opens on a fresh clone: the Vite alias for `@tauri-apps/plugin-dialog` no longer assumes a nested `frontend/node_modules`, which bun's workspace hoisting leaves empty (#1818) — thanks @flutterkage2k!
+- Slow backend startups remain running with progress updates, and Retry interrupts startup without stale timeout failures (#1809)
+- Backend connection errors report crashes only when recorded evidence exists, and diagnostic waits honor cancellation (#1810)
+- A CUDA or ROCm GPU with less VRAM than the engine needs now gets the CPU compute-time budget instead of the shorter accelerated one, since it pages to system RAM and renders slower than the CPU would — applied to local generation, voice conversion, and remote worker deadlines alike (#1806) — thanks @VishvakR!
+- Gallery previews no longer fail with "the voice engine returned no audible audio" on perfectly good renders: the degenerate-buzz guard measured spectral flatness over the whole clip (so the value tracked clip length) against a threshold calibrated on a synthetic signal, and rejected real speech in every language tested (#1819) — thanks @flutterkage2k!
+- Speak tilde separators in integer, signed, and decimal ranges in English, Korean, Japanese, and Chinese (#1821) — thanks @flutterkage2k!
+- Keep recording and conversion work safe while switching methods, synchronize dubbing language controls, and localize timeline controls and timing warnings (#1841)
+- Audiobook is now a Write → Cast → Produce tab workspace matching the voice workspace, with the warnings/progress/result rail pinned below (#1841)
+- Gallery uses a workspace header with zone tabs, hairline section dividers, theme-token cards, and borderless import rows (#1841)
+- Gallery cards reset native button faces, cluster icon actions in the header so Use voice never wraps, and use a roomier grid floor (#1841)
+- Gallery filters gain name search, removable iconified pills with clear-all, and dimension icons on every facet (#1841)
+- Dubbing playback starts before waveform decoding, automatic cast names are readable, and transcript timestamps have more room (#1823)
+- The title-bar engine button stays compact and stable while cycling labels, with engine names aligned right (#1823)
+- Long dubbing segment errors wrap in a bounded scrollable notice instead of widening the editor (#1823)
+- Voice dropdowns match their field width, use theme accents, and show recent voices only once (#1823)
+- Language menus no longer show a pale frame around their search header (#1823)
+- The notification count stays inside the title bar instead of clipping above the bell (#1823)
+- The workspace engine menu opens beside its button instead of at the opposite edge of the page (#1823)
+- Cloning reuses the dubbing language picker with flags, search, and single selection, opening above the pinned synthesis controls (#1823)
+- The first-run welcome line uses an instruction accepted by OmniVoice and VoiceDesign engines (#1861) — thanks @psiberfunk!
+- The header status dot now honors OS Reduce Motion instead of pulsing regardless (#1862) — thanks @psiberfunk!
+- Onboarding reads Hugging Face tokens locally, preserves Windows CLI logins, and requires successful discovery before replacing saved credentials (#1852) — thanks @psiberfunk!
+- The logs panel no longer reports “All clear” before log retrieval succeeds or while logs contain warnings or errors (#1870) — thanks @motodriver!
+- MOSS accelerator routing and status match runtime selection, with CPU fallback when device probing fails (#1830) — thanks @li-lizhe!
+- Confucius accelerator routing tolerates failed device probes, and dots.tts keeps safe default precision on non-CUDA hosts (#1831) — thanks @li-lizhe!
+- On macOS, the header status dot and kicker no longer render underneath the overlaid traffic lights (#1863) — thanks @psiberfunk!
+- The capture widget can hide after recording and recover from being left visible while idle (#1865) — thanks @psiberfunk!
+- macOS retains the shared desktop window sizing, resize limits, and file-drop behavior when native chrome is applied (#1865) — thanks @psiberfunk!
+- On macOS, the header no longer shows Windows-style minimize/maximize/close buttons alongside the native traffic lights (#1865) — thanks @psiberfunk!
+- Release retries replace their own partially uploaded installers without colliding with existing assets (#1871)
+- Timed-out voice engines finish process cleanup before retrying, and old timeout callbacks cannot kill replacement engines (#1872)
+- Fast macOS process exits no longer turn a completed shutdown into a permission error (#1809)
+- The bootstrap splash no longer shows fabricated first-run install steps on a warm start or repair sync — a step now renders done only once it was actually observed (#1894)
+- A deliberate, clean quit killed by the desktop shell's short shutdown grace no longer gets reported as a crash on next launch — the run sentinel now clears before the slower shutdown steps instead of after (#1895)
+- Model Catalogue engine rows stack into one column on narrow shells instead of clipping actions off-screen (#1891)
+- Simplified Chinese locale completed: all 486 missing keys translated and the parity ratchet tightened to zero (#1877) — thanks @yearth!
 - The generation compute-time budget is now a Settings control (Performance & Device) instead of an env-var-only setting the timeout error recommended with no UI path — the error copy points there too, and long CPU/MPS renders get an upfront heads-up before they start (#1787)
 - Windows: the backend can now start when the install path contains non-English characters (e.g. a CJK username) on a non-UTF-8 system code page — a new or broken Python environment now builds at an ASCII-safe path automatically (a healthy existing one is never relocated), and a specific error message names the cause and a working fix if the interpreter still crashes in `site` (#1783)
 - Exports and other native-picker actions no longer 403 with "Invalid or expired desktop authorization" when the desktop app and backend resolve different data directories, e.g. dev mode or a custom data folder (#1781)

README.md

@@ -57,6 +57,16 @@
 | **Storage** | Voices, projects, settings, and outputs stay on the machine by default |
 | **License** | AGPL-3.0 application; downloaded models keep their upstream terms |
 
+The Voice workspace starts with three tabs: **From audio** for cloning, **By design** for creating a voice, and **Convert** for speech-to-speech conversion. Each tab displays its own workflow, with Synthesize Audio or Convert pinned below the scrolling form. The top-bar **Engines** panel combines engine selection, loaded models, and unload/flush controls; <kbd>Ctrl</kbd>/<kbd>Cmd</kbd>+<kbd>E</kbd> opens it. The searchable language picker shares Dubbing’s flags and language list layout, selects one output language, and retains Auto and the full cloning catalogue. Language options flow into multiple columns when space allows. Expand **Workspaces** in the sidebar to reveal navigation labels; Escape collapses it.
+
+Dubbing starts with file upload or URL import and nearby language choices. Its **Projects** panel lists previous dubs so they can be reopened by clicking anywhere on a card; action buttons operate independently. Advanced import options include captions and optional YouTube sign-in. Dubbing places playback controls over the video with background blur and combines the waveform and timed transcript in one compact editing surface. Drag the zoomed waveform left or right to pan; click to seek. Translation language and ISO-code controls stay synchronized; Auto clears any previous language code and dialect. Transcript items group editable text, timing and status, and voice controls into three readable rows that wrap with the panel width. Output Options stays compact with the active settings shown in its summary; expand it to change output, timing, or voice matching. Transcript, glossary, and paste controls share a toolbar above the segment editor. Project details, workflow steps, and Generate/Verify/Export actions use an unfilled header.
+
+The Audiobook Script editor fills the available workspace beneath its markup toolbar; Voices and Book settings stay in their own tabs.
+
+Output settings use aligned rows; review status appears before the collapsible transcript and glossary. Glossary terms have labelled entry fields and an explicit edit action. Launchpad arranges recent files and saved voices side by side when space allows, with responsive card grids and visible Open actions.
+
+The casting board shows icon-based voice cards and searchable selectors for each speaker. Drag a card onto a speaker or choose a voice from that speaker’s menu.
+
 <a id="install"></a>
 
 ## Install
@@ -68,7 +78,7 @@ Download a package from the [latest release](https://github.com/debpalash/VoiceS
 | macOS 13.3+ | Apple Silicon DMG | [Install on macOS](docs/install/macos.md) |
 | Windows 10/11 | x64 MSI; choose the current-user build when listed to install without admin access | [Install on Windows](docs/install/windows.md#install-pre-built-msi) |
 | Linux | AppImage, x86_64 with glibc 2.39+ | [Install on Linux](docs/install/linux.md) |
-| Docker | CUDA, ROCm, CPU, and worker-only GPU profiles | [Run with Docker](docs/install/docker.md) |
+| Docker | Linux/AMD64 images; CUDA, ROCm, CPU, and worker-only GPU profiles | [Run with Docker](docs/install/docker.md) |
 
 First launch creates a managed Python environment and downloads the default model. Later launches reuse both.
 
@@ -77,6 +87,11 @@ First launch creates a managed Python environment and downloads the default mode
 
 ### Quick Docker run
 
+The published images are **`linux/amd64` only**. On Apple Silicon, use the
+[native macOS app](docs/install/macos.md) for GPU acceleration. ARM64 hosts
+should read the [architecture requirements](docs/install/docker.md#architecture)
+before pulling an image.
+
 ```bash
 docker run -d -p 127.0.0.1:3900:3900 -v omnivoice-data:/app/omnivoice_data --name voicestudio palashdeb/omnivoice-studio:stable
 ```
@@ -129,7 +144,7 @@ The desktop launcher configures Python dependencies on first run via `uv` automa
 |---|---|
 | **Voice Cloning** | Zero-shot synthesis from a short reference clip ([guide](docs/engines/README.md)) |
 | **Voice Design** | Create a voice from age, accent, pitch, style, and delivery instructions ([expressive speech](docs/expressive-speech.md)) |
-| **Video Dubbing** | Transcribe, translate, preserve speakers, synthesize, and export video ([export guide](docs/dubbing/export.md)) |
+| **Video Dubbing** | Transcribe, translate, preserve speakers, synthesize, and export video; compact translation settings include track selection, and completed dubs flag timing issues for review ([export guide](docs/dubbing/export.md)) |
 | **Stories and audiobooks** | Multi-voice scripts · EPUB/PDF import · chapter rendering · `.m4b` export |
 | **[Dictation Widget](docs/features/dictation.md)** | System-wide shortcut, live transcription, optional local-LLM cleanup |
 | **Vocal Isolation** | Demucs speech/background separation |
@@ -221,7 +236,7 @@ Engine support is capability-specific. Check cloning, language, platform, memory
 | [**Sherpa-ONNX**](docs/engines/sherpa-onnx.md) | 20+ | No | No | CUDA/CPU | CPU | CUDA/CPU | Apache-2.0 |
 | [**IndexTTS 2.5** ⚡](docs/engines/indextts.md) | ZH · EN · JA · ES · AR | Yes | No | CUDA/CPU | CPU | CUDA/CPU | Bilibili model license¹ |
 | [**OmniVoice GGUF** ⚡](docs/engines/omnivoice-gguf.md) | 600+ | Yes | Yes | CUDA/CPU | MPS/CPU | CUDA/CPU | [AGPL-3.0](LICENSE) app · [review the derivative model terms](https://huggingface.co/Serveurperso/OmniVoice-GGUF#license)³ |
-| [**OmniVoice (subprocess)** ⚡](docs/engines/omnivoice-subprocess.md) | 600+ | Yes | Yes | CUDA/CPU | MPS | CUDA/CPU | [AGPL-3.0](LICENSE) app · [Apache-2.0 code, CC-BY-NC weights](https://huggingface.co/k2-fsa/OmniVoice#license)³ |
+| [**OmniVoice (subprocess; opt-in off MPS)** ⚡](docs/engines/omnivoice-subprocess.md) | 600+ | Yes | Yes | CUDA/CPU | MPS via default OmniVoice | CUDA/CPU | [AGPL-3.0](LICENSE) app · [Apache-2.0 code, CC-BY-NC weights](https://huggingface.co/k2-fsa/OmniVoice#license)³ |
 | [**PocketTTS** ⚡](docs/engines/pockettts.md) | EN · FR · DE · PT · IT · ES | Yes | No | CPU | CPU | CPU | CC-BY-4.0, gated² |
 | [**Supertonic 3** ⚡](docs/engines/supertonic3.md) | 31 | No | No | CPU | CPU | CPU | OpenRAIL-M |
 | [**MOSS-TTS-v1.5** ⚡](docs/engines/moss-tts-v15.md) | 31 | Yes | No | CUDA/CPU | CPU | CUDA/CPU | Apache-2.0 |

backend/api/public_engine_metadata.py

@@ -32,17 +32,129 @@ def _public_routing_reason(status: object, diagnostic: object) -> str:
     return _ROUTING_BY_STATUS.get(status, _ROUTING_UNAVAILABLE)
 
 
+# Categories for WHY an engine is unavailable. The probe's own sentence cannot
+# cross the boundary — it carries exception text, local paths and sometimes
+# credentials — but "Engine unavailable. Check installation and configuration."
+# told the user nothing at all, and "Last error: A previous engine check
+# failed." reads like a crash rather than "you have not installed this yet"
+# (#1866). Classifying the private diagnostic into an owned sentence keeps the
+# boundary intact and still names the kind of problem and the place to fix it.
+_UNAVAILABLE_NOT_INSTALLED = (
+    "This engine's package isn't installed yet. Install it from "
+    "Model Catalogue → Engines."
+)
+# An engine gated behind an in-app license review (Supertonic-3, PocketTTS).
+# The Model Catalogue shows its Accept button only when the reason matches
+# /license not accepted/i (EngineCompatibilityMatrix.reasonMentionsLicense), so
+# this sentence must keep those words: collapsing it into the generic line hid
+# the only way to enable those engines.
+_UNAVAILABLE_LICENSE = (
+    "License not accepted yet. Review and accept it in "
+    "Model Catalogue → Engines to enable this engine."
+)
+# An engine that cannot run on this machine at all: Apple-Silicon-only MLX,
+# PyTorch with no Intel Mac build. "Isn't installed yet" or "check
+# installation" sent people after an install that could never work.
+_UNAVAILABLE_PLATFORM = (
+    "This engine doesn't run on this computer's platform. Its guide lists "
+    "the platforms it supports."
+)
+# Apple Silicon whose PyTorch cannot use the GPU (MPS): the platform is
+# right, the installation is not. MLX-Audio / MLX-Whisper need MPS (#390).
+_UNAVAILABLE_NO_MPS = (
+    "This engine needs Apple's GPU (MPS), and this installation's PyTorch "
+    "can't use it. Updating macOS or reinstalling VoiceStudio usually "
+    "restores it."
+)
+_UNAVAILABLE_NEEDS_CONFIG = (
+    "This engine needs to be configured before it can run. Open "
+    "Model Catalogue → Engines to finish setting it up."
+)
+_UNAVAILABLE_FILE_MISSING = (
+    "A file this engine needs is missing or unreadable. Reinstall it from "
+    "Model Catalogue → Engines."
+)
+
+# The same two cases for an engine the app cannot install for you. "Install it
+# from Model Catalogue → Engines" sent people to a page with no Install button
+# for that engine — most of the catalogue — which reads as the app being
+# broken. The row's own guide link (``docs_url``) is the real next step.
+_UNAVAILABLE_NOT_INSTALLED_MANUAL = (
+    "This engine isn't installed yet, and it has no one-click install. "
+    "Its guide lists the install steps."
+)
+_UNAVAILABLE_FILE_MISSING_MANUAL = (
+    "A file this engine needs is missing or unreadable. Its guide lists the "
+    "install steps."
+)
+_MANUAL_INSTALL_VARIANT = {
+    _UNAVAILABLE_NOT_INSTALLED: _UNAVAILABLE_NOT_INSTALLED_MANUAL,
+    _UNAVAILABLE_FILE_MISSING: _UNAVAILABLE_FILE_MISSING_MANUAL,
+}
+
+# Matched against the lowered probe text. Ordered most specific first: a
+# missing file often also says "not installed", and the file case has the more
+# useful remedy of the two.
+_UNAVAILABLE_SIGNATURES = (
+    # First: its probe text also says "Open Model Catalogue", and the
+    # license is the one gap only the user can close.
+    (_UNAVAILABLE_LICENSE, ("license not accepted",)),
+    # Before the install and file checks: a platform reason often also says
+    # "unavailable" or names a missing wheel, and no install can fix it. Not
+    # "apple silicon only": mlx-audio says that on an M-series Mac too, when
+    # the package is merely missing and installing does help.
+    (_UNAVAILABLE_PLATFORM, (
+        "requires apple silicon", "not supported on this platform",
+        "unavailable on intel macs", "no macos x86_64 wheel",
+        "no windows install", "not supported on windows",
+    )),
+    (_UNAVAILABLE_NO_MPS, ("torch mps unavailable",)),
+    (_UNAVAILABLE_FILE_MISSING, (
+        "file is missing", "file is empty", "file is unreadable",
+        "script missing", "binary", "not found at",
+    )),
+    (_UNAVAILABLE_NEEDS_CONFIG, (
+        "environment variable", "configure a server endpoint", "api key",
+        "unconfigured", "set the", "base url",
+    )),
+    (_UNAVAILABLE_NOT_INSTALLED, (
+        "not installed", "package missing", "not available", "no module named",
+        "import ", "unavailable:", "failed to load",
+    )),
+)
+
+
+def _public_unavailable_reason(diagnostic: object) -> str:
+    """Map a private availability probe to an accurate stable category."""
+    private = diagnostic.lower() if isinstance(diagnostic, str) else ""
+    for public, markers in _UNAVAILABLE_SIGNATURES:
+        if any(marker in private for marker in markers):
+            return public
+    return _UNAVAILABLE
+
+
 def public_backends(entries: list[dict]) -> list[dict]:
     """Copy registry entries while replacing service diagnostics.
 
     Availability probes may contain exception text, local paths, tracebacks, or
-    credentials. Installation hints are registry-authored and remain intact.
+    credentials. Registry-authored fields are not probe output and remain
+    intact: ``install_hint``, ``setup_snippet`` and ``docs_url`` are all
+    VoiceStudio-owned constants keyed on the engine id, so an unavailable row
+    still has something actionable to show and somewhere to send the user
+    (#1866) even though ``reason``/``last_error`` are replaced here.
     """
     safe: list[dict] = []
     for entry in entries:
         item = dict(entry)
         if item.get("reason") is not None:
-            item["reason"] = _UNAVAILABLE
+            reason = _public_unavailable_reason(item["reason"])
+            # Only a row that explicitly says it has NO one-click install gets
+            # the manual wording. Rows without the field (ASR, LLM,
+            # translation — some of which have installers of their own) keep
+            # the line that points at Model Catalogue.
+            if item.get("one_click_install") is False:
+                reason = _MANUAL_INSTALL_VARIANT.get(reason, reason)
+            item["reason"] = reason
         if item.get("last_error") is not None:
             item["last_error"] = _PREVIOUS_FAILURE
         if item.get("routing_reason") is not None:

backend/api/routers/archetypes.py

@@ -57,11 +57,12 @@ _PREVIEW_SEED = 42
 # 32 reliably converges to speech across the gallery's instruct/script space
 # at a one-time (cached) render cost.
 _PREVIEW_NUM_STEP = 32
-# Spectral-flatness floor below which a render is a degenerate tonal artifact
-# rather than speech. Real, mastered speech sits ~0.04–0.07; a tonal buzz
-# collapses to <0.005. 0.015 separates the two with wide margin and sits well
-# below even breathy/whisper voices (which are broadband → high flatness).
-_DEGENERATE_FLATNESS = 0.015
+# Reject near-pure tonal artifacts using mean framed spectral flatness.
+# Calibrated against the tracked speech demos exercised by
+# test_archetype_preview_quality.py: the quietest (Mandarin dubbing, 44.1 kHz)
+# measures ~7.7e-6, while the worst tested tonal buzz measures ~3.3e-9.
+# 1e-7 leaves >10x margin on both sides without rejecting low-flatness speech.
+_DEGENERATE_FLATNESS = 1e-7
 
 
 def _preview_key(a: dict) -> str:
@@ -248,24 +249,46 @@ def _is_blank_audio(audio_tensor) -> bool:
         return False
 
 
+_FLATNESS_FRAME = 1024
+_FLATNESS_HOP = 512
+#: Frames quieter than this fraction of the loudest frame's energy are the gaps
+#: between words, not speech; their spectrum is the noise floor and averaging it
+#: in drags the measurement toward the value of whatever silence sounds like.
+_FLATNESS_FRAME_FLOOR = 1e-4
+
+
 def _spectral_flatness(audio_tensor) -> Optional[float]:
-    """Geometric-mean / arithmetic-mean of the power spectrum.
+    """Mean per-frame geometric-mean / arithmetic-mean of the power spectrum.
 
     ~1.0 for broadband noise, →0 for a pure tone. The degenerate diffusion
-    renders this guards against are near-pure tonal buzzes (flatness <0.005),
-    distinct from both silence (caught by ``_is_blank_audio``) and real speech
-    (~0.04+). Returns ``None`` if it can't be computed so callers don't act on
-    a bad measurement.
+    renders this guards against are near-pure tonal buzzes, distinct from both
+    silence (caught by ``_is_blank_audio``) and real speech. Returns ``None``
+    if it can't be computed so callers don't act on a bad measurement.
+
+    Measured over short frames and averaged — the standard definition. A single
+    FFT of the whole clip (what this used to do) is not the same quantity: its
+    frequency resolution grows with clip length, so speech harmonics carve
+    ever-deeper nulls into the spectrum and the geometric mean collapses. That
+    made the result depend on how long the clip was rather than on what it
+    sounded like, and put real speech below the rejection threshold.
     """
     try:
         import torch
 
         t = audio_tensor if isinstance(audio_tensor, torch.Tensor) else torch.as_tensor(audio_tensor)
-        t = t.detach().to("cpu", dtype=torch.float32).flatten()
-        if t.numel() < 1024 or not torch.isfinite(t).all():
+        t = t.detach().to("cpu", dtype=torch.float32)
+        if t.ndim > 1:
+            t = t.mean(dim=0)
+        t = t.flatten()
+        if t.numel() < _FLATNESS_FRAME or not torch.isfinite(t).all():
             return None
-        spec = torch.fft.rfft(t * torch.hann_window(t.numel())).abs().pow(2) + 1e-12
-        return float(torch.exp(torch.mean(torch.log(spec))) / torch.mean(spec))
+        frames = t.unfold(0, _FLATNESS_FRAME, _FLATNESS_HOP)
+        spec = torch.fft.rfft(frames * torch.hann_window(_FLATNESS_FRAME)).abs().pow(2) + 1e-12
+        energy = spec.sum(dim=1)
+        spec = spec[energy > energy.max() * _FLATNESS_FRAME_FLOOR]
+        if spec.shape[0] == 0:
+            return None
+        return float((torch.exp(spec.log().mean(dim=1)) / spec.mean(dim=1)).mean())
     except Exception:  # never let the checker itself block a render
         return None
 

backend/api/routers/audiobook.py

@@ -718,6 +718,10 @@ def _remote_chapter_call(chapter, *, engine_id, default_voice, voice_map,
         "expressive": opts.to_manifest(), "watermark": bool(watermark_enabled()),
     }
     signature = hashlib.sha256(json.dumps(params, sort_keys=True, default=str).encode()).hexdigest()
+    # The worker synthesizes from ``spans``, but the gateway and scheduler read
+    # top-level ``text`` to scale the remote execution deadline. Add this after
+    # the signature so existing content-addressed remote cache keys still hit.
+    params["text"] = "\n".join(row["text"] for row in rows)
     wav_path = os.path.join(cache_dir, f"remote-{signature}.wav")
 
     def decode(result):
@@ -739,7 +743,7 @@ async def _run_chapter(chapter, *, operation="audiobook", decision, job, default
                        voice_map, lexicon, cache_dir):
     """Run one chapter through the gateway; local preparation stays lazy."""
     from services import gpu_gateway
-    from services.tts_backend import active_backend_id
+    from services.tts_backend import active_backend_id, get_backend_class
 
     engine_id = active_backend_id()
     remote, remote_cache = _remote_chapter_call(
@@ -753,15 +757,27 @@ async def _run_chapter(chapter, *, operation="audiobook", decision, job, default
         return remote_cache, float(info.duration), True, None
 
     async def prepare_local():
+        from services.model_manager import generate_timeout_s
+
         synth, sr, resolve, local_engine = await _prepare_synth(
             default_voice, language=language, opts=opts, voice_map=voice_map
         )
+        try:
+            timeout_engine = get_backend_class(local_engine)
+        except ValueError:
+            # Tests and third-party integrations may inject a synth under a
+            # non-catalogue id. Keep the canonical host/text policy available;
+            # registered production engines still add their routing metadata.
+            timeout_engine = None
         return gpu_gateway.LocalCall(
             fn=lambda: _render_chapter_cached(
                 chapter, synth, sr, local_engine, resolve, cache_dir, lexicon,
                 language, opts, voice_map,
             ),
             what="Audiobook chapter",
+            timeout=generate_timeout_s(
+                remote.params["text"], engine=timeout_engine
+            ),
         )
 
     return await gpu_gateway.run(

backend/api/routers/capture.py

@@ -28,6 +28,17 @@ router = APIRouter()
 logger = logging.getLogger("omnivoice.capture")
 
 
+def _timing(value):
+    """A segment timing, or ``None`` when the engine could not determine one.
+
+    ``dict.get(key, 0)`` hands back a stored ``None`` rather than the default,
+    because the key is present — so rounding it raised and took a transcript
+    that was otherwise fine down with it (#1904). Pass the null through instead:
+    the segment list renders whichever half of the range is known.
+    """
+    return round(value, 2) if isinstance(value, (int, float)) else None
+
+
 def _truthy(value: Optional[str]) -> bool:
     """Parse a multipart form flag. Treats '1'/'true'/'yes'/'on'/'auto'
     (any case) as on; everything else — including None — as off."""
@@ -162,10 +173,15 @@ async def transcribe_audio(
         from services.text_polish import polish_text
         full_text = polish_text(full_text)
 
-        # Calculate audio duration from segments if available
+        # Calculate audio duration from segments if available. A segment whose
+        # timing the engine could not determine carries end=None (sherpa's
+        # _sherpa_result when the sample rate yields no duration, and every
+        # plain-text OpenAI-compatible response), so measure only the ones that
+        # have a number and keep 0.0 when none do.
         duration = 0.0
         if segments:
-            duration = max(s.get("end", 0) for s in segments)
+            ends = [e for e in (s.get("end") for s in segments) if isinstance(e, (int, float))]
+            duration = max(ends) if ends else 0.0
 
         detected_lang = result.get("language", language or "unknown")
 
@@ -194,8 +210,8 @@ async def transcribe_audio(
             "text": full_text,
             "segments": [
                 {
-                    "start": round(s.get("start", 0), 2),
-                    "end": round(s.get("end", 0), 2),
+                    "start": _timing(s.get("start", 0)),
+                    "end": _timing(s.get("end", 0)),
                     "text": s.get("text", "").strip(),
                 }
                 for s in segments

backend/api/routers/capture_ws.py

@@ -55,6 +55,17 @@ from services.text_polish import polish_text
 router = APIRouter()
 logger = logging.getLogger("omnivoice.capture_ws")
 
+
+def _timing(value):
+    """A segment timing, or ``None`` when the engine could not determine one.
+
+    ``dict.get(key, 0)`` returns a stored ``None`` rather than the default, so
+    rounding it raised (#1904). The null is the honest answer here — this module
+    emits it deliberately for un-endpointed utterances — and the segment list
+    renders whichever half of the range is known.
+    """
+    return round(value, 2) if isinstance(value, (int, float)) else None
+
 SPEECH_PROTOCOL = "voicestudio.speech.v1"
 PLATFORM_STREAM_PATH = "/v1/audio/transcriptions/stream"
 
@@ -349,6 +360,7 @@ async def ws_transcribe(websocket: WebSocket):
     audio_chunks: list[bytes] = []
     total_bytes = 0
     last_audio_time = time.monotonic()
+    paused = False
     running = True
     partial_text = ""
     # Track whether the client initiated the disconnect. When True the
@@ -366,7 +378,7 @@ async def ws_transcribe(websocket: WebSocket):
         message as the authoritative result and skip the duplicate HTTP
         POST that used to run on every dictation.
         """
-        nonlocal total_bytes, last_audio_time, running, client_disconnected
+        nonlocal total_bytes, last_audio_time, running, client_disconnected, paused
         try:
             while running:
                 msg = await websocket.receive()
@@ -397,6 +409,10 @@ async def ws_transcribe(websocket: WebSocket):
                     total_bytes += len(data)
                     last_audio_time = time.monotonic()
                     continue
+                if msg.get("text") in ("PAUSE", "RESUME"):
+                    paused = msg["text"] == "PAUSE"
+                    last_audio_time = time.monotonic()
+                    continue
                 if _is_end_control(msg.get("text")):
                     # Client signals end-of-audio but stays connected for `final`.
                     running = False
@@ -429,6 +445,9 @@ async def ws_transcribe(websocket: WebSocket):
             if not running:
                 break
 
+            if paused:
+                continue
+
             # Check silence timeout
             if time.monotonic() - last_audio_time > SILENCE_TIMEOUT_S and total_bytes > MIN_BUFFER_BYTES:
                 running = False
@@ -1187,13 +1206,19 @@ async def _transcribe_buffer_full(
             from services.refinement import collapse_repetitive_artifacts
             full_text = collapse_repetitive_artifacts(full_text)
 
-            duration = max((s.get("end", 0) for s in segments), default=0.0)
+            # end=None means the engine could not determine the timing — this
+            # module writes exactly that in its own streaming payloads, and
+            # sherpa's _sherpa_result does too when the sample rate yields no
+            # duration. Measure only real numbers, and pass the nulls through
+            # rather than rounding them (#1904).
+            ends = [e for e in (s.get("end") for s in segments) if isinstance(e, (int, float))]
+            duration = max(ends) if ends else 0.0
 
             return {
                 "text": full_text,
                 "segments": [
-                    {"start": round(s.get("start", 0), 2),
-                     "end": round(s.get("end", 0), 2),
+                    {"start": _timing(s.get("start", 0)),
+                     "end": _timing(s.get("end", 0)),
                      "text": s.get("text", "").strip()}
                     for s in segments
                 ],

backend/api/routers/dictation.py

@@ -86,6 +86,18 @@ def list_dictation_models():
     }
 
 
+@router.get("/dictation/readiness", dependencies=[Depends(require_local)])
+def dictation_readiness(model_id: str | None = None) -> dict:
+    """Check capture's model selection without loading or downloading weights."""
+    from services.asr_backend import asr_model_missing_error
+
+    missing = asr_model_missing_error(
+        purpose="dictation",
+        sherpa_model_id=model_id or _read_prefs()["model_id"],
+    )
+    return {"ready": missing is None, "missing": missing}
+
+
 @router.get("/dictation/prefs", dependencies=[Depends(require_local)])
 def get_dictation_prefs():
     return _read_prefs()

backend/api/routers/dub_core.py

@@ -540,12 +540,30 @@ _DUB_SOURCE_LANG_CODES = frozenset({
 
 
 def _source_lang_override(value: str | None) -> str | None:
-    """Normalize a user-selected source language; auto/und means detect."""
+    """Normalize a user-selected source language; auto/und means detect.
+
+    A rejection NAMES the code it rejected. "Invalid source language code" on
+    its own cannot be acted on or reported usefully: it does not say which of
+    the ninety-odd codes was wrong, so neither the user nor a maintainer
+    reading the auto-filed issue can tell whether the picker offered something
+    the backend does not accept, or a stale preference from an older build is
+    still being sent (#1960).
+
+    The value is a language code the user chose from a menu — not private
+    data — and the neighbouring engine validator already echoes its input the
+    same way.
+    """
     code = (value or "").strip().lower()
     if code in {"", "auto", "und"}:
         return None
     if code not in _DUB_SOURCE_LANG_CODES:
-        raise HTTPException(status_code=400, detail="Invalid source language code")
+        raise HTTPException(
+            status_code=400,
+            detail=(
+                f"Invalid source language code: {code!r}. Pick a language from "
+                "the Dubbing source-language menu, or leave it on auto-detect."
+            ),
+        )
     return code
 
 

backend/api/routers/engines.py

@@ -42,10 +42,26 @@ _FAMILIES = {
 }
 
 
+def _catalogue_active_id(family: str, module) -> str:
+    """Return the active id represented by the public engine catalogue."""
+    active = module.active_backend_id()
+    if family != "tts" or active != "omnivoice-subprocess":
+        return active
+
+    from core.device_caps import detect_host_caps
+
+    try:
+        return "omnivoice" if detect_host_caps().family == "mps" else active
+    except Exception:
+        return active
+
+
 def _family_payload(family: str, module):
     """Public inventory plus whether an environment pin owns this family."""
     return {
-        "active": module.active_backend_id(),
+        # MPS hides the explicit compatibility row, so legacy configs report
+        # the visible canonical equivalent as active to picker consumers.
+        "active": _catalogue_active_id(family, module),
         "env_override": bool(os.environ.get(f"OMNIVOICE_{family.upper()}_BACKEND")),
         "backends": public_backends(module.list_backends()),
     }
@@ -188,6 +204,11 @@ async def uninstall_translation_engine(engine_id: str):
     pkg = entry.get("pip_package")
     if not pkg:
         return {"status": "no_op", "engine": engine_id}
+    # The builtin flag is a promise someone has to remember to make; this
+    # check does not depend on it (#2019).
+    blocked = translation_engines.uninstall_blocker(engine_id)
+    if blocked:
+        raise HTTPException(status_code=blocked[0], detail=blocked[1])
     rc, out = await translation_engines.run_pip(["uninstall", "-y", pkg])
     if rc != 0:
         raise HTTPException(status_code=500, detail=f"pip uninstall {pkg} failed ({rc}): {out[-1000:]}")
@@ -230,6 +251,11 @@ def install_sidecar_engine(engine_id: str):
     from services import sidecar_install
     try:
         return sidecar_install.start_install(engine_id)
+    except sidecar_install.HostUnsupported as exc:
+        # The engine has an installer, but not one that can work on this
+        # machine. 409, not 404: the route is right, the host is the problem,
+        # and the message (a VoiceStudio-owned sentence) says what to do.
+        raise HTTPException(status_code=409, detail=str(exc))
     except KeyError:
         raise HTTPException(
             status_code=404,
@@ -354,6 +380,9 @@ def engine_health(engine_id: str):
         )
 
     t0 = perf_counter()
+    # Stable exception class when the probe itself raised, None when it merely
+    # returned not-available. Never the exception text — see the log line below.
+    raised_class: str | None = None
     if hasattr(cls, "health_check"):
         # SubprocessBackend path — spawn sidecar (if not running) and ping.
         # ``health_check`` already swallows its own exceptions per Plan
@@ -364,6 +393,7 @@ def engine_health(engine_id: str):
             ok, msg = instance.health_check()
         except Exception as exc:
             ok, msg = False, f"{type(exc).__name__}: {exc}"
+            raised_class = type(exc).__name__
     else:
         # In-process backend — `is_available()` is the classmethod-level
         # liveness check. Cheap and side-effect-free for every shipping
@@ -372,6 +402,7 @@ def engine_health(engine_id: str):
             ok, msg = cls.is_available()
         except Exception as exc:
             ok, msg = False, f"{type(exc).__name__}: {exc}"
+            raised_class = type(exc).__name__
 
     # Engine-owned output can contain much more than shaped HF tokens: local
     # paths, arbitrary credentials, source lines, or a nested traceback.
@@ -379,7 +410,38 @@ def engine_health(engine_id: str):
 
     latency_ms = (perf_counter() - t0) * 1000.0
     if not ok:
-        logger.warning("Engine health check failed; details withheld")
+        # The response tells the user to "check the backend log for details",
+        # and docs/engines/*.md asks a user diagnosing an unavailable engine to
+        # copy that engine's log lines. The old line named neither the engine
+        # nor anything about the probe, so neither instruction could be
+        # followed (#1866).
+        #
+        # `probe=` reports what the PROBE DID, not what went wrong. It cannot
+        # classify the cause: SubprocessBackend.health_check() swallows its own
+        # exceptions per Plan 02-01's contract, so a dead sidecar and a package
+        # that was never installed both arrive here as `returned-unavailable`.
+        # Separating those needs structured failure metadata from the probes
+        # themselves, which is a wider change than this one.
+        #
+        # Still no diagnostic text and still not the caller-supplied id: the
+        # engine id comes off the resolved registry class and a raised probe
+        # contributes only its exception class, the same shape
+        # core.public_errors.public_failure() logs as `class=`.
+        # tests/test_response_safety.py pins that boundary and passes
+        # unchanged.
+        #
+        # The id is a class attribute off the registry rather than caller
+        # input, but this line is a log-injection surface either way, so it is
+        # flattened to a single token before it goes in.
+        engine_label = str(getattr(cls, "id", None) or cls.__name__)
+        engine_label = "".join(
+            c if (c.isalnum() or c in "-_.") else "-" for c in engine_label
+        )[:64]
+        logger.warning(
+            "Engine health check failed; engine=%s probe=%s, details withheld",
+            engine_label or "unknown",
+            f"raised:{raised_class}" if raised_class else "returned-unavailable",
+        )
     return {
         "id": engine_id,
         "ok": bool(ok),
@@ -589,7 +651,15 @@ def select_engine(req: SelectEngineRequest):
     if not family:
         raise HTTPException(400, f"Unknown family: {req.family}. Expected one of tts/asr/llm.")
     module, pref_key = family
-    available = {b["id"]: b for b in module.list_backends()}
+    # MPS intentionally hides the redundant explicit OmniVoice sidecar from
+    # the picker, but existing scripts and saved preferences may still submit
+    # that supported compatibility id directly.
+    rows = (
+        module.list_backends(include_hidden=True)
+        if req.family == "tts"
+        else module.list_backends()
+    )
+    available = {b["id"]: b for b in rows}
     if req.backend_id not in available:
         raise HTTPException(400, f"Unknown {req.family} backend: {req.backend_id!r}")
     entry = available[req.backend_id]

backend/api/routers/generation.py

@@ -804,16 +804,34 @@ def _oom_friendly_reraise(e):
     ) from e
 
 
-def _generate_timeout_s(text: str, *, execution_device=None) -> float:
+def _generate_timeout_s(
+    text: str,
+    *,
+    execution_device=None,
+    min_vram_gb=0.0,
+    hardware_family=None,
+    vram_gb=None,
+) -> float:
     """Wall-clock budget for one generate, scaled to the request.
 
     Thin alias for the canonical helper, which moved to
     ``services.model_manager.generate_timeout_s`` (#1190) so /v1/audio/speech,
     batch, dub and archetype previews share it instead of each re-deriving (or,
     as they did, silently keeping the flat 300s).
+
+    ``min_vram_gb`` is the engine's declared VRAM floor. A GPU below it pages to
+    system RAM and renders slower than this machine's CPU, so it must not be
+    budgeted as fast hardware (#1804) — the same figure the dispatch already
+    hands the guard so a timeout message can name the card (#1226/#1222).
     """
     from services.model_manager import generate_timeout_s
-    return generate_timeout_s(text, execution_device=execution_device)
+    return generate_timeout_s(
+        text,
+        execution_device=execution_device,
+        min_vram_gb=min_vram_gb,
+        hardware_family=hardware_family,
+        vram_gb=vram_gb,
+    )
 
 
 def _run_inference(
@@ -1434,6 +1452,8 @@ async def generate_speech(
     # local fallback call's timeout device-neutral so the closure is valid
     # without pretending the control plane describes the remote worker.
     _routing = {"effective_device": None}
+    _routing_hardware_family = None
+    _routing_vram_gb = None
 
     if not _remote:
         # Single-active-engine memory discipline: hand back any OTHER resident
@@ -1490,11 +1510,16 @@ async def generate_speech(
         # 4090 from a Mac control plane would be refused by a gate describing
         # a machine that is about to do nothing.
         from core.device_caps import detect_host_caps
-        from services.engine_routing import resolve_routing, routing_notice
-        _routing = resolve_routing(
-            getattr(backend_cls, "gpu_compat", ("cpu",)), detect_host_caps(),
-            _engine_min_vram_gb,
+        from services.engine_routing import (
+            routing_notice,
+            runtime_compute_profile_async,
         )
+        _routing = await runtime_compute_profile_async(
+            backend_cls, detect_host_caps()
+        )
+        _engine_min_vram_gb = _routing["min_vram_gb"]
+        _routing_hardware_family = _routing.get("runtime_hardware_family")
+        _routing_vram_gb = _routing.get("runtime_vram_gb")
         if _routing["routing_status"] == "unavailable":
             # The engine needs an accelerator this host lacks and has no CPU path.
             raise HTTPException(status_code=400, detail=_routing["routing_reason"])
@@ -1718,7 +1743,13 @@ async def generate_speech(
             local=gpu_gateway.LocalCall(
                 _remote_only_local_call(_target_label),
                 what="TTS generate",
-                timeout=_generate_timeout_s(text, execution_device=_routing["effective_device"]),
+                timeout=_generate_timeout_s(
+                    text,
+                    execution_device=_routing["effective_device"],
+                    min_vram_gb=_engine_min_vram_gb,
+                    hardware_family=_routing_hardware_family,
+                    vram_gb=_routing_vram_gb,
+                ),
                 min_vram_gb=_engine_min_vram_gb,
             ),
             remote=_remote_call,
@@ -2012,7 +2043,13 @@ async def generate_speech(
                                 ),
                                 what="TTS generate",
                                 min_vram_gb=_engine_min_vram_gb,
-                                timeout=_generate_timeout_s(text, execution_device=_routing["effective_device"]),
+                                timeout=_generate_timeout_s(
+                                    text,
+                                    execution_device=_routing["effective_device"],
+                                    min_vram_gb=_engine_min_vram_gb,
+                                    hardware_family=_routing_hardware_family,
+                                    vram_gb=_routing_vram_gb,
+                                ),
                                 on_abandon=release,
                             )
                         )
@@ -2032,7 +2069,13 @@ async def generate_speech(
                                 ),
                                 what="TTS generate",
                                 min_vram_gb=_engine_min_vram_gb,
-                                timeout=_generate_timeout_s(text, execution_device=_routing["effective_device"]),
+                                timeout=_generate_timeout_s(
+                                    text,
+                                    execution_device=_routing["effective_device"],
+                                    min_vram_gb=_engine_min_vram_gb,
+                                    hardware_family=_routing_hardware_family,
+                                    vram_gb=_routing_vram_gb,
+                                ),
                                 on_abandon=release,
                             )
                         )
@@ -2072,7 +2115,13 @@ async def generate_speech(
                                 # Budget scaled to THIS chunk (#1190) — the flat
                                 # 300s here is what made long streamed renders fail
                                 # even after the v0.3.22 scaled budget shipped.
-                                timeout=_generate_timeout_s(chunk_text, execution_device=_routing["effective_device"]),
+                                timeout=_generate_timeout_s(
+                                    chunk_text,
+                                    execution_device=_routing["effective_device"],
+                                    min_vram_gb=_engine_min_vram_gb,
+                                    hardware_family=_routing_hardware_family,
+                                    vram_gb=_routing_vram_gb,
+                                ),
                                 on_abandon=release,
                             )
                         )
@@ -2232,7 +2281,13 @@ async def generate_speech(
                     _REMOTE_OP,
                     local=gpu_gateway.LocalCall(
                         _local_render, what="TTS generate",
-                        timeout=_generate_timeout_s(text, execution_device=_routing["effective_device"]),
+                        timeout=_generate_timeout_s(
+                            text,
+                            execution_device=_routing["effective_device"],
+                            min_vram_gb=_engine_min_vram_gb,
+                            hardware_family=_routing_hardware_family,
+                            vram_gb=_routing_vram_gb,
+                        ),
                         min_vram_gb=_engine_min_vram_gb,
                         on_abandon=release,
                     ),
@@ -2357,7 +2412,16 @@ async def generate_speech(
         raise HTTPException(status_code=503, detail=str(e)) from e
     except ValueError as e:
         logger.error("Validation failed: %s", e)
-        raise HTTPException(status_code=400, detail=str(e)) from e
+        # Most ValueErrors here are VoiceStudio's own validation messages and
+        # are exactly what the user should read. A few are raw library text
+        # naming parameters and files the user cannot act on — those get the
+        # owned remedy for their class instead (#1879). Unclassified ones keep
+        # passing through, so this cannot swallow a good message.
+        from core.failure import classify, public_hint_for_topic
+
+        _topic = classify(str(e))
+        _owned = public_hint_for_topic(_topic) if _topic else ""
+        raise HTTPException(status_code=400, detail=_owned or str(e)) from e
     except Exception as e:
         tb = traceback.format_exc()
         logger.error("Inference failed: %s\n%s", e, tb)

backend/api/routers/openai_compat.py

@@ -325,10 +325,9 @@ async def create_speech(req: SpeechRequest):
 
     # Routing gate (#21 — no silent CPU fallback), identical to REST /generate.
     from core.device_caps import detect_host_caps
-    from services.engine_routing import resolve_routing, routing_notice
-    _routing = resolve_routing(
-        getattr(backend, "gpu_compat", ("cpu",)), detect_host_caps(),
-        getattr(backend, "min_vram_gb", 0.0),
+    from services.engine_routing import routing_notice, runtime_compute_profile_async
+    _routing = await runtime_compute_profile_async(
+        backend, detect_host_caps()
     )
     if _routing["routing_status"] == "unavailable":
         raise HTTPException(status_code=400, detail=_routing["routing_reason"])

backend/api/routers/pronunciation.py

@@ -32,7 +32,11 @@ from pydantic import BaseModel
 
 from api.dependencies import require_admin
 from core.db import db_conn
-from services.pronunciation import apply_pronunciation, entries_for_language
+from services.pronunciation import (
+    apply_pronunciation,
+    entries_for_language,
+    inert_entries_for_language,
+)
 
 logger = logging.getLogger("omnivoice.pronunciation")
 router = APIRouter(dependencies=[Depends(require_admin)])
@@ -250,11 +254,18 @@ def test_substitution(req: PronTestRequest):
         ).fetchall()
     substituted = apply_pronunciation(req.text, rows, req.language)
     applied = entries_for_language(rows, req.language)
+    # IPA/CMU rows are validated and stored but not applied yet, so a term that
+    # DOES match can still change nothing. Reporting them separately keeps the
+    # dry run honest — otherwise it says "no entries match", which is wrong and
+    # sends the user to re-type an entry that was already correct (#1949).
+    inert = inert_entries_for_language(rows, req.language)
     return {
         "input": req.text,
         "substituted": substituted,
         "changed": substituted != req.text,
         "applied_terms": sorted(applied.keys(), key=len, reverse=True),
+        # Present but not honoured: [{term, type}, …]. Empty on the happy path.
+        "inert_entries": inert,
     }
 
 

backend/api/routers/settings.py

@@ -35,11 +35,11 @@ class _HFTokenBody(BaseModel):
     token: str = Field(..., min_length=1, description="HuggingFace access token")
 
 
-def _state_response() -> dict:
+def _state_response(*, validate: bool = False) -> dict:
     """Return the same shape the React panel renders. Never includes raw token."""
     from services import token_resolver
 
-    s = token_resolver.state()
+    s = token_resolver.state(validate=validate)
     return {
         "active": s["active"],
         "sources": [asdict(row) for row in s["sources"]],
@@ -65,8 +65,7 @@ def save_hf_token(body: _HFTokenBody):
 
 @router.delete("/hf-token")
 def clear_hf_token(also_clear_hf_cli: bool = Query(False)):
-    """Clear the App-source token. Optionally also call huggingface_hub.logout
-    to clear the canonical HF file. Returns the updated cascade state."""
+    """Clear the App token and optionally recognized local Hub token files."""
     from services import token_resolver
     try:
         token_resolver.clear_app_token(also_clear_hf_cli=also_clear_hf_cli)
@@ -82,13 +81,12 @@ def get_hf_token_state(fresh: bool = Query(False)):
 
     ``fresh=1`` drops the resolver's whoami validation cache first so the
     response re-runs whoami for every source — this is what the panel's
-    "Test now" button sends. Plain GETs (panel mounts) keep the 300s cache
-    so repeat Settings visits don't hammer the HF API.
+    "Test now" button sends. Plain GETs only inspect local token presence.
     """
     from services import token_resolver
     if fresh:
         token_resolver.invalidate_cache()
-    return _state_response()
+    return _state_response(validate=fresh)
 
 
 # ── Performance settings (INST-12) ────────────────────────────────────────
@@ -150,7 +148,7 @@ def _compute_device_state() -> dict:
     caps = device_caps.detect_host_caps()
     env_pin = (os.environ.get("OMNIVOICE_DEVICE") or "").strip().lower()
     auto_family = next(
-        (f for f in ("cuda", "rocm", "xpu", "mps") if f in caps.available_families),
+        (f for f in device_caps.ACCELERATOR_PRIORITY if f in caps.available_families),
         "cpu",
     )
     value = device_caps.requested_device_override()

backend/api/routers/setup/download.py

@@ -381,11 +381,60 @@ def _is_retryable_download_error(exc: BaseException) -> bool:
     return is_hf_connectivity_error(str(exc))
 
 
+def _segmented_retry_plan(
+    exc: BaseException, attempt: int, max_attempts: int
+) -> tuple[bool, bool]:
+    """What to do after the segmented accelerator failed on ``attempt``.
+
+    Returns ``(disable_accelerator, reraise)``.
+
+    A dropped connection is not the accelerator's fault, so the error is
+    re-raised for the outer retry: the next attempt re-enters
+    :func:`_segmented_snapshot`, which resumes from the ``.part`` manifest.
+    Falling straight through to ``snapshot_download`` instead would finish the
+    install from a separate ``.incomplete`` file and strand that manifest — the
+    restart-from-zero this exists to prevent.
+
+    The final attempt is always reserved for the plain path, so the accelerator
+    can never be the reason an install fails outright. The two flags are
+    decoupled for that handover: the attempt that exhausts the accelerator still
+    re-raises, so the plain path starts on the LAST attempt rather than the
+    second-to-last. Disabling and falling through in the same attempt would
+    abandon the resumable manifest one attempt early and restart through a
+    separate file — which is the failure this whole helper exists to avoid.
+    """
+    if not _is_retryable_download_error(exc):
+        return True, False  # the accelerator cannot work here at all
+    if attempt >= max_attempts:
+        # Nothing left to hand over to: take the plain path now rather than
+        # re-raising out of the loop with no fallback ever tried.
+        return True, False
+    return attempt >= max_attempts - 1, True
+
+
+def _segmented_retry_note(disable: bool, reraise: bool) -> str:
+    """How to describe the outcome of :func:`_segmented_retry_plan` in the log.
+
+    Three distinct states, and reading only ``disable`` conflates two of them:
+    the attempt that exhausts the accelerator is disabled AND re-raises, so the
+    fallback starts on the NEXT attempt, not this one.
+    """
+    if not disable:
+        return "kept for the next attempt (resumes from its manifest)"
+    if reraise:
+        return "exhausted — retrying once more, then snapshot_download takes over"
+    return "disabled for this install — falling back to snapshot_download now"
+
+
 @router.post("/models/install")
 async def install_model(req: InstallModelRequest):
     """Download one HF repo snapshot; progress goes through the shared
     ``/setup/download-stream`` SSE feed."""
-    if req.repo_id not in [m["repo_id"] for m in KNOWN_MODELS]:
+    model_spec = next(
+        (model for model in KNOWN_MODELS if model["repo_id"] == req.repo_id),
+        None,
+    )
+    if model_spec is None:
         raise HTTPException(
             status_code=400,
             detail=(
@@ -393,6 +442,7 @@ async def install_model(req: InstallModelRequest):
                 + ", ".join(m["repo_id"] for m in KNOWN_MODELS)
             ),
         )
+    allow_patterns = list(model_spec.get("allow_patterns") or []) or None
     target = (req.target or "").strip()
     if target != "local":
         from services import gpu_gateway  # noqa: PLC0415
@@ -450,6 +500,8 @@ async def install_model(req: InstallModelRequest):
                 "revision": revision_for(req.repo_id),
                 "max_workers": _download_max_workers(),
             }
+            if allow_patterns:
+                dl_kwargs["allow_patterns"] = allow_patterns
             _tqdm_cls = hf_progress.tracked_tqdm_class()
             if _tqdm_cls is not None:
                 dl_kwargs["tqdm_class"] = _tqdm_cls
@@ -493,6 +545,8 @@ async def install_model(req: InstallModelRequest):
                 "revision": dl_kwargs["revision"],
                 "dry_run": True,
             }
+            if allow_patterns:
+                _preflight_kwargs["allow_patterns"] = allow_patterns
             if _endpoint:
                 _preflight_kwargs["endpoint"] = _endpoint
             try:
@@ -548,6 +602,11 @@ async def install_model(req: InstallModelRequest):
 
             _max_attempts = 5
             _attempt = 0
+            # The accelerator is retried across attempts so its manifest-based
+            # resume actually gets used; it is disabled for the rest of the
+            # install only when it fails for a reason that is NOT transient
+            # network trouble (i.e. the accelerator itself is unusable here).
+            _segmented_off = False
             while True:
                 if req.repo_id in _cancelled:
                     raise _InstallCancelled()
@@ -555,11 +614,17 @@ async def install_model(req: InstallModelRequest):
                 try:
                     # Segmented accelerator (FDL-09, default ON): parallel
                     # byte-range fetch with real live progress, for the
-                    # legacy-LFS path. Any failure falls through to
-                    # snapshot_download — the accelerator can never compromise a
-                    # correct install.
+                    # legacy-LFS path. A failure that is not transient network
+                    # trouble falls through to snapshot_download, and so does the
+                    # install's last attempt — the accelerator can never
+                    # compromise a correct install (see _segmented_retry_plan).
                     _snapshot_path = None
-                    if _attempt == 1 and _segmented_enabled() and not _xet_active():
+                    if (
+                        not _segmented_off
+                        and not allow_patterns
+                        and _segmented_enabled()
+                        and not _xet_active()
+                    ):
                         try:
                             _snapshot_path = _segmented_snapshot(
                                 req.repo_id,
@@ -569,10 +634,16 @@ async def install_model(req: InstallModelRequest):
                         except _InstallCancelled:
                             raise
                         except Exception as _seg_err:
-                            logger.info(
-                                "segmented download for %s failed (%s); falling back to snapshot_download",
-                                req.repo_id, _seg_err,
+                            _segmented_off, _seg_reraise = _segmented_retry_plan(
+                                _seg_err, _attempt, _max_attempts
                             )
+                            logger.info(
+                                "segmented download for %s failed (%s); accelerator %s",
+                                req.repo_id, _seg_err,
+                                _segmented_retry_note(_segmented_off, _seg_reraise),
+                            )
+                            if _seg_reraise:
+                                raise
                             _snapshot_path = None
                     if _snapshot_path is None:
                         _snapshot_path = snapshot_download(**dl_kwargs)  # nosec B615 -- immutable revision_for pin

backend/api/routers/setup/wizard.py

@@ -19,6 +19,7 @@ import sys
 from fastapi import APIRouter
 
 from api.schemas import SetupStatusResponse, PreflightResponse
+from core.device_caps import KERNEL_RISK_MARKER
 # MIN_FREE_GB + disk_free_bytes are single-sourced in ``.models`` (the lowest
 # module in the setup import graph) so the wizard gate, the /models header, and
 # the per-install disk guard can't drift apart.
@@ -174,8 +175,8 @@ def _detect_gpu() -> dict:
     return info
 
 
-def _probe_network(host: str = "huggingface.co", port: int = 443, timeout: float = 2.0) -> bool:
-    """Tiny TCP connect test."""
+def _probe_network(host: str = "huggingface.co", port: int = 443, timeout: float = 8.0) -> bool:
+    """Tiny TCP connect test. 8s default — high-latency / China paths often exceed 2–3s."""
     import socket
     try:
         with socket.create_connection((host, port), timeout=timeout):
@@ -498,10 +499,14 @@ def preflight():
         _why = gpu_routing.get("routing_reason")
         if _rs == "accelerated" and not _why:
             r_status, r_detail, r_fix = "pass", f"{_eng} → {_dev} (accelerated)", None
-        elif _rs == "accelerated":  # driver/arch caveat
+        elif _rs == "accelerated" and KERNEL_RISK_MARKER in (_why or ""):
             r_status, r_detail, r_fix = "warn", f"{_eng} → {_dev}: {_why}", (
                 "GPU selected but may fail at kernel launch — update drivers / "
                 "reinstall torch for this GPU architecture.")
+        elif _rs == "accelerated":  # low-VRAM caveat — not a driver/arch issue
+            r_status, r_detail, r_fix = "warn", f"{_eng} → {_dev}: {_why}", (
+                "Unload other models before generating, keep the text short, "
+                "or pick a lighter engine.")
         elif _rs == "cpu_fallback":
             r_status, r_detail, r_fix = "warn", (
                 f"{_eng} runs on CPU here: {_why or 'no GPU path for this host'}"), (

backend/api/routers/system.py

@@ -297,6 +297,142 @@ def _tail_file(path: str, tail: int):
     return all_lines[-tail:], len(all_lines)
 
 
+# Must track main.py's _WindowsSafeRotatingFileHandler(backupCount=3). The
+# handler rolls omnivoice.log at 2 MB into .1/.2/.3, so up to 6 MB of history
+# lives in files this module used to ignore entirely.
+_LOG_BACKUP_COUNT = 3
+
+
+def _rotated_log_paths(base: str) -> list[str]:
+    """Existing `<base>.1 … .N`, newest first."""
+    return [p for p in (f"{base}.{i}" for i in range(1, _LOG_BACKUP_COUNT + 1)) if os.path.exists(p)]
+
+
+def _tail_rolling(base: str, tail: int):
+    """Tail `base`, reaching into its rotated siblings when it runs short.
+
+    A rollover leaves omnivoice.log nearly empty, and the Backend tab then
+    showed a handful of lines — or none — while the failure the user was asked
+    to copy sat in omnivoice.log.1. Reading the current file first keeps the
+    common case at one file read; the backups are only touched when they are
+    the only place the requested lines can come from.
+
+    Returns (lines oldest-first, total lines across the files read, paths read
+    oldest-first). The total counts only the files it had to open — it stops as
+    soon as `tail` is satisfied, so it is "how much is behind these lines",
+    not the size of the whole rotation set.
+    """
+    chunks: list[list[str]] = []
+    paths: list[str] = []
+    total = 0
+    remaining = tail
+    candidates = [p for p in [base, *_rotated_log_paths(base)] if os.path.exists(p)]
+    for path in candidates:
+        if remaining <= 0:
+            break
+        try:
+            lines, count = _tail_file(path, remaining)
+        except FileNotFoundError:
+            # A rollover can rename a candidate between the existence check
+            # above and this open, and the handler holds no lock we can take
+            # from a route. Skip the vanished file rather than 500 the whole
+            # panel over one member of the set — the previous single-file
+            # version failed the request outright in the same situation.
+            #
+            # A roll landing mid-walk can also shift which chunk a file holds,
+            # so a tail taken at that instant may repeat or miss a block. The
+            # panel re-polls every 5s and the next read is clean; buying strict
+            # consistency here would mean reaching into logging's internals.
+            continue
+        except PermissionError as exc:
+            # Windows only, and only the sharing violation: the handler still
+            # holds the file it is rolling. Any other permission failure is a
+            # real misconfiguration and must not be hidden.
+            if os.name == "nt" and getattr(exc, "winerror", None) == 32:
+                continue
+            raise
+        if count == 0:
+            continue
+        chunks.append(lines)
+        paths.append(path)
+        total += count
+        remaining -= len(lines)
+    # Files were visited newest-first; the reader wants oldest-first.
+    out: list[str] = []
+    for chunk in reversed(chunks):
+        out.extend(chunk)
+    return out, total, list(reversed(paths))
+
+def _tauri_plugin_log_candidates():
+    """The `tauri-plugin-log` files — the shell's own log, and the only thing
+    the Tauri tab actually displays.
+
+    Split out from :func:`_tauri_log_candidates` so Clear can touch these and
+    leave the backend stdout/stderr redirect alone. See
+    :func:`clear_tauri_logs`.
+    """
+    home = os.path.expanduser("~")
+    bid = "com.debpalash.omnivoice-studio"
+    if sys.platform == "darwin":
+        return [
+            os.path.join(home, "Library/Logs", bid, "tauri.log"),
+            os.path.join(home, "Library/Logs", bid, "VoiceStudio.log"),
+        ]
+    if sys.platform.startswith("linux"):
+        data_dir = os.environ.get("XDG_DATA_HOME") or os.path.join(home, ".local/share")
+        return [
+            os.path.join(data_dir, bid, "logs", "tauri.log"),
+            os.path.join(home, ".config", bid, "logs", "tauri.log"),
+        ]
+    if sys.platform.startswith("win"):
+        appdata = os.environ.get("APPDATA", home)
+        localappdata = os.environ.get("LOCALAPPDATA") or os.path.join(home, "AppData", "Local")
+        return [
+            os.path.join(localappdata, bid, "logs", "tauri.log"),
+            os.path.join(appdata, bid, "logs", "tauri.log"),
+        ]
+    return []
+
+
+def _backend_redirect_log_candidates():
+    """`backend.log` / `backend_err.log` — the spawned backend's stdout and
+    stderr, written by `src-tauri/src/backend.rs::backend_log_path()`.
+
+    Deliberately NOT cleared by the Tauri tab's Clear button.
+    `open_err_log_for_run()` opens `backend_err.log` **append-only** so "a
+    respawn must not destroy the previous run's evidence" (#1510), rotates it
+    to `.1` rather than truncating, and its spawn diagnostics are described
+    there as "retained in backend_err.log across runs and lands verbatim in bug
+    reports". A native death (a Windows access violation, a SIGSEGV) writes
+    nothing to the Python log by construction, so this file is the only record
+    of it.
+
+    `OMNIVOICE_LOG_DIR` is honoured first, in the same precedence
+    `backend_log_path()` uses. The backend is a child of the shell, so an
+    ambient override reaches both — and a resolver that ignored it would look
+    in the per-OS default while the writer wrote somewhere else, which is the
+    divergence class this file already has one of (see #1782).
+    """
+    override = (os.environ.get("OMNIVOICE_LOG_DIR") or "").strip()
+    if override:
+        return [
+            os.path.join(override, "backend.log"),
+            os.path.join(override, "backend_err.log"),
+        ]
+    home = os.path.expanduser("~")
+    if sys.platform == "darwin":
+        base = os.path.join(home, "Library/Logs/OmniVoice")
+    elif sys.platform.startswith("linux"):
+        state_dir = os.environ.get("XDG_STATE_HOME") or os.path.join(home, ".local/state")
+        base = os.path.join(state_dir, "OmniVoice")
+    elif sys.platform.startswith("win"):
+        localappdata = os.environ.get("LOCALAPPDATA") or os.path.join(home, "AppData", "Local")
+        base = os.path.join(localappdata, "OmniVoice", "Logs")
+    else:
+        return []
+    return [os.path.join(base, "backend.log"), os.path.join(base, "backend_err.log")]
+
+
 def _tauri_log_candidates():
     """Likely paths for Tauri-side logs, most useful first.
 
@@ -308,40 +444,15 @@ def _tauri_log_candidates():
       `com.debpalash.omnivoice-studio` (frontend/src-tauri/tauri.conf.json).
     - backend.rs::backend_log_path() redirects the spawned backend's
       stdout/stderr to `backend.log` / `backend_err.log` under
-      `~/Library/Logs/OmniVoice` (macOS), `$XDG_STATE_HOME/VoiceStudio` falling
+      `~/Library/Logs/OmniVoice` (macOS), `$XDG_STATE_HOME/OmniVoice` falling
       back to `~/.local/state/OmniVoice` (Linux), and
       `%LOCALAPPDATA%\\OmniVoice\\Logs` (Windows). This is where uvicorn
       startup banners and hard-crash tracebacks land — keep all three OS
       shapes listed or sidecar crashes become invisible off-macOS.
     """
-    home = os.path.expanduser("~")
-    bid = "com.debpalash.omnivoice-studio"
-    if sys.platform == "darwin":
-        return [
-            os.path.join(home, "Library/Logs", bid, "tauri.log"),
-            os.path.join(home, "Library/Logs", bid, "VoiceStudio.log"),
-            os.path.join(home, "Library/Logs/OmniVoice/backend.log"),
-            os.path.join(home, "Library/Logs/OmniVoice/backend_err.log"),
-        ]
-    if sys.platform.startswith("linux"):
-        data_dir = os.environ.get("XDG_DATA_HOME") or os.path.join(home, ".local/share")
-        state_dir = os.environ.get("XDG_STATE_HOME") or os.path.join(home, ".local/state")
-        return [
-            os.path.join(data_dir, bid, "logs", "tauri.log"),
-            os.path.join(home, ".config", bid, "logs", "tauri.log"),
-            os.path.join(state_dir, "OmniVoice", "backend.log"),
-            os.path.join(state_dir, "OmniVoice", "backend_err.log"),
-        ]
-    if sys.platform.startswith("win"):
-        appdata = os.environ.get("APPDATA", home)
-        localappdata = os.environ.get("LOCALAPPDATA") or os.path.join(home, "AppData", "Local")
-        return [
-            os.path.join(localappdata, bid, "logs", "tauri.log"),
-            os.path.join(appdata, bid, "logs", "tauri.log"),
-            os.path.join(localappdata, "OmniVoice", "Logs", "backend.log"),
-            os.path.join(localappdata, "OmniVoice", "Logs", "backend_err.log"),
-        ]
-    return []
+    # Composed from the two halves so the read path keeps seeing every file
+    # while Clear can be narrowed to the shell's own log.
+    return _tauri_plugin_log_candidates() + _backend_redirect_log_candidates()
 
 
 @router.get("/system/logs")
@@ -356,12 +467,24 @@ async def system_logs(tail: int = 200):
     except Exception:
         tail = 200
 
-    path = LOG_PATH if os.path.exists(LOG_PATH) else CRASH_LOG_PATH
-    if not os.path.exists(path):
+    if os.path.exists(LOG_PATH) or _rotated_log_paths(LOG_PATH):
+        base = LOG_PATH
+    else:
+        base = CRASH_LOG_PATH
+    if not os.path.exists(base) and not _rotated_log_paths(base):
         return {"lines": [], "path": LOG_PATH, "exists": False}
+    path = base
     try:
-        lines, total = await asyncio.to_thread(_tail_file, path, tail)
-        return {"lines": lines, "path": path, "exists": True, "total_lines": total}
+        lines, total, paths = await asyncio.to_thread(_tail_rolling, base, tail)
+        return {
+            "lines": lines,
+            "path": path,
+            "exists": True,
+            "total_lines": total,
+            # Which files the tail actually came from, oldest first. A bug
+            # report can then say whether it crossed a rollover.
+            "paths": paths,
+        }
     except Exception as e:
         raise HTTPException(
             status_code=500,
@@ -467,9 +590,23 @@ def _read_from_pos(path: str, pos: int) -> list[str]:
 
 @router.post("/system/logs/clear")
 async def clear_system_logs():
-    """Truncate the rolling runtime log and the crash log (what the Backend tab reads)."""
+    """Truncate the rolling runtime log and the crash log (what the Backend tab reads).
+
+    Includes the rotated siblings. Truncating only omnivoice.log left up to
+    6 MB in .1/.2/.3, so Clear freed almost nothing and — now that the tail
+    reaches into those files — would have looked like it did nothing at all.
+    """
     cleared_any = False
-    for p in (LOG_PATH, CRASH_LOG_PATH):
+    # The full fixed name set rather than a snapshot of what exists: enumerating
+    # first leaves a window where a rollover creates a backup after the scan and
+    # its history survives a Clear that reported success. Names the handler can
+    # ever write are known up front, so there is nothing to enumerate.
+    targets = [
+        LOG_PATH,
+        *(f"{LOG_PATH}.{i}" for i in range(1, _LOG_BACKUP_COUNT + 1)),
+        CRASH_LOG_PATH,
+    ]
+    for p in targets:
         if os.path.exists(p):
             try:
                 await asyncio.to_thread(_truncate_file, p)
@@ -502,10 +639,20 @@ def _truncate_file(path: str):
 
 @router.post("/system/logs/tauri/clear")
 async def clear_tauri_logs():
-    """Truncate whichever Tauri-side log files we know about. OS-level rotation may recreate them."""
+    """Truncate the shell's own log files. OS-level rotation may recreate them.
+
+    The backend stdout/stderr redirect is deliberately excluded. This button
+    lives on a tab that shows `tauri.log`, and truncating `backend_err.log`
+    from it destroyed evidence the user was never shown — the one record of a
+    native death, which writes nothing to the Python log. `backend.rs`'s
+    `open_err_log_for_run()` opens that file append-only precisely so "a
+    respawn must not destroy the previous run's evidence" (#1510) and rotates
+    it to `.1` instead of truncating, so it manages its own size and does not
+    need clearing from here.
+    """
     cleared = []
     failed = 0
-    for p in _tauri_log_candidates():
+    for p in _tauri_plugin_log_candidates():
         if os.path.exists(p):
             try:
                 await asyncio.to_thread(_truncate_file, p)
@@ -910,7 +1057,7 @@ async def set_env_var(body: dict):
     Persistent keys (proxy, FFMPEG_PATH, translation provider keys, …) are
     saved to ``prefs.json`` so they survive backend restarts (restored at
     startup in ``main.py``). HF_TOKEN is persisted via
-    ``huggingface_hub.login()`` (and cleared via ``logout()``). Other keys
+    ``huggingface_hub.login()`` (and cleared with the shared token-file helper). Other keys
     are set on ``os.environ`` for the running process.
 
     The loopback-origin gate that previously lived inline here is now applied
@@ -985,14 +1132,14 @@ async def set_env_var(body: dict):
         # Mirror the persistence on clear — wipe the saved token file too.
         if key == "HF_TOKEN":
             try:
-                from huggingface_hub import logout as _hf_logout
-                _hf_logout()
-                logger.info("HF token cleared from $HF_HOME/token via logout()")
-            except Exception as e:
-                logger.warning("Could not clear HF token file: %s", e)
+                from services.token_resolver import clear_hf_cli_tokens
+                clear_hf_cli_tokens()
+                logger.info("Local Hugging Face token files cleared")
+            except Exception:
+                raise HTTPException(status_code=500, detail="Could not clear local Hugging Face token files") from None
 
     # HF_TOKEN persistence is handled above via huggingface_hub.login()/
-    # logout() — it never touches prefs.json. Everything else in
+    # clear_hf_cli_tokens() — it never touches prefs.json. Everything else in
     # PERSISTENT_KEYS (proxy, FFMPEG_PATH, translation provider keys, …) is
     # saved to prefs.json so it survives backend restarts (restored at
     # startup in main.py). Non-persistent keys stay process-local.
@@ -1100,7 +1247,7 @@ def asr_backends():
 def hf_token_state():
     """Return the 3-source HF token cascade state for the Settings UI
     (Wave 2 React panel consumes this). Never returns the raw token —
-    only a masked preview, whoami username, and per-source validity.
+    only a masked preview and local presence; no outbound validation.
     """
     from dataclasses import asdict
     from services import token_resolver

backend/api/routers/tts_stream.py

@@ -174,11 +174,14 @@ async def ws_tts(websocket: WebSocket):
                 # close on `unavailable`, a one-time `routing` frame on
                 # cpu_fallback / accelerated-with-caveat (before any audio).
                 from core.device_caps import detect_host_caps
-                from services.engine_routing import resolve_routing, routing_notice
+                from services.engine_routing import (
+                    routing_notice,
+                    runtime_compute_profile_async,
+                )
                 from core.scrub import scrub_text
-                _routing = resolve_routing(
-                    getattr(backend, "gpu_compat", ("cpu",)), detect_host_caps(),
-                    getattr(backend, "min_vram_gb", 0.0))
+                _routing = await runtime_compute_profile_async(
+                    backend, detect_host_caps()
+                )
                 if _routing["routing_status"] == "unavailable":
                     await websocket.send_json({
                         "type": "error",

backend/api/routers/voice_convert.py

@@ -307,6 +307,16 @@ async def convert_speech(
             GpuPoolBusyError,
             run_on_gpu_pool_guarded,
         )
+        from core.device_caps import detect_host_caps
+        from services.engine_routing import runtime_compute_profile_async
+        compute_profile = await runtime_compute_profile_async(
+            backend, detect_host_caps()
+        )
+        if compute_profile["routing_status"] == "unavailable":
+            raise HTTPException(
+                status_code=400,
+                detail=compute_profile["routing_reason"],
+            )
 
         start_time = time.time()
         _render = functools.partial(
@@ -323,8 +333,14 @@ async def convert_speech(
             audio_tensor = await run_on_gpu_pool_guarded(
                 _render,
                 what="Voice convert",
-                timeout=_generate_timeout_s(text),
-                min_vram_gb=getattr(type(backend), "min_vram_gb", 0.0),
+                timeout=_generate_timeout_s(
+                    text,
+                    execution_device=compute_profile["effective_device"],
+                    min_vram_gb=compute_profile["min_vram_gb"],
+                    hardware_family=compute_profile.get("runtime_hardware_family"),
+                    vram_gb=compute_profile.get("runtime_vram_gb"),
+                ),
+                min_vram_gb=compute_profile["min_vram_gb"],
             )
         except GpuPoolBusyError as e:
             raise HTTPException(

backend/config/models.yaml

@@ -28,6 +28,9 @@
 #                            their own (weights live in referenced sub-repos). Such
 #                            a cache is legitimately tiny, so the truncated-download
 #                            (weights-missing) detector must NOT flag it incomplete.
+#   allow_patterns (optional) — restrict installation to these repository paths.
+#                               Use for multi-package repos so an explicit install
+#                               never downloads unrelated model variants.
 # ─────────────────────────────────────────────────────────────────────────
 
 models:
@@ -40,6 +43,14 @@ models:
     required: true
     curated_on: [all]
 
+  - repo_id: "audio-cpp/audio.cpp-gguf"
+    label: "Breeze-TTS-2 Q8_0 for audio.cpp (English + Chinese, clone + design)"
+    role: TTS
+    size_gb: 4.73
+    allow_patterns:
+      - "Breeze-TTS-2-GGUF/breeze-tts-2-q8_0.gguf"
+    note: "Optional audio.cpp model. Research/non-commercial weights and self-hosted outputs; install only after reviewing the license."
+
   # ── ASR (optional — curated per platform) ─────────────────────────────
   # No ASR model is required to boot: TTS-only installs work. Dubbing,
   # dictation, and clone-reference transcription prompt for the curated

backend/core/config.py

@@ -15,6 +15,18 @@ def get_app_data_dir():
         return os.path.expanduser("~/.omnivoice")
 
 
+def _configured_hf_token_path():
+    """Match Hub's token location without importing or refreshing credentials."""
+    default_cache = os.path.join(os.path.expanduser("~"), ".cache")
+    hf_home = os.environ.get("HF_HOME", os.path.join(os.environ.get("XDG_CACHE_HOME", default_cache), "huggingface"))
+    return os.path.expandvars(os.path.expanduser(os.environ.get("HF_TOKEN_PATH", os.path.join(hf_home, "token"))))
+
+
+# Snapshot recognized locations before automatic model-cache redirection.
+# Explicit cache/token overrides restrict clearing to their selected location.
+HF_CLI_TOKEN_PATHS = (_configured_hf_token_path(),)
+
+
 def _ensure_short_hf_cache_on_windows():
     """Redirect HuggingFace cache to a short path on Windows.
 
@@ -38,6 +50,14 @@ def _ensure_short_hf_cache_on_windows():
         return
     short_cache = os.path.join(local_app, "OmniVoice", "hf_cache")
     os.makedirs(short_cache, exist_ok=True)
+    if "HF_TOKEN_PATH" not in os.environ:
+        global HF_CLI_TOKEN_PATHS
+        canonical = HF_CLI_TOKEN_PATHS[0]
+        legacy = os.path.join(short_cache, "token")
+        HF_CLI_TOKEN_PATHS = tuple(dict.fromkeys((canonical, legacy)))
+        # Keep existing app-written logins usable without copying credentials.
+        selected = canonical if os.path.exists(canonical) or not os.path.exists(legacy) else legacy
+        os.environ.setdefault("HF_TOKEN_PATH", selected)
     os.environ["HF_HOME"] = short_cache
     os.environ["HF_HUB_CACHE"] = short_cache
 

backend/core/db.py

@@ -274,7 +274,8 @@ _BASE_SCHEMA = """
         started_at REAL,
         finished_at REAL,
         lease_expires_at REAL,
-        grace_expires_at REAL
+        grace_expires_at REAL,
+        deadlines_json TEXT
     );
     CREATE INDEX IF NOT EXISTS idx_remote_attempts_task ON remote_task_attempts(task_id);
     CREATE INDEX IF NOT EXISTS idx_remote_attempts_worker ON remote_task_attempts(worker_id, state);

backend/core/device_caps.py

@@ -35,7 +35,8 @@ import sys
 from dataclasses import dataclass
 from typing import Literal
 
-DeviceFamily = Literal["cuda", "rocm", "mps", "xpu", "cpu"]
+DeviceFamily = Literal["cuda", "rocm", "mps", "xpu", "npu", "cpu"]
+ACCELERATOR_PRIORITY = ("cuda", "rocm", "xpu", "npu", "mps")
 
 # Stable substring stamped onto notes that represent a real kernel-launch risk
 # (arch/driver mismatch) — as opposed to advisory notes (multi-GPU, VRAM query
@@ -533,9 +534,13 @@ def _probe() -> HostCaps:
         # is the whole truth in that case (CodeRabbit, #1425).
         notes.extend(why_no_gpu(torch))
 
-    # ── Intel XPU via IPEX ───────────────────────────────────────────────
+    # Older builds register XPU through IPEX; modern torch exposes it directly.
     try:
         import intel_extension_for_pytorch  # noqa: F401
+    except Exception:
+        # Optional IPEX may be absent or incompatible; still probe native torch XPU.
+        pass
+    try:
         if hasattr(torch, "xpu") and torch.xpu.is_available():
             detected.append("xpu")
             if not device_name:
@@ -546,7 +551,23 @@ def _probe() -> HostCaps:
                     pass
             notes.append("XPU VRAM not queried (unreliable across IPEX versions)")
     except Exception:
-        # IPEX absent or XPU probe failed — no XPU on this host.
+        # XPU probe failed — no usable XPU on this host.
+        pass
+
+    # Vendor extensions may register an NPU with torch. Probe only an already
+    # registered backend; never install or import an optional vendor package.
+    try:
+        if hasattr(torch, "npu") and torch.npu.is_available():
+            detected.append("npu")
+            if not device_name:
+                try:
+                    device_name = torch.npu.get_device_name(0)
+                except Exception:
+                    # An unavailable display name does not invalidate a usable NPU.
+                    pass
+            notes.append("NPU VRAM not queried")
+    except Exception:
+        # Missing or broken vendor backends mean no usable NPU; continue probing.
         pass
 
     # ── Apple Silicon MPS ────────────────────────────────────────────────
@@ -579,7 +600,7 @@ def _probe() -> HostCaps:
 
     # Preferred family by priority; cpu when nothing accelerated was detected.
     family: DeviceFamily = "cpu"
-    for pref in ("cuda", "rocm", "xpu", "mps"):
+    for pref in ACCELERATOR_PRIORITY:
         if pref in detected:
             family = pref  # type: ignore[assignment]
             break

backend/core/diagnose.py

@@ -28,6 +28,7 @@ import shutil
 import sys
 
 from core.config import DATA_DIR
+from core.device_caps import KERNEL_RISK_MARKER
 from core.scrub import scrub_text
 from core.version import APP_VERSION
 
@@ -189,7 +190,7 @@ def _check_ram() -> dict:
 def _check_engines() -> dict:
     try:
         from services.tts_backend import list_backends, active_backend_id
-        backends = list_backends()
+        backends = list_backends(include_hidden=True)
         active = active_backend_id()
     except Exception as e:
         return _check("engines", "TTS engines", WARN, f"could not enumerate: {e}")
@@ -230,11 +231,16 @@ def _check_gpu_routing() -> dict:
     host = v.get("host_family", "cpu")
 
     if status == "accelerated":
-        if reason:  # driver/arch caveat — accelerated but at risk
+        if reason and KERNEL_RISK_MARKER in reason:  # driver/arch caveat — at risk
             return _check("gpu_routing", "GPU routing", WARN,
                           f"{engine} -> {dev}: {reason}",
                           "The GPU is selected but may fail at kernel launch — "
                           "update drivers / reinstall torch for this GPU arch.")
+        if reason:  # low-VRAM caveat — not a driver/arch issue
+            return _check("gpu_routing", "GPU routing", WARN,
+                          f"{engine} -> {dev}: {reason}",
+                          "Unload other models before generating, keep the text "
+                          "short, or pick a lighter engine.")
         return _check("gpu_routing", "GPU routing", OK, f"{engine} -> {dev} (accelerated)")
     if status == "cpu_fallback":
         return _check("gpu_routing", "GPU routing", WARN,
@@ -374,7 +380,12 @@ def run_diagnostics(include_network: bool = True, deep: bool = False) -> dict:
         try:
             module = importlib.import_module(f"services.{family}_backend")
             active = module.active_backend_id()
-            row = next((item for item in module.list_backends() if item.get("id") == active), None)
+            rows = (
+                module.list_backends(include_hidden=True)
+                if family == "tts"
+                else module.list_backends()
+            )
+            row = next((item for item in rows if item.get("id") == active), None)
             if row is not None:
                 engine_execution.append({
                     "family": family,

backend/core/failure.py

@@ -97,6 +97,9 @@ _HINTS: dict[str, str] = {
     "TRANSFORMERS_IMPORT": "Your transformers install is incomplete, or a package it loads models through (torchaudio, torchvision) is missing or mismatched with your torch — a torch/torchvision version mismatch fails with exactly this wording. Reinstall them together at the pinned versions (`uv pip install --python .venv --reinstall torch==2.8.0 torchaudio==2.8.0 torchvision==0.23.0 transformers` in the project folder), then restart the backend. If only transcription is affected, switching ASR to faster-whisper (Model Catalogue → Models) also works around it.",
     "WINDOWS_APP_CONTROL_BLOCKED": "Windows refused to load a file VoiceStudio needs — an Application Control policy (Smart App Control, WDAC, or AppLocker) blocked it. On a personal PC: Windows Security → App & browser control → Smart App Control → Off (Windows only lets you turn it off once — re-enabling requires a Windows reset), then restart VoiceStudio. On a managed/work PC, ask IT to allow the VoiceStudio install folder.",
     "WINDOWS_PAGING_FILE_TOO_SMALL": "Windows ran out of virtual memory while mapping the model into memory — its paging file is smaller than the model needs. This is not the same as your RAM being full, and closing other apps usually won't fix it: Windows has to be allowed to back the mapping. Set a bigger paging file — Settings → System → About → Advanced system settings → Performance → Settings → Advanced → Virtual memory → Change: untick \"Automatically manage\", pick your system drive, choose \"Custom size\" and set both Initial and Maximum to at least 32768 MB (more than the model's size), then OK and restart Windows. A smaller/quantized engine (OmniVoice GGUF, Supertonic-3) also avoids the large mapping entirely.",
+    "WINDOWS_UNTRUSTED_MOUNT": "Windows refused to walk a folder on the way to this file because the path crosses a mount point it does not trust (WinError 448). That is a Windows rule about the VOLUME, not about VoiceStudio or the file itself — it turns up on Dev Drives, on mounted VHD/ReFS volumes, and on junctions pointing into another user profile, so retrying the same link cannot help. Point VoiceStudio at a folder on an ordinary local drive instead: Settings → Storage → data directory, or the download/output folder named in the message. If that folder has to stay where it is, trust the volume with `fsutil devdrv trust <drive>:` from an elevated prompt and restart.",
+    "INPUT_TOO_SHORT": "The input was too short for this engine to process — its first convolution needs more frames than the text (or the reference clip) produced. This is a hard limit of the model, not a transient failure, so retrying the same input will fail the same way. Give it a few more words, or a longer reference clip: a short phrase rather than one or two characters, and about a second of speech rather than a fragment.",
+    "CLONE_REFERENCE_MISSING": "This engine was asked to clone a voice but got no reference audio to clone FROM, and the model folder carries no built-in voice either. Pick a voice profile that has a saved reference clip, or record/upload a few seconds of clean speech as the reference, then generate again. A designed voice with no saved reference cannot be cloned from — synthesize with it directly instead.",
     "MEDIA_TOOL_MISSING": "VoiceStudio's media engine (ffmpeg/ffprobe) wasn't on the system path when a component went looking for it. Open Settings → Audio tools and use Download/Repair to fetch the bundled copy, then retry — a restart picks it up for everything. If you'd rather use a system install, install ffmpeg (macOS: `brew install ffmpeg`; Windows: `winget install Gyan.FFmpeg`; Linux: your package manager) and restart VoiceStudio, or point FFMPEG_PATH / OMNIVOICE_FFPROBE_PATH at the binaries in Settings.",
     "AUDIO_IO_FAILED": "An audio file couldn't be read or written at the OS level. Check the drive isn't full, that the output and temp folders exist and are writable, and that antivirus or OneDrive isn't locking them (add a VoiceStudio exclusion if you use one).",
     "VIDEO_DOWNLOAD_OS_ERROR": "The OS refused a file operation while saving the downloaded video — this is a disk/folder problem, not a network one, so retrying the same link won't help. The download is written to a job folder under your VoiceStudio data directory (Settings → Storage shows the path): check that drive isn't full, that the folder exists and is writable, and that antivirus or a cloud-sync client (OneDrive, Dropbox) isn't locking it — add a VoiceStudio exclusion if you use one. If your data directory sits on a synced or network drive, move it to a local one.",
@@ -308,6 +311,23 @@ _CONTEXT_FREE_HINT_CLASSES = frozenset({
     # a Windows virtual-memory setting rather than a connectivity problem, and
     # the detailed hint we already had for it never reached them.
     "WINDOWS_PAGING_FILE_TOO_SMALL",
+    # #1957: triggered by WinError 448 or the literal "untrusted mount
+    # point" — both unmistakable, and it reaches the user as a bare
+    # download failure with only the OS sentence attached.
+    "WINDOWS_UNTRUSTED_MOUNT",
+    # #1826: torch's own conv wording, which nothing else produces, and it
+    # reaches the user through the generic 500.
+    "INPUT_TOO_SHORT",
+    # #1879: matched on wording no other failure produces, and it reaches the
+    # user as a bare 400 carrying only the library sentence.
+    "CLONE_REFERENCE_MISSING",
+    # Its trigger is a VoiceStudio-authored sentence — "the TTS model cache
+    # for … is incomplete" plus "could not be auto-repaired" / "weights
+    # missing" — so it cannot be produced by an unrelated library. The 500
+    # handler is the surface a corrupt cache actually reaches, and dropping
+    # its hint there would leave the user with no way to know a redownload
+    # is the fix.
+    "MODEL_CACHE_CORRUPT",
 })
 
 
@@ -569,6 +589,34 @@ def classify(reason: str) -> str:
         or "application control policy" in low
     ):
         return "WINDOWS_APP_CONTROL_BLOCKED"
+    # #1957: the path to a download or output file crosses a mount point
+    # Windows will not traverse (Dev Drive, mounted VHD/ReFS, a junction into
+    # another profile). Matched on the numeric code first because the OS
+    # translates the sentence, with the English phrase as a fallback.
+    if "[winerror 448]" in low or "untrusted mount point" in low:
+        return "WINDOWS_UNTRUSTED_MOUNT"
+    # #1826: a degenerate-length input reaches a conv layer whose kernel is
+    # wider than the tensor, and torch says so in its own terms — "Calculated
+    # padded input size per channel: (1). Kernel size: (2). Kernel size can't
+    # be greater than actual input size". That arrived doubly wrapped in
+    # "Underlying error:" and told the user nothing they could act on, when
+    # the fix is simply "type more than one character".
+    if "kernel size can't be greater than actual input size" in low or (
+        "calculated padded input size per channel" in low
+    ):
+        return "INPUT_TOO_SHORT"
+    # #1879: mlx-audio (and the Chatterbox-family models under it) raise a
+    # bare ValueError naming their own parameters — "No conditionals
+    # available. Either provide audio_prompt/audio_prompt_sr ... or ensure
+    # conds.safetensors is in the model directory." The generate route passed
+    # that straight through as the 400 detail, so the user was told to supply
+    # an argument they have no way to name and to check for a file they have
+    # never heard of. What actually happened is "you asked to clone without a
+    # reference clip".
+    if "no conditionals available" in low or (
+        "audio_prompt" in low and "conds.safetensors" in low
+    ):
+        return "CLONE_REFERENCE_MISSING"
     # #1221: libsndfile failed an OS-level audio read/write. Its own wording is
     # a bare "System error.", so match the library name — audio_io already
     # prefixes the target path and free space onto the write-path failures.

backend/core/parent_liveness.py

@@ -4,7 +4,14 @@ from __future__ import annotations
 import os
 import sys
 import threading
-from typing import BinaryIO, Callable
+import time
+from typing import Any, BinaryIO, Callable, Optional
+
+# Poll cadence for the Windows pipe watcher. Exit latency after the desktop
+# closes its end is bounded by this; the desktop's own kill-on-close Job is the
+# hard backstop, so a quarter second is plenty and costs nothing measurable.
+WINDOWS_PIPE_POLL_INTERVAL_S = 0.25
+_FILE_TYPE_PIPE = 3  # winbase.h FILE_TYPE_PIPE
 
 
 def _watch_parent_pipe(reader: BinaryIO, exit_process: Callable[[int], None]) -> None:
@@ -18,6 +25,64 @@ def _watch_parent_pipe(reader: BinaryIO, exit_process: Callable[[int], None]) ->
     exit_process(0)
 
 
+def _watch_parent_pipe_handle(
+    handle: int,
+    exit_process: Callable[[int], None],
+    *,
+    peek: Optional[Callable[[int], Any]] = None,
+    read_file: Optional[Callable[[int, int], Any]] = None,
+    sleep: Callable[[float], None] = time.sleep,
+    interval: float = WINDOWS_PIPE_POLL_INTERVAL_S,
+) -> None:
+    """Windows twin of :func:`_watch_parent_pipe` that never leaves a read
+    pending on the pipe.
+
+    A synchronous ``ReadFile`` parked on the stdin pipe — whether issued through
+    the C runtime's ``read()`` or straight to the kernel — deadlocks the
+    OpenBLAS DLL initializer that ``import torch`` reaches (numpy's
+    ``_multiarray_umath``) in the startup worker: every desktop-spawned backend
+    on Windows froze at "Loading ML runtime (PyTorch)" while the identical
+    command from a terminal, with no stdin pipe and no watchdog, started in
+    seconds. A thread that merely sleeps does not trigger it; only the pending
+    read on that pipe does. So instead of blocking in a read, poll with
+    ``PeekNamedPipe``: it returns immediately, holds no I/O on the file object,
+    drains any keepalive bytes the desktop might write, and fails with
+    ``ERROR_BROKEN_PIPE`` the moment the desktop closes its end — which is the
+    same EOF signal the POSIX reader gets.
+    """
+    if peek is None or read_file is None:
+        import _winapi  # Windows-only stdlib module; the caller gates on the platform
+
+        peek = peek or _winapi.PeekNamedPipe
+        read_file = read_file or _winapi.ReadFile
+    try:
+        while True:
+            available, _ = peek(handle)
+            if available:
+                # Bytes are already buffered, so this read cannot block.
+                read_file(handle, available)
+            else:
+                sleep(interval)
+    except OSError:
+        # ERROR_BROKEN_PIPE (109) is how the closed parent end surfaces here.
+        pass
+    exit_process(0)
+
+
+def _windows_pipe_handle(reader: Any) -> Optional[int]:
+    """The OS handle behind ``reader`` when it is a pipe, else None."""
+    try:
+        import msvcrt
+        import _winapi
+
+        handle = msvcrt.get_osfhandle(reader.fileno())
+        if _winapi.GetFileType(handle) != _FILE_TYPE_PIPE:
+            return None
+        return handle
+    except (OSError, ValueError, AttributeError, ImportError):
+        return None
+
+
 def arm_desktop_parent_watchdog() -> bool:
     """Use stdin EOF as an unforgeable parent-liveness signal for desktop runs."""
     if os.environ.get("OMNIVOICE_DESKTOP_CONTAINED") != "1":
@@ -25,9 +90,18 @@ def arm_desktop_parent_watchdog() -> bool:
     reader = getattr(sys.stdin, "buffer", None)
     if reader is None:
         return False
+    target: Callable[..., None] = _watch_parent_pipe
+    args: tuple = (reader, os._exit)
+    if os.name == "nt":
+        handle = _windows_pipe_handle(reader)
+        if handle is not None:
+            target = _watch_parent_pipe_handle
+            args = (handle, os._exit)
+        # A non-pipe stdin (file, NUL) cannot have a read pending against a
+        # pipe file object, so the blocking reader stays correct there.
     threading.Thread(
-        target=_watch_parent_pipe,
-        args=(reader, os._exit),
+        target=target,
+        args=args,
         name="desktop-parent-watchdog",
         daemon=True,
     ).start()

backend/core/public_errors.py

@@ -94,6 +94,16 @@ def stream_generation_failure(error: BaseException | object) -> dict[str, object
     replace the failure being diagnosed.
     """
     payload = stream_failure("generation_failed")
+    if isinstance(error, BaseException):
+        # The exception's TYPE NAME, never its message. Two failures that both
+        # render the floor message "Generation failed. Check the selected
+        # engine and try again." are indistinguishable in an auto-filed report,
+        # so every unclassified streaming failure arrives as the same issue and
+        # none of them can be triaged (#1800). A class name is VoiceStudio-safe
+        # by the same reasoning that already puts it on the wire as
+        # `error_class` in the dub routes and on the analytics allowlist: it is
+        # a Python type, not user text, and no substring of `error` is copied.
+        payload["error_class"] = type(error).__name__
     try:
         enriched = public_exception_response(error, fallback=str(payload["detail"]))
     except Exception:
@@ -149,12 +159,31 @@ def public_exception_response(error: BaseException, *, fallback: str) -> dict[st
     Classification may inspect the private diagnostic locally, but response
     values come exclusively from VoiceStudio-owned constants. No substring of
     ``error`` is copied into the payload.
+
+    Every caller is a CONTEXT-FREE surface — the global 500 handler, the
+    streaming generate error frame, the dub GPU-OOM 503 — so the topic is
+    filtered through ``failure._CONTEXT_FREE_HINT_CLASSES`` before its hint is
+    attached. Without that filter a topic whose trigger is a generic phrase
+    stamps a confidently wrong remediation on an unrelated failure: #1943 is a
+    macOS mlx-audio TTS 500 that came back advising the user that "the
+    connection to the video server dropped mid-download", because
+    VIDEO_DOWNLOAD_NETWORK triggers on a bare "timed out" / "connection
+    reset". The allowlist already existed and already named that class as the
+    example of what must not appear here; only :func:`failure.append_hint`
+    honoured it, and this helper replaced ``append_hint`` on the 500 path
+    without carrying the rule across.
+
+    HF_MIRROR_UNREACHABLE is allowed alongside it: its hint is dynamic (it
+    names the configured mirror) and its trigger requires that a mirror is
+    configured at all, so it cannot fire on an unrelated failure (#874).
     """
-    from core.failure import classify, public_hint_for_topic
+    from core.failure import _CONTEXT_FREE_HINT_CLASSES, classify, public_hint_for_topic
 
     try:
         topic = classify(str(error))
-        hint = public_hint_for_topic(topic)
+        if topic and topic not in _CONTEXT_FREE_HINT_CLASSES and topic != "HF_MIRROR_UNREACHABLE":
+            topic = ""
+        hint = public_hint_for_topic(topic) if topic else ""
     except Exception:
         topic = ""
         hint = ""

backend/core/torch_indexes.py

@@ -0,0 +1,43 @@
+"""The PyTorch wheel index VoiceStudio installs CUDA builds from.
+
+A local-version pin such as ``torch==2.9.1+cu128`` exists only on PyTorch's
+own index, never on PyPI. The app's own ``pyproject.toml`` routes torch there
+through ``[tool.uv.sources]``, but a sidecar engine is installed with
+``uv pip install`` into its own venv, which knows nothing about that config —
+so every CUDA-pinned sidecar install has to name the index itself.
+
+MOSS-TTS-v1.5's install did not, and its ``[torch-runtime]`` extra
+(``torch==2.9.1+cu128``) could never resolve: ``uv pip compile`` reports it
+unsatisfiable without this index and resolves it with it. One definition here,
+imported by the one-click installer and by the engine's own bootstrap, so the
+two cannot drift apart again. ``tests/test_sidecar_install.py`` pins the URL
+to the ``pytorch-cuda`` index declared in the app's ``pyproject.toml``.
+"""
+
+PYTORCH_CU128_INDEX_URL = "https://download.pytorch.org/whl/cu128"
+
+# `unsafe-best-match`: the PyTorch index also mirrors common dependencies
+# (numpy, pillow, sympy, …) at a narrower range of versions than PyPI. uv's
+# default first-index strategy would stop at whichever index lists a name first
+# and could pin an old mirror copy or fail outright. The index is PyTorch's
+# official one, so the dependency-confusion risk the name warns about does not
+# apply to it.
+UV_PIP_CU128_ARGS: tuple[str, ...] = (
+    "--extra-index-url",
+    PYTORCH_CU128_INDEX_URL,
+    "--index-strategy",
+    "unsafe-best-match",
+)
+
+PYTORCH_CPU_INDEX_URL = "https://download.pytorch.org/whl/cpu"
+
+# For an engine that runs torch only on the CPU (PocketTTS). On Linux, PyPI's
+# torch is the CUDA build and pulls ~15 NVIDIA packages the engine never uses;
+# this index serves `+cpu` builds for Linux and Windows and the regular build
+# for macOS.
+UV_PIP_CPU_ARGS: tuple[str, ...] = (
+    "--extra-index-url",
+    PYTORCH_CPU_INDEX_URL,
+    "--index-strategy",
+    "unsafe-best-match",
+)

backend/engines/audiocpp/__init__.py

@@ -0,0 +1,590 @@
+"""audio.cpp TTS backend — Breeze-TTS-2 via a managed native server.
+
+audio.cpp (0xShug0/audio.cpp) is a pure-C++ ggml runtime: prebuilt
+``audiocpp_server`` binaries for Windows/macOS/Linux, no Python venv, no
+``transformers`` pin — so this engine needs neither the venv-isolation
+(``engines.dots_tts``) nor the per-generate CLI-spawn (``engines
+.omnivoice_gguf``) patterns. The parent instead:
+
+1. resolves the binary + GGUF model (``bootstrap.py``),
+2. spawns ONE long-lived ``audiocpp_server`` on 127.0.0.1 (lazy model load,
+   so model memory is only held after the first generate), and
+3. speaks its OpenAI-style ``POST /v1/audio/speech`` per generate.
+
+v1 serves the ``breeze_tts`` family only (Breeze-TTS-2, en+zh, voice clone
++ voice design + voice direction). The server is task-agnostic on the
+speech route — reference-audio presence selects clone/direction vs design —
+so a single ``task: tts`` model entry covers all three modes.
+
+License honesty: Breeze-TTS-2 weights (``BreezeBlue/Breeze-TTS-2`` and the
+audio.cpp GGUF repack) are RESEARCH AND NON-COMMERCIAL ONLY
+(``BreezeBlue Research and Non-Commercial License``); only the audio.cpp
+code is Apache-2.0. There is no in-tree acceptance dialog for this engine
+yet (settings ``/license`` allow-list), so the restriction is surfaced in
+the display name, the install hint, and ``docs/engines/audio-cpp.md`` —
+not silently.
+"""
+from __future__ import annotations
+
+import atexit
+import base64
+import io
+import json
+import logging
+import os
+import secrets
+
+# Used only for stream constants; spawn_owned performs the process launch.
+import subprocess  # nosec B404
+import threading
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+from core.contained_subprocess import spawn_owned
+from services.tts_backend import TTSBackend, TTSInputError
+
+if TYPE_CHECKING:
+    import torch
+
+logger = logging.getLogger("omnivoice.audiocpp")
+
+#: Engine id in the TTS registry.
+ENGINE_ID = "audiocpp"
+
+#: How long to wait for ``/health`` after spawning the server (first spawn
+#: extracts nothing heavy — the model loads lazily on first generate).
+_HEALTH_TIMEOUT_S = 120.0
+
+#: Finish the inner HTTP request before the canonical generation guard can
+#: abandon its worker thread. This leaves enough time to terminate the owned
+#: native process and release its model memory synchronously.
+_TERMINATE_GRACE_S = 5.0
+_TERMINATE_KILL_S = 5.0
+_GENERATE_TIMEOUT_MARGIN_S = (
+    _TERMINATE_GRACE_S + _TERMINATE_KILL_S + 5.0
+)
+
+
+# ── pure request/config builders (unit-tested, no I/O) ──────────────────────
+
+
+def _cpu_thread_count() -> int:
+    """Use up to 16 physical cores, with a stdlib fallback."""
+    try:
+        import psutil
+
+        cores = psutil.cpu_count(logical=False)
+    except (ImportError, OSError):
+        cores = None
+    return min(16, max(1, cores or os.cpu_count() or 1))
+
+
+def _device_min_vram_gb(device) -> float:
+    """Dedicated-memory comfort floor for one discovered native device."""
+    return 6.0 if (
+        device
+        and device.kind == "GPU"
+        and (
+            device.backend == "vulkan"
+            or device.hardware_family in {"cuda", "rocm"}
+        )
+    ) else 0.0
+
+
+def build_server_config(
+    *, model_id: str, family: str, model_path: str, port: int,
+    backend: str = "cpu", device: int = 0,
+    execution_target: str | None = None,
+) -> dict:
+    """``server.json`` dict for the managed ``audiocpp_server``.
+
+    ``lazy_load`` defers the ~4.73 GiB GGUF load to the first generate;
+    ``max_loaded_models: 1`` bounds residency to the one model we serve.
+    """
+    return {
+        "host": "127.0.0.1",
+        "port": port,
+        "backend": backend,
+        "device": device,
+        # The pinned CPU runtime scales strongly through 16 workers while
+        # producing byte-identical audio.
+        "threads": _cpu_thread_count()
+        if (execution_target or backend) == "cpu" else 1,
+        "lazy_load": True,
+        "max_loaded_models": 1,
+        "models": [
+            {
+                "id": model_id,
+                "family": family,
+                "path": model_path,
+                "task": "tts",
+                "mode": "offline",
+            }
+        ],
+    }
+
+
+def build_speech_payload(
+    *, model_id: str, text: str, ref_audio: str | None = None,
+    ref_text: str | None = None, instructions: str | None = None,
+    guidance_scale: float | None = None, seed: int | None = None,
+) -> dict:
+    """``POST /v1/audio/speech`` JSON body.
+
+    Field spellings verified against ``app/server/runtime.cpp``
+    (``build_speech_request``): ``instructions`` (plural, OpenAI spelling)
+    feeds the ``instruction`` request option; ``reference_text`` and
+    ``guidance_scale``/``seed`` pass through top-level; ``voice_ref`` takes
+    a ``{"type": "path", ...}`` object so the reference stays on disk
+    (the 5 MiB base64 cap never bites). ``response_format: json`` returns
+    the WAV base64-in-JSON — one round trip, no binary framing.
+    """
+    payload: dict[str, Any] = {
+        "model": model_id,
+        "input": text,
+        "response_format": "json",
+    }
+    if instructions:
+        payload["instructions"] = instructions
+    if ref_audio:
+        payload["voice_ref"] = {"type": "path", "path": str(ref_audio)}
+        if ref_text:
+            payload["reference_text"] = ref_text
+    if guidance_scale is not None:
+        payload["guidance_scale"] = float(guidance_scale)
+    if seed is not None:
+        payload["seed"] = int(seed)
+    return payload
+
+
+def decode_speech_json(obj: dict) -> tuple[int, object]:
+    """``(sample_rate, mono float32 numpy)`` from a ``response_format=json``
+    speech body. Raises ``ValueError`` on a server error payload."""
+    if not isinstance(obj, dict):
+        raise TypeError(f"audio.cpp speech reply is not JSON: {obj!r:.120}")
+    if "audio" not in obj:
+        raise ValueError(f"audio.cpp speech failed: {obj.get('error', obj)!r:.300}")
+    import numpy as np
+    import soundfile as sf
+
+    wav_bytes = base64.b64decode(obj["audio"])
+    wav, sr = sf.read(io.BytesIO(wav_bytes), dtype="float32", always_2d=False)
+    wav = np.asarray(wav, dtype=np.float32)
+    if wav.ndim > 1:
+        wav = wav.mean(axis=-1)
+    return int(sr), wav
+
+
+# ── backend ─────────────────────────────────────────────────────────────────
+
+
+class AudioCPPBackend(TTSBackend):
+    """Breeze-TTS-2 through a parent-managed ``audiocpp_server``."""
+
+    id = ENGINE_ID
+    display_name = (
+        "audio.cpp · Breeze-TTS-2 (native GGUF, en+zh, clone+design; "
+        "weights research/non-commercial)"
+    )
+    supports_voice_design = True
+    applies_own_mastering = True  # model-decoded 24 kHz studio output
+    gpu_compat = ("cpu",)
+    runs_out_of_process = True
+    # Same marker SubprocessBackend sets: this engine lives in another OS
+    # process. Consumers only branch the matrix label and the self-test
+    # route (spawn-and-ping instead of in-process synth) — both correct
+    # here; nothing assumes the stdio protocol from it.
+    _is_subprocess_isolated = True
+    _DEFAULT_SAMPLE_RATE = 24000  # Breeze-TTS-2 native rate
+
+    def __init__(self) -> None:
+        self._proc: Any | None = None
+        self._port: int | None = None
+        self._server_model_id: str | None = None
+        self._sr = self._DEFAULT_SAMPLE_RATE
+        self._lock = threading.RLock()
+        self._server_json: Path | None = None
+        self._selection = None
+        self._device = None
+        self._provider = None
+
+    # ── availability ────────────────────────────────────────────────────
+
+    @classmethod
+    def is_available(cls) -> tuple[bool, str]:
+        from engines.audiocpp import bootstrap
+
+        try:
+            bootstrap.resolve_server_binary()
+            bootstrap.resolve_model_file()
+        except RuntimeError as exc:
+            return False, str(exc)
+        return True, "ready"
+
+    @classmethod
+    def runtime_compute_profile(cls, caps) -> dict:
+        from dataclasses import replace
+
+        from engines.audiocpp import bootstrap
+        from services.engine_routing import low_vram_caveat
+
+        try:
+            selection = bootstrap.resolve_compute_selection(caps)
+            targets = bootstrap.runtime_targets()
+        except RuntimeError as exc:
+            return {
+                "gpu_compat": cls.gpu_compat,
+                "min_vram_gb": 0.0,
+                "effective_device": "cpu",
+                "routing_status": "unavailable",
+                "routing_reason": str(exc),
+                "runtime_backend": None,
+                "runtime_device_index": None,
+                "runtime_device_name": None,
+                "runtime_hardware_family": None,
+                "runtime_vram_gb": None,
+                "runtime_device_verified": False,
+            }
+        selected = selection.device
+        accelerated = selected.target != "cpu"
+        min_vram_gb = _device_min_vram_gb(selected)
+        dedicated = min_vram_gb > 0
+        reason = selection.fallback_reason
+        if accelerated and dedicated and reason is None:
+            selected_caps = replace(
+                caps,
+                device_name=selected.name,
+                vram_gb=selection.verified_vram_gb,
+            )
+            reason = low_vram_caveat(
+                selected_caps,
+                min_vram_gb,
+                family=selected.hardware_family,
+                vram_gb=selection.verified_vram_gb,
+            )
+        status = "accelerated" if accelerated else (
+            "cpu_fallback" if selection.fallback_reason else "cpu_only"
+        )
+        return {
+            "gpu_compat": targets,
+            "min_vram_gb": min_vram_gb,
+            "effective_device": selected.target,
+            "routing_status": status,
+            "routing_reason": reason,
+            "runtime_backend": selected.backend,
+            "runtime_device_index": selected.index,
+            "runtime_device_name": selected.name,
+            "runtime_hardware_family": selected.hardware_family,
+            "runtime_vram_gb": selection.verified_vram_gb,
+            "runtime_device_verified": selection.verified_vram_gb > 0,
+        }
+
+    # ── TTSBackend protocol ─────────────────────────────────────────────
+
+    @property
+    def sample_rate(self) -> int:
+        return self._sr
+
+    @property
+    def supported_languages(self) -> list[str]:
+        return ["en", "zh"]
+
+    def model_identity(self) -> str | None:
+        from engines.audiocpp import bootstrap
+
+        return f"{bootstrap.FAMILY}/{bootstrap.package_filename()}"
+
+    # ── server lifecycle ────────────────────────────────────────────────
+
+    def _base_url(self) -> str:
+        return f"http://127.0.0.1:{self._port}"
+
+    def _ensure_loaded(self) -> None:
+        """Spawn the server (once) and wait for ``/health``. Idempotent."""
+        with self._lock:
+            if self._proc is not None and self._proc.poll() is None:
+                return
+            self._proc = None  # stale handle — respawn below
+            from engines.audiocpp import bootstrap
+
+            binary = bootstrap.resolve_server_binary()
+            selection = bootstrap.resolve_compute_selection()
+            model_file = bootstrap.resolve_model_file()
+            self._port = bootstrap.server_port()
+            # The random model id is a per-launch challenge. Before sending
+            # speech text or a reference path, _verify_server_identity asks
+            # /v1/models to prove this is the child configured by this process,
+            # not an unrelated listener that pre-bound the loopback port.
+            self._server_model_id = f"{bootstrap.MODEL_ID}-{secrets.token_hex(16)}"
+            config = build_server_config(
+                model_id=self._server_model_id,
+                family=bootstrap.FAMILY,
+                model_path=str(model_file),
+                port=self._port,
+                backend=selection.device.backend,
+                device=selection.device.index,
+                execution_target=selection.device.target,
+            )
+            self._selection = selection
+            self._device = selection.device.target
+            self._provider = selection.device.backend
+            from core.config import DATA_DIR
+
+            workdir = Path(str(DATA_DIR)) / "audiocpp"
+            workdir.mkdir(parents=True, exist_ok=True)
+            self._server_json = workdir / "server.json"
+            flags = os.O_WRONLY | os.O_CREAT | os.O_TRUNC
+            config_fd = os.open(self._server_json, flags, 0o600)
+            try:
+                if os.name != "nt":
+                    os.fchmod(config_fd, 0o600)
+                with os.fdopen(config_fd, "w", encoding="utf-8") as config_fh:
+                    config_fd = -1
+                    json.dump(config, config_fh, indent=2)
+            finally:
+                if config_fd >= 0:
+                    os.close(config_fd)
+            log_path = workdir / "server.log"
+            logger.info(
+                "audio.cpp: starting %s (backend=%s, device=%d, port=%d, model=%s)",
+                binary.name, selection.device.backend, selection.device.index,
+                self._port, model_file.name,
+            )
+            with open(log_path, "ab") as log_fh:
+                self._proc = spawn_owned(
+                    [str(binary), "--config", str(self._server_json)],
+                    stdout=log_fh,
+                    stderr=subprocess.STDOUT,
+                    stdin=subprocess.DEVNULL,
+                )
+            atexit.register(self._terminate_server)
+            self._wait_for_health()
+
+    def _wait_for_health(self) -> None:
+        if self._proc is None or self._port is None:
+            raise RuntimeError("managed audio.cpp server was not started")
+        deadline = time.monotonic() + _HEALTH_TIMEOUT_S
+        last_err = "unknown"
+        url = self._base_url() + "/health"
+        while time.monotonic() < deadline:
+            if self._proc.poll() is not None:
+                raise RuntimeError(
+                    "audiocpp_server exited during startup "
+                    f"(code {self._proc.returncode}). See the server log next "
+                    "to server.json under the app data audiocpp/ directory — "
+                    "the managed port may already be in use."
+                )
+            try:
+                # ``url`` is always the hard-coded loopback host plus a
+                # validated integer port; arbitrary schemes are impossible.
+                with urllib.request.urlopen(url, timeout=5) as resp:  # nosec B310
+                    if resp.status == 200:
+                        self._verify_server_identity()
+                        if self._proc.poll() is None:
+                            logger.info(
+                                "audio.cpp: managed server is healthy on loopback"
+                            )
+                            return
+                    last_err = f"HTTP {resp.status}"
+            except Exception as exc:  # noqa: BLE001 — still starting; retry
+                last_err = f"{type(exc).__name__}: {exc}"
+            time.sleep(1.0)
+        self._terminate_server()
+        raise RuntimeError(
+            f"audiocpp_server did not become healthy within "
+            f"{_HEALTH_TIMEOUT_S:.0f}s (last: {last_err})."
+        )
+
+    def _get_json(self, path: str, timeout: float = 5.0) -> dict:
+        """GET one loopback JSON endpoint without sending request content."""
+        if self._port is None:
+            raise RuntimeError("managed audio.cpp server port is missing")
+        req = urllib.request.Request(self._base_url() + path, method="GET")
+        with urllib.request.urlopen(req, timeout=timeout) as resp:  # nosec B310
+            obj = json.loads(resp.read().decode("utf-8"))
+        if not isinstance(obj, dict):
+            raise TypeError("audio.cpp returned an invalid JSON response")
+        return obj
+
+    def _verify_server_identity(self) -> None:
+        """Prove the loopback listener owns this launch's random model id."""
+        if self._proc is None or self._proc.poll() is not None:
+            raise RuntimeError("managed audio.cpp server is not running")
+        expected = self._server_model_id
+        if not expected:
+            raise RuntimeError("managed audio.cpp server identity is missing")
+        obj = self._get_json("/v1/models")
+        data = obj.get("data", [])
+        if not isinstance(data, list):
+            raise TypeError("managed audio.cpp server identity is invalid")
+        model_ids = {
+            item.get("id") for item in data
+            if isinstance(item, dict)
+        }
+        if expected not in model_ids or self._proc.poll() is not None:
+            raise RuntimeError(
+                "loopback listener did not prove managed audio.cpp ownership"
+            )
+
+    def _post_json(self, path: str, payload: dict, timeout: float) -> dict:
+        """Verify child ownership, then POST JSON to the managed server."""
+        if self._port is None:
+            raise RuntimeError("managed audio.cpp server port is missing")
+        self._verify_server_identity()
+        body = json.dumps(payload).encode("utf-8")
+        req = urllib.request.Request(
+            self._base_url() + path,
+            data=body,
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            # ``req`` targets only ``_base_url()`` (127.0.0.1 + validated
+            # integer port), never a caller-provided URL.
+            with urllib.request.urlopen(req, timeout=timeout) as resp:  # nosec B310
+                return json.loads(resp.read().decode("utf-8"))
+        except urllib.error.HTTPError as exc:
+            detail = exc.read().decode("utf-8", errors="replace")[:500]
+            raise RuntimeError(
+                f"audio.cpp {path} failed (HTTP {exc.code}): {detail}"
+            ) from exc
+        except urllib.error.URLError as exc:
+            if isinstance(exc.reason, TimeoutError):
+                raise TimeoutError("audio.cpp request timed out") from exc
+            raise
+
+    def _terminate_server(self) -> None:
+        proc, self._proc = self._proc, None
+        self._server_model_id = None
+        if proc is None:
+            return
+        try:
+            proc.terminate()
+            proc.wait(timeout=_TERMINATE_GRACE_S)
+        except Exception:  # noqa: BLE001 — kill as last resort, never raise
+            try:
+                proc.kill()
+                proc.wait(timeout=_TERMINATE_KILL_S)
+            except Exception as exc:  # noqa: BLE001 — process is already failing
+                logger.debug("audio.cpp: final server kill failed: %s", exc)
+
+    # ── generate ────────────────────────────────────────────────────────
+
+    def generate(self, text: str, **kw) -> torch.Tensor:
+        import torch
+        from services.model_manager import (
+            GENERATE_PROGRESS_GRACE_S,
+            generate_timeout_s,
+            report_generate_progress,
+        )
+
+        if not text or not text.strip():
+            raise TTSInputError(
+                "audio.cpp: the input contains no speakable text — "
+                "send at least one word."
+            )
+        ref_audio = kw.get("ref_audio")
+        ref_text = kw.get("ref_text")
+        if ref_text and not ref_audio:
+            logger.info(
+                "audio.cpp: ref_text supplied without ref_audio; ignoring."
+            )
+            ref_text = None
+
+        # Voice design: our `description=` (no ref) and voice direction
+        # (`instruct=` + ref) both ride the server's `instructions` field —
+        # verified spelling against app/server/runtime.cpp.
+        instruct = kw.get("instruct") or kw.get("description") or None
+
+        language = kw.get("language")
+        if language and str(language).strip().lower() not in {
+            "auto", "en", "english", "zh", "chinese",
+        }:
+            logger.info(
+                "audio.cpp (Breeze-TTS-2) is en+zh only; ignoring "
+                "language=%r.", language,
+            )
+        if kw.get("speed", 1.0) != 1.0:
+            logger.info("audio.cpp: speed is not supported; ignoring.")
+
+        request_started = time.monotonic()
+        with self._lock:
+            self._ensure_loaded()
+            selected = self._selection.device if self._selection else None
+            min_vram_gb = _device_min_vram_gb(selected)
+            request_budget = generate_timeout_s(
+                text,
+                execution_device=selected.target if selected else "cpu",
+                min_vram_gb=min_vram_gb,
+                hardware_family=selected.hardware_family if selected else None,
+                vram_gb=self._selection.verified_vram_gb
+                if self._selection else 0.0,
+            )
+            if not self._server_model_id:
+                raise RuntimeError("managed audio.cpp server identity is missing")
+            payload = build_speech_payload(
+                model_id=self._server_model_id,
+                text=text,
+                ref_audio=str(ref_audio) if ref_audio else None,
+                ref_text=ref_text,
+                instructions=instruct,
+                guidance_scale=kw.get("guidance_scale", 1.0),
+                seed=kw.get("seed"),
+            )
+            # Device discovery and server startup can consume part of the soft
+            # budget. This fresh synthesis lease gives the lazy model load and
+            # request a bounded window. The inner request always expires early
+            # enough to reap the owned server before the outer guard abandons us.
+            report_generate_progress()
+            soft_remaining = request_budget - (time.monotonic() - request_started)
+            timeout = (
+                max(soft_remaining, GENERATE_PROGRESS_GRACE_S)
+                - _GENERATE_TIMEOUT_MARGIN_S
+            )
+            if timeout <= 0:
+                self._terminate_server()
+                raise TimeoutError(
+                    "audio.cpp startup exhausted the generation time budget"
+                )
+            try:
+                obj = self._post_json(
+                    "/v1/audio/speech", payload, timeout=timeout,
+                )
+            except TimeoutError:
+                self._terminate_server()
+                raise RuntimeError(
+                    "audio.cpp generation timed out; its managed server was reset"
+                ) from None
+        sr, wav_np = decode_speech_json(obj)
+        self._sr = sr
+        wav = torch.from_numpy(wav_np).float()
+        if wav.ndim == 0:
+            raise RuntimeError("audio.cpp produced empty audio")
+        return wav.unsqueeze(0)
+
+    # ── lifecycle ───────────────────────────────────────────────────────
+
+    def unload(self) -> None:
+        """Free the model server-side, then stop it. Idempotent."""
+        with self._lock:
+            if self._port is not None and self._proc is not None \
+                    and self._proc.poll() is None:
+                try:
+                    self._post_json("/v1/tasks/unload_all_models", {}, timeout=30)
+                except Exception as exc:  # noqa: BLE001 — best effort
+                    logger.warning("audio.cpp: server unload failed: %s", exc)
+            self._port = None
+            self._terminate_server()
+        super().unload()
+
+
+__all__ = [
+    "ENGINE_ID",
+    "AudioCPPBackend",
+    "build_server_config",
+    "build_speech_payload",
+    "decode_speech_json",
+]

backend/engines/audiocpp/bootstrap.py

@@ -0,0 +1,738 @@
+"""audio.cpp binary probe + model resolution.
+
+audio.cpp (0xShug0/audio.cpp) is a pure-C++ ggml inference engine with
+prebuilt release binaries — no Python venv, no ``transformers`` pin, so
+none of the dependency-isolation machinery in ``engines._venv_probe`` or
+``services.subprocess_backend`` applies. The parent instead:
+
+1. locates a user-installed ``audiocpp_server`` (env var, user dir, or this
+   package's ``bin/``), and
+2. resolves an explicitly installed GGUF model file from a direct path or
+   the shared Hugging Face cache.
+
+Probe order for the server binary (existing installs win, zero migration):
+
+    1. ``${OMNIVOICE_AUDIOCPP_BIN}`` — absolute path to the binary itself.
+    2. ``${OMNIVOICE_AUDIOCPP_DIR}/audiocpp_server[.exe]`` — a user-managed
+       install dir (e.g. an extracted release zip, or a self-built tree).
+    3. ``backend/engines/audiocpp/bin/audiocpp_server[.exe]`` — an explicitly
+       installed local copy.
+
+``is_installed()`` is a cheap file-existence check — no spawn, no network.
+VoiceStudio never downloads executable code for this engine.
+"""
+from __future__ import annotations
+
+import errno
+import functools
+import logging
+import os
+import platform
+import subprocess  # nosec B404 -- fixed argv probes a user-selected executable
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+
+logger = logging.getLogger("omnivoice.audiocpp.bootstrap")
+
+#: Pinned audio.cpp release. BreezeTTS-2 support landed in 0.7.2 — older
+#: binaries have no ``breeze_tts`` family, so the floor is also the pin.
+VERSION = "v0.7.2"
+
+#: GitHub repo serving the prebuilt binaries.
+GH_REPO = "0xShug0/audio.cpp"
+
+#: HuggingFace repo serving the GGUF model packages (not gated).
+HF_MODEL_REPO = "audio-cpp/audio.cpp-gguf"
+
+# Immutable repository revision used for the v0.7.2 Breeze-TTS-2 package.
+# Pinning prevents a later upstream file replacement from silently changing
+# the model exercised by this backend.
+HF_MODEL_REVISION = "dc6fecccc2b0c6bdda0a8b2f38fa61394fee0b9c"
+
+#: Model id used in the generated ``server.json`` and in speech requests.
+MODEL_ID = "breeze-tts-2"
+
+#: audio.cpp family name for BreezeTTS 2 (``--family`` / server ``family``).
+FAMILY = "breeze_tts"
+
+#: GGUF package directory inside :data:`HF_MODEL_REPO`.
+PACKAGE_DIR = "Breeze-TTS-2-GGUF"
+
+#: Default package (Q8_0, the upstream-recommended GGUF). ``bf16`` is
+#: available via ``OMNIVOICE_AUDIOCPP_PACKAGE``.
+DEFAULT_PACKAGE = "breeze-tts-2-q8_0.gguf"
+
+#: Env var pointing directly at the ``audiocpp_server`` binary.
+BIN_ENV = "OMNIVOICE_AUDIOCPP_BIN"
+
+#: Env var pointing at a directory containing ``audiocpp_server``.
+DIR_ENV = "OMNIVOICE_AUDIOCPP_DIR"
+
+#: Env var overriding the GGUF package filename (e.g. the bf16 package).
+PACKAGE_ENV = "OMNIVOICE_AUDIOCPP_PACKAGE"
+
+#: Optional advanced overrides for a binary that exposes several runtimes or
+#: devices. Device indices are local to the selected backend registry.
+BACKEND_ENV = "OMNIVOICE_AUDIOCPP_BACKEND"
+DEVICE_ENV = "OMNIVOICE_AUDIOCPP_DEVICE"
+
+#: Env var overriding the loopback port the managed server binds.
+PORT_ENV = "OMNIVOICE_AUDIOCPP_PORT"
+
+#: Default loopback port. High and engine-specific to avoid clashing with
+#: the app itself or a user-run ``audiocpp_server`` (default 8080).
+DEFAULT_PORT = 17860
+
+#: This package's owned binary dir (probe 3).
+_PKG_BIN_DIR: Path = Path(__file__).parent / "bin"
+
+# Recommended (asset filename, sha256) per platform slug, from the v0.7.2
+# release. Windows and Linux use the vendor-neutral Vulkan build, which also
+# exposes the native CPU backend. Upstream publishes the macOS builds under
+# the Metal package name. No linux-aarch64 prebuilt exists in v0.7.2.
+_ASSETS: dict[str, tuple[str, str]] = {
+    "windows-x64": (
+        "audio-v0.7.2-bin-windows-x64-vulkan.zip",
+        "15b8232eae740e21e507d87f827a89966de9451b085a45932d9e214e032962c1",
+    ),
+    "linux-x64": (
+        "audio-v0.7.2-bin-ubuntu-x64-vulkan.tar.gz",
+        "fee1f978cee76453cf17f00196554bc2ee294645739538af0726a143b6a69a23",
+    ),
+    "darwin-arm64": (
+        "audio-v0.7.2-bin-macos-arm64-metal.tar.gz",
+        "c01e4f82971bedbe341697e63a9cebd5a5d1f72d5a9bcb51a3191f95ddab7a95",
+    ),
+    "darwin-x64": (
+        "audio-v0.7.2-bin-macos-x64-metal.tar.gz",
+        "3862270f33439077225324169313f727064f727305b54d8ce920244d75ddcc24",
+    ),
+}
+
+#: Binary filename per platform.
+_BINARY_NAMES = {"windows-x64": "audiocpp_server.exe"}
+
+_REGISTRY_BACKENDS = {
+    "CPU": "cpu",
+    "CUDA": "cuda",
+    "MUSA": "cuda",
+    "HIP": "hip",
+    "ROCm": "hip",
+    "Vulkan": "vulkan",
+    "Metal": "metal",
+    "MTL": "metal",
+}
+_BACKEND_ALIASES = {
+    "cpu": "cpu",
+    "cuda": "cuda",
+    "hip": "hip",
+    "rocm": "hip",
+    "vulkan": "vulkan",
+    "metal": "metal",
+}
+
+
+@dataclass(frozen=True)
+class AudioCPPDevice:
+    """One immutable device from audio.cpp's backend-local registry."""
+
+    registry: str
+    backend: str
+    index: int
+    name: str
+    kind: str
+    target: str
+    hardware_family: str
+
+
+@dataclass(frozen=True)
+class AudioCPPSelection:
+    """The runtime/device chosen for the next managed server."""
+
+    device: AudioCPPDevice
+    fallback_reason: str | None = None
+    verified_vram_gb: float = 0.0
+
+
+@dataclass(frozen=True)
+class _ProbeOutcome:
+    devices: tuple[AudioCPPDevice, ...] = ()
+    error: str | None = None
+
+
+def _cpu_probe_fallback(error: RuntimeError) -> AudioCPPSelection:
+    """A usable automatic fallback when native device discovery fails."""
+    return AudioCPPSelection(
+        AudioCPPDevice(
+            registry="CPU",
+            backend="cpu",
+            index=0,
+            name="Host CPU",
+            kind="CPU",
+            target="cpu",
+            hardware_family="cpu",
+        ),
+        f"{error}; running on CPU",
+    )
+
+
+def _vulkan_hardware_family(name: str) -> str:
+    low = name.casefold()
+    if any(token in low for token in ("nvidia", "geforce", "quadro", "tesla")):
+        return "cuda"
+    if any(token in low for token in ("amd", "radeon")):
+        return "rocm"
+    if any(token in low for token in ("intel", "arc ")):
+        return "xpu"
+    return "vulkan"
+
+
+def _device_families(registry: str, name: str, kind: str) -> tuple[str, str]:
+    # Software adapters such as Vulkan llvmpipe may be listed by a GPU
+    # registry but still execute on the CPU. Keep their runtime backend for
+    # explicit overrides while reporting and routing them as CPU work.
+    if kind == "CPU":
+        return "cpu", "cpu"
+    if registry in {"CUDA", "MUSA"}:
+        return "cuda", "cuda"
+    if registry in {"HIP", "ROCm"}:
+        return "rocm", "rocm"
+    if registry in {"Metal", "MTL"}:
+        return "mps", "mps"
+    if registry == "Vulkan":
+        return "vulkan", _vulkan_hardware_family(name)
+    return "cpu", "cpu"
+
+
+def parse_device_list(output: str) -> tuple[AudioCPPDevice, ...]:
+    """Parse the stable stdout contract of ``--list-devices``.
+
+    Backend diagnostics are emitted on stderr and deliberately never enter
+    this parser. Unknown future registries are ignored; malformed entries for
+    a registry we understand fail closed instead of selecting the wrong GPU.
+    """
+    devices: list[AudioCPPDevice] = []
+    seen: set[tuple[str, int]] = set()
+    for raw in str(output or "").splitlines():
+        line = raw.strip()
+        registry, colon, detail = line.partition(":")
+        if not colon or registry not in _REGISTRY_BACKENDS:
+            continue
+        index_text, space, remainder = detail.strip().partition(" ")
+        if not space or not index_text.isascii() or not index_text.isdecimal():
+            raise RuntimeError(
+                f"malformed audio.cpp {registry} device entry"
+            )
+        index = int(index_text)
+        remainder = remainder.strip()
+        kind_start = remainder.rfind("[")
+        if kind_start < 0 or not remainder.endswith("]"):
+            raise RuntimeError(
+                f"malformed audio.cpp {registry} device entry"
+            )
+        name_field = remainder[:kind_start].strip()
+        if name_field:
+            if len(name_field) < 2 or name_field[0] != '"' or name_field[-1] != '"':
+                raise RuntimeError(
+                    f"malformed audio.cpp {registry} device entry"
+                )
+            name = name_field[1:-1]
+        else:
+            name = ""
+        kind = remainder[kind_start + 1:-1].strip().upper()
+        if kind not in {"CPU", "GPU", "IGPU", "ACCEL", "META"}:
+            raise RuntimeError("unknown audio.cpp device kind")
+        # Registry aliases such as HIP/ROCm share one backend-local index
+        # namespace and therefore cannot safely describe different devices.
+        key = (_REGISTRY_BACKENDS[registry], index)
+        if key in seen:
+            raise RuntimeError(
+                f"duplicate audio.cpp device entry: {registry}:{index}"
+            )
+        seen.add(key)
+        target, hardware_family = _device_families(registry, name, kind)
+        devices.append(AudioCPPDevice(
+            registry=registry,
+            backend=_REGISTRY_BACKENDS[registry],
+            index=index,
+            name=name,
+            kind=kind,
+            target=target,
+            hardware_family=hardware_family,
+        ))
+    if not devices:
+        raise RuntimeError("audio.cpp reported no recognized compute devices")
+    return tuple(devices)
+
+
+def _platform_slug() -> str:
+    system = sys.platform
+    machine = platform.machine().lower()
+    if system == "win32":
+        return "windows-x64"
+    if system == "darwin":
+        return "darwin-arm64" if machine in ("arm64", "aarch64") else "darwin-x64"
+    if machine in ("x86_64", "amd64"):
+        return "linux-x64"
+    return f"linux-{machine}"
+
+
+def binary_name(slug: str | None = None) -> str:
+    """``audiocpp_server`` filename for ``slug`` (``.exe`` on Windows)."""
+    return _BINARY_NAMES.get(slug or _platform_slug(), "audiocpp_server")
+
+
+def _probe_paths() -> list[Path]:
+    out: list[Path] = []
+    direct = os.environ.get(BIN_ENV, "").strip()
+    if direct:
+        out.append(Path(direct))
+    user_dir = os.environ.get(DIR_ENV, "").strip()
+    if user_dir:
+        out.append(Path(user_dir) / binary_name())
+    out.append(_PKG_BIN_DIR / binary_name())
+    return out
+
+
+def is_installed() -> bool:
+    """Cheap precedence-aware check for a usable server binary."""
+    try:
+        resolve_server_binary()
+    except RuntimeError:
+        return False
+    return True
+
+
+def resolve_server_binary() -> Path:
+    """Resolve the ``audiocpp_server`` binary. Raises ``RuntimeError`` with
+    install instructions when none is found."""
+    for cand in _probe_paths():
+        if cand.is_file():
+            if os.name == "nt" or os.access(cand, os.X_OK):
+                return cand
+            raise RuntimeError(
+                "audiocpp_server is not executable. Run `chmod +x "
+                "audiocpp_server` on the configured binary, then restart "
+                "VoiceStudio. See docs/engines/audio-cpp.md."
+            )
+    slug = _platform_slug()
+    asset = _ASSETS.get(slug)
+    if asset is None:
+        raise RuntimeError(
+            f"audio.cpp ships no prebuilt binary for this platform ({slug}). "
+            "Build from https://github.com/0xShug0/audio.cpp and set "
+            f"{BIN_ENV} to your audiocpp_server binary. See "
+            "docs/engines/audio-cpp.md."
+        )
+    raise RuntimeError(
+        "audiocpp_server not found. Download "
+        f"https://github.com/{GH_REPO}/releases/download/{VERSION}/{asset[0]} "
+        f"(SHA-256 {asset[1]}), verify and extract it, and set {BIN_ENV} to the "
+        "audiocpp_server binary (or "
+        f"{DIR_ENV} to its directory). See docs/engines/audio-cpp.md."
+    )
+
+
+@functools.lru_cache(maxsize=4)
+def _probe_device_outcome(binary: str) -> _ProbeOutcome:
+    try:
+        proc = subprocess.run(  # nosec B603 -- executable is the resolved engine binary
+            [binary, "--list-devices"],
+            capture_output=True,
+            text=True,
+            timeout=10,
+            check=False,
+        )
+    except subprocess.TimeoutExpired:
+        return _ProbeOutcome(
+            error="audiocpp_server device discovery timed out after 10 seconds"
+        )
+    except OSError as exc:
+        return _ProbeOutcome(
+            error=(
+                "audiocpp_server device discovery could not start: "
+                f"{type(exc).__name__}"
+            )
+        )
+    if proc.returncode != 0:
+        return _ProbeOutcome(
+            error=(
+                "audiocpp_server device discovery failed "
+                f"(code {proc.returncode}). Check the audio.cpp server log "
+                "for details."
+            )
+        )
+    try:
+        return _ProbeOutcome(devices=parse_device_list(proc.stdout))
+    except RuntimeError as exc:
+        return _ProbeOutcome(error=str(exc))
+
+
+def _probe_devices(binary: str) -> tuple[AudioCPPDevice, ...]:
+    outcome = _probe_device_outcome(binary)
+    if outcome.error:
+        raise RuntimeError(outcome.error)
+    return outcome.devices
+
+
+def probe_devices() -> tuple[AudioCPPDevice, ...]:
+    """Return the installed binary's devices without loading a model."""
+    return _probe_devices(str(resolve_server_binary()))
+
+
+def _priority(device: AudioCPPDevice) -> tuple[int, int]:
+    if device.kind == "META":
+        # Tensor-parallel meta devices are valid explicit targets, but their
+        # resource footprint is not safe to choose implicitly over CPU.
+        rank = 8
+    elif device.backend != "cpu" and device.kind == "CPU":
+        # Native CPU is the predictable fallback. Software adapters remain
+        # available to an explicit backend override but never win auto mode.
+        rank = 7
+    elif device.backend == "cuda":
+        rank = 0
+    elif device.backend == "hip":
+        rank = 1
+    elif device.backend == "metal":
+        rank = 2
+    elif device.backend == "vulkan" and device.kind == "GPU":
+        rank = 3
+    elif device.backend == "vulkan" and device.kind in {"IGPU", "ACCEL"}:
+        rank = 4
+    elif device.backend == "cpu":
+        rank = 6
+    else:
+        rank = 5
+    return rank, device.index
+
+
+def select_device(
+    devices: tuple[AudioCPPDevice, ...],
+    *,
+    requested_family: str = "auto",
+    backend_override: str | None = None,
+    device_override: int | None = None,
+    preferred_name: str = "",
+) -> AudioCPPSelection:
+    """Resolve one device with explicit overrides and discrete-GPU priority."""
+    if backend_override:
+        normalized = _BACKEND_ALIASES.get(backend_override.strip().lower())
+        if normalized is None:
+            valid = ", ".join(_BACKEND_ALIASES)
+            raise RuntimeError(
+                f"unknown audio.cpp backend '{backend_override}' (valid: {valid})"
+            )
+        candidates = [device for device in devices if device.backend == normalized]
+        if device_override is not None:
+            candidates = [
+                device for device in candidates if device.index == device_override
+            ]
+        if not candidates:
+            suffix = "" if device_override is None else f" device {device_override}"
+            available = ", ".join(
+                f"{device.backend}:{device.index}" for device in devices
+            )
+            raise RuntimeError(
+                f"audio.cpp backend '{backend_override}'{suffix} is unavailable "
+                f"(available: {available})"
+            )
+        # An explicit runtime request should still prefer a compute device to
+        # a software adapter when no backend-local index was supplied. META is
+        # valid here because the user explicitly chose this registry.
+        return AudioCPPSelection(min(
+            candidates,
+            key=lambda device: (device.kind == "CPU", _priority(device)),
+        ))
+
+    if device_override is not None:
+        raise RuntimeError(
+            f"{DEVICE_ENV} requires {BACKEND_ENV} because device indices are "
+            "backend-local"
+        )
+
+    family = (requested_family or "auto").strip().lower()
+    if family != "auto":
+        candidates = [
+            device for device in devices if device.hardware_family == family
+        ]
+        if candidates:
+            preferred = preferred_name.casefold().strip()
+            if preferred:
+                named = [
+                    device for device in candidates
+                    if device.name
+                    and (
+                        preferred in device.name.casefold()
+                        or device.name.casefold() in preferred
+                    )
+                ]
+                if named:
+                    candidates = named
+            return AudioCPPSelection(min(candidates, key=_priority))
+        cpu = [device for device in devices if device.backend == "cpu"]
+        if cpu:
+            return AudioCPPSelection(
+                min(cpu, key=_priority),
+                f"requested {family.upper()} device is not exposed by the "
+                "installed audio.cpp binary; running on CPU",
+            )
+        raise RuntimeError(
+            f"requested {family.upper()} device is not exposed by the "
+            "installed audio.cpp binary"
+        )
+
+    return AudioCPPSelection(min(devices, key=_priority))
+
+
+def resolve_compute_selection(caps=None) -> AudioCPPSelection:
+    """Select the runtime from engine env overrides, Settings, then auto."""
+    backend_override = os.environ.get(BACKEND_ENV, "").strip() or None
+    raw_device = os.environ.get(DEVICE_ENV, "").strip()
+    device_override: int | None = None
+    if raw_device:
+        try:
+            device_override = int(raw_device)
+        except ValueError as exc:
+            raise RuntimeError(
+                f"{DEVICE_ENV} must be a non-negative integer"
+            ) from exc
+        if device_override < 0:
+            raise RuntimeError(f"{DEVICE_ENV} must be a non-negative integer")
+
+    if caps is None:
+        from core.device_caps import detect_host_caps
+
+        caps = detect_host_caps()
+    requested = getattr(caps, "requested_family", "auto") or "auto"
+    try:
+        devices = probe_devices()
+    except RuntimeError as exc:
+        if backend_override or raw_device or requested != "auto":
+            raise
+        return _cpu_probe_fallback(exc)
+    selection = select_device(
+        devices,
+        requested_family=requested,
+        backend_override=backend_override,
+        device_override=device_override,
+        preferred_name=getattr(caps, "device_name", "") or "",
+    )
+    # HostCaps measures the preferred accelerator's device 0. Reuse that VRAM
+    # only when the selected native registry has exactly one device with the
+    # same normalized name. Multi-GPU peers with identical names stay unknown.
+    selected_name = " ".join(selection.device.name.casefold().split())
+    host_name = " ".join(
+        str(getattr(caps, "device_name", "") or "").casefold().split()
+    )
+    peers = [
+        device for device in devices
+        if device.backend == selection.device.backend
+        and " ".join(device.name.casefold().split()) == host_name
+    ]
+    if (
+        selected_name
+        and selected_name == host_name
+        and len(peers) == 1
+        and float(getattr(caps, "vram_gb", 0.0) or 0.0) > 0
+    ):
+        return AudioCPPSelection(
+            selection.device,
+            selection.fallback_reason,
+            float(caps.vram_gb),
+        )
+    return selection
+
+
+def runtime_targets(devices: tuple[AudioCPPDevice, ...] | None = None) -> tuple[str, ...]:
+    """Actual compute backends compiled into the selected binary."""
+    if devices is not None:
+        found = devices
+    else:
+        try:
+            found = probe_devices()
+        except RuntimeError:
+            if (
+                os.environ.get(BACKEND_ENV, "").strip()
+                or os.environ.get(DEVICE_ENV, "").strip()
+            ):
+                raise
+            return ("cpu",)
+    ordered: list[str] = []
+    for device in sorted(found, key=_priority):
+        if device.target not in ordered:
+            ordered.append(device.target)
+    return tuple(ordered)
+
+
+def invalidate() -> None:
+    """Forget cached binary capability discovery after an install change."""
+    _probe_device_outcome.cache_clear()
+
+
+def default_asset() -> tuple[str, str] | None:
+    """``(filename, sha256)`` of the release asset for this host, or None
+    when upstream ships no prebuilt for it."""
+    return _ASSETS.get(_platform_slug())
+
+
+def server_port() -> int:
+    """Loopback port for the managed server (env override or default)."""
+    raw = os.environ.get(PORT_ENV, "").strip()
+    if raw:
+        try:
+            port = int(raw)
+            if 1 <= port <= 65535:
+                return port
+            logger.warning("Ignoring %s=%r: out of range.", PORT_ENV, raw)
+        except ValueError:
+            logger.warning("Ignoring %s=%r: not a number.", PORT_ENV, raw)
+    return DEFAULT_PORT
+
+
+def package_filename() -> str:
+    """GGUF package filename (env override or the Q8_0 default)."""
+    return os.environ.get(PACKAGE_ENV, "").strip() or DEFAULT_PACKAGE
+
+
+def _materialize_gguf_cache_path(model_file: Path) -> Path:
+    """Return a real ``.gguf`` path when the HF snapshot is a symlink.
+
+    audio.cpp canonicalizes model paths before inspecting the suffix. The
+    Hugging Face cache points the friendly ``.gguf`` snapshot name at an
+    extensionless content-addressed blob, so passing that symlink makes the
+    server reject a valid model. A hard link beside the snapshot keeps the
+    required suffix without copying a multi-gigabyte model or escaping the
+    snapshot's cleanup lifecycle.
+    """
+    resolved = model_file.resolve()
+    if resolved.suffix.lower() == ".gguf":
+        return model_file
+    if model_file.suffix.lower() != ".gguf":
+        raise RuntimeError(f"audio.cpp model must be a .gguf file: {model_file}")
+
+    def _link(alias: Path) -> Path:
+        for attempt in range(2):
+            try:
+                os.link(resolved, alias)
+            except FileExistsError:
+                if (
+                    not alias.is_symlink()
+                    and alias.is_file()
+                    and os.path.samefile(resolved, alias)
+                ):
+                    return alias
+                if attempt == 0 and alias.is_symlink():
+                    alias.unlink()
+                    continue
+                raise RuntimeError(
+                    f"audio.cpp model alias points at a different file: {alias}"
+                ) from None
+            return alias
+        raise RuntimeError(f"audio.cpp model alias could not be created: {alias}")
+
+    alias = model_file.with_name(
+        f".{model_file.stem}-{HF_MODEL_REVISION[:12]}.audiocpp.gguf"
+    )
+    try:
+        return _link(alias)
+    except OSError as exc:
+        if exc.errno == errno.EXDEV:
+            # An explicit symlink may live on a different filesystem from its
+            # target. Put the suffix-preserving hard link beside the resolved
+            # file so no multi-gigabyte copy is needed.
+            target_alias = resolved.with_name(
+                f".{resolved.name}-{HF_MODEL_REVISION[:12]}.audiocpp.gguf"
+            )
+            try:
+                return _link(target_alias)
+            except OSError as target_exc:
+                exc = target_exc
+        raise RuntimeError(
+            "audio.cpp cannot materialize the Hugging Face cache symlink as "
+            f"a .gguf hard link: {exc}"
+        ) from exc
+
+
+def resolve_model_file() -> Path:
+    """Resolve an explicitly installed Breeze-TTS-2 GGUF file.
+
+    An explicit ``OMNIVOICE_AUDIOCPP_MODEL`` path wins (file or directory
+    containing the package file). Otherwise only the local Hugging Face cache
+    is inspected. Downloads must be started explicitly from Model Catalogue →
+    Models, so generation can never silently transfer the 4.73 GiB package.
+    """
+    override = os.environ.get("OMNIVOICE_AUDIOCPP_MODEL", "").strip()
+    if override:
+        cand = Path(override)
+        if cand.is_file():
+            return _materialize_gguf_cache_path(cand)
+        if cand.is_dir():
+            inner = cand / package_filename()
+            if inner.is_file():
+                return _materialize_gguf_cache_path(inner)
+        raise RuntimeError(
+            f"OMNIVOICE_AUDIOCPP_MODEL={override} is not a GGUF file or a "
+            "directory containing one."
+        )
+
+    from huggingface_hub import snapshot_download
+    from huggingface_hub.utils import LocalEntryNotFoundError
+
+    try:
+        cached = Path(
+            snapshot_download(
+                repo_id=HF_MODEL_REPO,
+                # Full immutable commit SHA declared above; Bandit cannot follow
+                # the module constant through this call.
+                revision=HF_MODEL_REVISION,  # nosec B615
+                allow_patterns=[f"{PACKAGE_DIR}/{package_filename()}"],
+                local_files_only=True,
+            )
+        )
+    except (LocalEntryNotFoundError, OSError) as exc:
+        raise RuntimeError(
+            "Breeze-TTS-2 is not installed. Install the audio.cpp Breeze-TTS-2 "
+            "model from Model Catalogue → Models, or set "
+            "OMNIVOICE_AUDIOCPP_MODEL to an existing GGUF file."
+        ) from exc
+    model_file = cached / PACKAGE_DIR / package_filename()
+    if not model_file.is_file():
+        raise RuntimeError(
+            f"Breeze-TTS-2 package {package_filename()} is not completely "
+            "installed. Reinstall it from Model Catalogue → Models."
+        )
+    return _materialize_gguf_cache_path(model_file)
+
+
+__all__ = [
+    "AudioCPPDevice",
+    "AudioCPPSelection",
+    "BACKEND_ENV",
+    "BIN_ENV",
+    "DEFAULT_PACKAGE",
+    "DEFAULT_PORT",
+    "DEVICE_ENV",
+    "DIR_ENV",
+    "FAMILY",
+    "HF_MODEL_REPO",
+    "HF_MODEL_REVISION",
+    "MODEL_ID",
+    "PACKAGE_DIR",
+    "PACKAGE_ENV",
+    "PORT_ENV",
+    "VERSION",
+    "_materialize_gguf_cache_path",
+    "binary_name",
+    "default_asset",
+    "invalidate",
+    "is_installed",
+    "package_filename",
+    "parse_device_list",
+    "probe_devices",
+    "resolve_compute_selection",
+    "resolve_model_file",
+    "resolve_server_binary",
+    "runtime_targets",
+    "server_port",
+]

backend/engines/confucius4/__init__.py

@@ -56,15 +56,15 @@ class Confucius4Backend(SubprocessBackend):
 
     id = "confucius4-tts"
     display_name = (
-        "Confucius4-TTS (LLM, 14 langs, cross-lingual zero-shot clone, CUDA/CPU, Apache-2.0)"
+        "Confucius4-TTS (LLM, 14 langs, cross-lingual zero-shot clone, Apache-2.0)"
     )
     supports_voice_design = False  # timbre comes from a reference clip
     # Upstream vocoder rate (config target_sample_rate) — confirmed 22 050 Hz by
     # a live run (2026-07-02); still re-read from the sidecar's ready/audio frames.
     _DEFAULT_SAMPLE_RATE = 22050
-    # CUDA fast path + CPU fallback, both exercised (CPU end-to-end validated).
-    # No MPS claim — upstream has no Metal path.
-    gpu_compat = ("cuda", "cpu")
+    # Match device propagation into upstream .to(device). XPU/NPU routing is
+    # contract-tested, not a claim of physical-hardware synthesis validation.
+    gpu_compat = ("cuda", "rocm", "xpu", "npu", "cpu")
 
     @classmethod
     def is_available(cls) -> tuple[bool, str]:

backend/engines/confucius4/main.py

@@ -104,7 +104,7 @@ def _ensure_clone_on_sys_path() -> None:
 
 
 def _load_model(stdout):
-    """Cold-construct the Confucius4 model (CUDA, else CPU — both validated)."""
+    """Cold-construct using an available torch accelerator, with CPU fallback."""
     global _model
     if _model is not None:
         return _model
@@ -115,7 +115,18 @@ def _load_model(stdout):
     import torch
     from confuciustts.cli.inference import ConfuciusTTS  # type: ignore[import-not-found]
 
-    device = "cuda" if torch.cuda.is_available() else "cpu"
+    try:
+        # Existing manually provisioned venvs may predate torch.accelerator.
+        current_accelerator = getattr(getattr(torch, "accelerator", None), "current_accelerator", None)
+        if current_accelerator is None:
+            device = torch.device("cuda") if torch.cuda.is_available() else None
+        else:
+            device = current_accelerator(check_available=True)
+        device = device.type if device is not None else "cpu"  # 'cuda', 'npu', 'mps', 'xpu', 'cpu'
+    except Exception:
+        device = "cpu"  # Broken accelerator drivers must not block CPU loading.
+    if device == "mps":
+        device = "cpu"  # MPS was slower than CPU in the existing validation run
     _send(stdout, {"op": "progress", "stage": "loading_model", "percent": 50})
 
     _model = ConfuciusTTS(config_path=_config_path(), device=device)

backend/engines/dots_tts/main.py

@@ -115,7 +115,12 @@ def _load_runtime(stdout):
     from dots_tts.runtime import DotsTtsRuntime  # type: ignore[import-not-found]
 
     repo = os.environ.get("OMNIVOICE_DOTS_TTS_MODEL", _DEFAULT_REPO)
-    default_precision = "bfloat16" if torch.cuda.is_available() else "float32"
+    # Match DotsTtsRuntime's own CUDA/CPU selection. Its _check_torch_env
+    # rejects half precision without CUDA, even when an XPU/NPU is available.
+    try:
+        default_precision = "bfloat16" if torch.cuda.is_available() else "float32"
+    except Exception:
+        default_precision = "float32"  # Probe failure must not force half precision.
     precision = os.environ.get("OMNIVOICE_DOTS_TTS_PRECISION", default_precision)
     optimize = os.environ.get("OMNIVOICE_DOTS_TTS_OPTIMIZE", "0") == "1"
 

backend/engines/moss_tts_v15/__init__.py

@@ -29,15 +29,11 @@ Do NOT import ``main.py`` from the parent process — it runs under a
 different venv (``transformers==5.0.0``) and importing it in-process would
 re-introduce the exact conflict this isolation exists to avoid.
 
-Hardware honesty (cross-platform rule): MOSS-TTS-v1.5's upstream documents
-only CUDA and CPU. There is **no documented or tested MPS path** — the
-custom ``trust_remote_code`` modelling code and the separate audio
-tokenizer are unverified on Apple Silicon. We therefore advertise
-``gpu_compat = ("cuda", "cpu")`` and the sidecar selects ``cuda`` when
-present else ``cpu`` — it never silently routes to MPS where it might
-crash. On Apple Silicon the engine honestly resolves to CPU (slow but
-correct), and the engine is opt-in regardless, so it never becomes a
-broken default on any platform.
+Hardware routing follows the sidecar's runtime-available PyTorch accelerator:
+CUDA/ROCm, XPU, or a registered NPU. MPS remains excluded; CPU is the fallback.
+XPU/NPU routing is covered with mocked device contracts, not physical-hardware
+synthesis certification; users need a compatible torch/vendor runtime in the
+isolated engine venv.
 """
 from __future__ import annotations
 
@@ -85,13 +81,13 @@ class MossTTSV15Backend(SubprocessBackend):
 
     id = "moss-tts-v15"
     display_name = (
-        "MOSS-TTS-v1.5 (8B, 31 langs, zero-shot clone, CUDA/CPU, Apache-2.0)"
+        "MOSS-TTS-v1.5 (8B, 31 langs, zero-shot clone, Apache-2.0)"
     )
     supports_voice_design = False  # requires ref audio for timbre cloning
     _DEFAULT_SAMPLE_RATE = 24000
-    # Honest hardware surface: upstream documents CUDA + CPU only. MPS is
-    # undocumented / untested, so we do NOT claim it (cross-platform rule).
-    gpu_compat = ("cuda", "cpu")
+    # Accelerator routing requires its matching runtime in the isolated venv.
+    # MPS remains untested and is deliberately excluded.
+    gpu_compat = ("cuda", "rocm", "xpu", "npu", "cpu")
 
     # ── availability ───────────────────────────────────────────────────────
 
@@ -111,7 +107,7 @@ class MossTTSV15Backend(SubprocessBackend):
             return False, (
                 "MOSS-TTS-v1.5 venv not found. Set OMNIVOICE_MOSS_TTS_V15_DIR "
                 "to your MOSS-TTS clone (the directory containing pyproject.toml) "
-                "and restart VoiceStudio. CUDA or CPU only (no MPS). See "
+                "and restart VoiceStudio. Install the matching PyTorch runtime. See "
                 "docs/engines/moss-tts-v15.md for the full install walk-through."
             )
         if not MOSS_TTS_V15_SIDECAR_SCRIPT.exists():
@@ -119,7 +115,7 @@ class MossTTSV15Backend(SubprocessBackend):
                 "MOSS-TTS-v1.5 sidecar script missing at "
                 f"{MOSS_TTS_V15_SIDECAR_SCRIPT} — reinstall VoiceStudio."
             )
-        return True, "ok (CUDA when present, else CPU)"
+        return True, "ok (runtime-available accelerator or CPU; no MPS)"
 
     @classmethod
     def venv_python(cls):

backend/engines/moss_tts_v15/bootstrap.py

@@ -222,8 +222,7 @@ def _bootstrap_engines_venv(clone_dir: Path) -> Path:
     Runs ``uv venv <engines_venv>`` then ``uv pip install --python
     <engines_venv>/bin/python -e "<clone>[torch-runtime]"``. Verifies the
     result by re-probing the import — a successful uv invocation that still
-    can't import the stack indicates a deeper environment problem (e.g. the
-    ``+cu128`` torch-runtime extra can't resolve on a non-CUDA host) and we
+    can't import the stack indicates a deeper environment problem, and we
     raise with whatever stderr we captured plus a docs pointer.
     """
     uv = _locate_uv()
@@ -254,6 +253,8 @@ def _bootstrap_engines_venv(clone_dir: Path) -> Path:
             f"{exc.stderr.decode('utf-8', errors='replace') if exc.stderr else exc}"
         ) from exc
 
+    from core.torch_indexes import UV_PIP_CU128_ARGS
+
     python_path = _venv_python_path(_ENGINES_VENV_DIR)
     try:
         subprocess.run(
@@ -261,6 +262,10 @@ def _bootstrap_engines_venv(clone_dir: Path) -> Path:
                 uv, "pip", "install",
                 "--python", str(python_path),
                 "-e", f"{clone_dir}[torch-runtime]",
+                # The extra pins torch==2.9.1+cu128, which exists only on
+                # PyTorch's index — without it this could never resolve, on
+                # any host (core.torch_indexes).
+                *UV_PIP_CU128_ARGS,
             ],
             check=True,
             timeout=_UV_PIP_INSTALL_TIMEOUT_S,
@@ -270,9 +275,10 @@ def _bootstrap_engines_venv(clone_dir: Path) -> Path:
     except subprocess.CalledProcessError as exc:
         raise RuntimeError(
             "uv pip install -e failed during MOSS-TTS-v1.5 bootstrap "
-            f"({clone_dir}). On a non-CUDA host the upstream '[torch-runtime]' "
-            "extra (cu128) cannot resolve — set up the venv manually per "
-            "docs/engines/moss-tts-v15.md. Error: "
+            # uv's own error names what failed; the PyTorch index is always
+            # supplied now, so a guess about the host would only mislead.
+            f"({clone_dir}). See docs/engines/moss-tts-v15.md for the manual "
+            "install. Error: "
             f"{exc.stderr.decode('utf-8', errors='replace') if exc.stderr else exc}"
         ) from exc
 

backend/engines/moss_tts_v15/main.py

@@ -138,11 +138,12 @@ _state = None
 def _load_model(stdout):
     """Cold-construct the MOSS-TTS-v1.5 processor + model.
 
-    Device selection is CUDA-or-CPU only — MOSS's upstream documents no MPS
-    path and the custom ``trust_remote_code`` modelling code is untested on
-    Apple Silicon, so we never route to MPS where it might crash. dtype is
-    bf16 on CUDA, fp32 on CPU (bf16 CPU ops are spotty). Emits progress
-    frames so the parent can surface the multi-GB cold-load latency.
+    Device selection uses the torch.accelerator API to support any backend
+    (CUDA, NPU, XPU, etc.) automatically. MPS is excluded — MOSS's upstream
+    ``trust_remote_code`` modelling code is untested on Apple Silicon. dtype is
+    bf16 on GPU-class accelerators, fp32 on CPU (bf16 CPU ops are spotty).
+    Emits progress frames so the parent can surface the multi-GB cold-load
+    latency.
     """
     global _state
     if _state is not None:
@@ -154,8 +155,22 @@ def _load_model(stdout):
     from transformers import AutoModel, AutoProcessor
 
     repo, revision = _model_source()
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    dtype = torch.bfloat16 if device == "cuda" else torch.float32
+    # current_accelerator() returns None on CPU-only builds (no accelerator
+    # compiled in) or when no accelerator is available; fall back to "cpu".
+    # Existing manually provisioned venvs may predate torch.accelerator.
+    current_accelerator = getattr(getattr(torch, "accelerator", None), "current_accelerator", None)
+    try:
+        if current_accelerator is None:
+            accel = torch.device("cuda") if torch.cuda.is_available() else None
+        else:
+            accel = current_accelerator(check_available=True)
+    except Exception:
+        # Optional drivers can fail during probing; CPU loading remains usable.
+        accel = None
+    device = accel.type if accel is not None else "cpu"  # 'cuda', 'npu', 'mps', 'xpu', 'cpu'
+    if device == "mps":
+        device = "cpu"  # MOSS is untested on MPS; fall back to CPU for safety
+    dtype = torch.bfloat16 if device != "cpu" else torch.float32
     # "sdpa" works on CUDA + CPU and needs no extra dep. flash_attention_2
     # (Ampere+ CUDA, optional flash-attn) is opt-in via env.
     attn = os.environ.get("OMNIVOICE_MOSS_TTS_V15_ATTN", "sdpa")

backend/engines/pockettts/__init__.py

@@ -48,6 +48,15 @@ from services.subprocess_backend import SubprocessBackend
 
 logger = logging.getLogger("omnivoice.engines.pockettts")
 
+_VENV_ENV_VAR = "OMNIVOICE_POCKETTTS_DIR"
+
+
+def _own_venv_python() -> "Path | None":
+    """The venv the one-click installer made for this engine, if any."""
+    from services.sidecar_install import engine_venv_python
+
+    return engine_venv_python(_VENV_ENV_VAR)
+
 if TYPE_CHECKING:
     import torch  # noqa: F401
 
@@ -121,16 +130,17 @@ class PocketTTSBackend(SubprocessBackend):
     def is_available(cls) -> tuple[bool, str]:
         if platform_error := cls._platform_error():
             return False, platform_error
-        # Optional-dep gate: the pocket-tts wheel is installed only when the user
-        # opted in. The interpreter is the parent's own (sys.executable), so
-        # there is no separate venv to validate.
-        try:
-            import pocket_tts  # type: ignore[import-not-found]  # noqa: F401
-        except Exception as e:
-            return False, (
-                f"pocket_tts package not installed or failed to import ({e}). "
-                f"Enable in Settings -> Engines (uv sync --extra pockettts)."
-            )
+        # Installed either into its own venv by the one-click installer, which
+        # verified `import pocket_tts` there before saving the path, or into the
+        # app's environment by `uv sync --extra pockettts`.
+        if _own_venv_python() is None:
+            try:
+                import pocket_tts  # type: ignore[import-not-found]  # noqa: F401
+            except Exception as e:
+                return False, (
+                    f"pocket_tts package not installed or failed to import ({e}). "
+                    "Install it from Model Catalogue → Engines."
+                )
 
         # The model repository has an additional gated-access agreement and
         # prohibited-use conditions beyond its CC-BY-4.0 license. Keep first
@@ -145,10 +155,10 @@ class PocketTTSBackend(SubprocessBackend):
 
     @classmethod
     def venv_python(cls) -> Path:
-        # Parent interpreter: pocket-tts deps (torch>=2.5, scipy, beartype) sit
-        # happily at the parent's pins, so this isolates for crash recovery, not
-        # dependency pins (same rationale as omnivoice-subprocess).
-        return Path(sys.executable)
+        # Its own venv when the one-click installer made one. Otherwise the
+        # parent interpreter, where `uv sync --extra pockettts` installs it
+        # (its deps sit happily at the parent's pins).
+        return _own_venv_python() or Path(sys.executable)
 
     @classmethod
     def sidecar_script(cls) -> Path:

backend/engines/supertonic3/backend.py

@@ -48,6 +48,15 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger("omnivoice.supertonic3")
 
+_VENV_ENV_VAR = "OMNIVOICE_SUPERTONIC3_DIR"
+
+
+def _own_venv_python() -> "Path | None":
+    """The venv the one-click installer made for this engine, if any."""
+    from services.sidecar_install import engine_venv_python
+
+    return engine_venv_python(_VENV_ENV_VAR)
+
 
 # Absolute path to the sidecar script ‑‑ same pattern as IndexTTS's
 # ``INDEXTTS_SIDECAR_SCRIPT``. SubprocessBackend spawns it with the
@@ -80,11 +89,11 @@ class Supertonic3Backend(SubprocessBackend):
 
     @classmethod
     def venv_python(cls) -> Path:
-        """Supertonic-3 lives in the main OmniVoice venv ‑‑ no dedicated
-        venv. ``sys.executable`` is the parent interpreter, which is the
-        same Python that ``uv sync --extra supertonic`` populated.
+        """Its own venv when the one-click installer made one. Otherwise the
+        parent interpreter, the same Python ``uv sync --extra supertonic``
+        populated.
         """
-        return Path(sys.executable)
+        return _own_venv_python() or Path(sys.executable)
 
     @classmethod
     def sidecar_script(cls) -> Path:
@@ -96,14 +105,16 @@ class Supertonic3Backend(SubprocessBackend):
     def is_available(cls) -> tuple[bool, str]:
         # 1. Optional-dep gate (TTS-02). The ``supertonic`` wheel is only
         #    installed when the user opted in via ``--extra supertonic``.
-        try:
-            import supertonic  # type: ignore[import-not-found]  # noqa: F401
-        except ImportError:
-            return False, (
-                "supertonic package not installed. Enable in "
-                "Model Catalogue → Engines (installs `supertonic` via `uv add --optional "
-                "supertonic supertonic==1.3.1`)."
-            )
+        #    Its own venv (made by the one-click installer, which verified the
+        #    import there) or the app's environment (`uv sync --extra`).
+        if _own_venv_python() is None:
+            try:
+                import supertonic  # type: ignore[import-not-found]  # noqa: F401
+            except ImportError:
+                return False, (
+                    "supertonic package not installed. Install it from "
+                    "Model Catalogue → Engines."
+                )
 
         # 2. License acceptance gate (TTS-05). Defence in depth: the
         #    settings_store helper handles the read; we just refuse

backend/engines/supertonic3/sidecar.py

@@ -137,9 +137,17 @@ def _resolve_pinned_sha() -> str:
         # Final fallback ‑‑ relative import for when the file is invoked
         # via ``python backend/engines/supertonic3/sidecar.py`` rather
         # than via ``python -m backend.engines.supertonic3.sidecar``.
-        sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
-        from engines.supertonic3.constants import PINNED_REVISION_SHA  # type: ignore[import-not-found]
-        return PINNED_REVISION_SHA
+        # Load constants.py by path. Importing it as `engines.supertonic3…`
+        # runs the package __init__, which imports the app's backend, and that
+        # is absent from the engine's own venv (one-click install).
+        import importlib.util
+
+        spec = importlib.util.spec_from_file_location(
+            "_supertonic3_constants", Path(__file__).resolve().with_name("constants.py"),
+        )
+        module = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(module)  # type: ignore[union-attr]
+        return module.PINNED_REVISION_SHA
 
 
 # ── model loading (lazy, on first synthesize) ─────────────────────────────

backend/main.py

@@ -624,7 +624,19 @@ def _phase_a_build_inner() -> None:
     _startup_progress.begin_step("ml_imports")
     import torchaudio
     warnings.filterwarnings("ignore", category=UserWarning)
-    torchaudio.set_audio_backend("soundfile")
+    # torchaudio 2.9 REMOVED set_audio_backend(); soundfile has been the only
+    # backend since 2.0, so the call was already a no-op there and is simply
+    # absent now. Unguarded it raises AttributeError inside `ml_imports`, and a
+    # failure in that phase takes the whole backend down — the desktop app sits
+    # on "starting backend" forever and /health stays 503.
+    #
+    # That is not a hypothetical version: RTX 50-series (Blackwell, sm_120)
+    # users have no choice but to move off the pinned torch 2.8.0, which has no
+    # sm_120 kernels, and the torch 2.9.x they land on brings torchaudio 2.9
+    # with it. So the one group forced to upgrade hit a hard startup crash for
+    # a line that does nothing (#1931).
+    if hasattr(torchaudio, "set_audio_backend"):
+        torchaudio.set_audio_backend("soundfile")
     from utils import hf_progress
     # HF tqdm patch before any library import that can trigger
     # hf_hub_download (transformers, mlx_whisper, …).
@@ -1080,6 +1092,33 @@ async def lifespan(app: FastAPI):
         app.state.startup_task = asyncio.create_task(_deferred_startup(app))
     yield
     # ── Graceful shutdown (SIGTERM from Tauri, Ctrl+C, etc.) ────────────
+    # Retire the run sentinel FIRST, before any bounded wait below (#1895):
+    # once uvicorn has begun graceful shutdown the exit is deliberate by
+    # definition, so the sentinel has already done its job. This is one
+    # os.remove, against a ~50s worst-case tail of bounded waits plus model
+    # unload / free_vram() / gc.collect() below. Measured on macOS: a normal
+    # shutdown takes 5.25s end to end, while the desktop shell allows 2s
+    # (bootstrap.rs terminate_process_tree) before SIGKILL — so the old
+    # placement at the very end was killed every time on any run that had
+    # reached a working state. Doing the deadline-sensitive step first makes
+    # correctness independent of how much of that tail runs, instead of
+    # depending on the shell-side deadline being long enough to cover it.
+    #
+    # SCOPE, explicitly: this only helps platforms where lifespan teardown
+    # actually BEGINS. On Windows it does not — tools.rs terminates the job
+    # object with no graceful phase at all, so this line is never reached and
+    # a deliberate quit is still misreported as a crash there. That needs the
+    # shell to signal deliberate intent before the hard kill, which is a
+    # separate Rust-side change and is tracked separately; nothing here
+    # should be read as fixing Windows.
+    #
+    # sentinel_cleared feeds the truthful "Shutdown: done."/degraded log at
+    # the end of this function; nothing below re-clears the sentinel, so a
+    # later failure can't mask this result.
+    try:
+        sentinel_cleared = run_sentinel.clear_sentinel()
+    except Exception:
+        sentinel_cleared = False
     # May run after a startup that never finished (SIGTERM mid-Phase-A/B), so
     # every handle is read from app.state with a None default and every
     # deferred-phase name is guarded.
@@ -1206,13 +1245,9 @@ async def lifespan(app: FastAPI):
         await close_http_client()
     except Exception:
         pass
-    # Last thing on a clean shutdown: retire the run sentinel so the next
-    # startup doesn't misread this exit as a crash (#1164). If clearing fails,
-    # retain the sentinel and report a degraded shutdown truthfully.
-    try:
-        sentinel_cleared = run_sentinel.clear_sentinel()
-    except Exception:
-        sentinel_cleared = False
+    # Sentinel was already retired at the TOP of this block (#1895) — report
+    # truthfully using that result rather than clearing (or re-checking) it
+    # again here, so a failure in the steps above can't mask it as "done."
     if sentinel_cleared:
         logger.info("Shutdown: done.")
     else:

backend/migrations/versions/0011_remote_attempt_deadlines.py

@@ -0,0 +1,18 @@
+"""Retain the dispatch-time deadline policy across worker/control-plane loss."""
+from alembic import op
+import sqlalchemy as sa
+
+revision = "0011_remote_attempt_deadlines"
+down_revision = "0010_remote_worker_schema"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    columns = op.get_bind().execute(sa.text("PRAGMA table_info(remote_task_attempts)"))
+    if not any(row[1] == "deadlines_json" for row in columns):
+        op.add_column("remote_task_attempts", sa.Column("deadlines_json", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("remote_task_attempts", "deadlines_json")

backend/services/asr_backend.py

@@ -1801,9 +1801,7 @@ class SherpaDictationBackend(ASRBackend):
 
     def __init__(self, model_id: str | None = None):
         from services import sherpa_dictation as _sd
-        mid = model_id or os.environ.get(
-            "OMNIVOICE_SHERPA_ASR_MODEL", _sd.DEFAULT_MODEL_ID
-        )
+        mid = model_id or sherpa_engine_model_id()
         spec = _sd.get_spec(mid)
         if spec is None:
             raise ValueError(
@@ -3047,6 +3045,31 @@ def get_sherpa_dictation_backend(model_id: str) -> "SherpaDictationBackend":
         return backend
 
 
+def sherpa_engine_model_id() -> str:
+    """The sherpa model the ``sherpa-onnx-asr`` engine loads when nothing pins
+    one explicitly: env var (power-user pin) → the dictation model the user
+    picked in Settings / the Engines menu → the catalogue default.
+
+    Unlike :func:`dictation_model_id` this ignores ``dictation.enabled`` — a
+    user who turned the hotkey off but chose the Sherpa engine for dub/batch
+    transcription still means *this* model — and never returns None: the
+    engine needs *some* model to construct. A demoted model (decoded nothing
+    on this host) falls through to the default rather than being re-picked.
+    """
+    from services import sherpa_dictation as _sd
+    explicit = os.environ.get("OMNIVOICE_SHERPA_ASR_MODEL")
+    if explicit:
+        return explicit
+    try:
+        from core import prefs
+        mid = prefs.get("dictation.model_id")
+    except Exception:  # noqa: BLE001 — prefs store unavailable → default
+        return _sd.DEFAULT_MODEL_ID
+    if _sd.is_sherpa_model(mid) and not _sd.is_demoted(mid):
+        return _sd.get_spec(mid).id
+    return _sd.DEFAULT_MODEL_ID
+
+
 def dictation_model_id() -> str | None:
     """The selected sherpa dictation model id, or None when dictation is off /
     no sherpa model is chosen. Env var wins (power-user pin), then prefs."""
@@ -3321,9 +3344,7 @@ def _offline_asr_repo(backend_id: str | None = None) -> str | None:
         # Unknown/none → fail open.
         try:
             from services import sherpa_dictation as _sd
-            spec = _sd.get_spec(
-                os.environ.get("OMNIVOICE_SHERPA_ASR_MODEL", _sd.DEFAULT_MODEL_ID)
-            )
+            spec = _sd.get_spec(sherpa_engine_model_id())
             return spec.repo_id if spec is not None else None
         except Exception:  # noqa: BLE001 — preflight must stay best-effort
             return None

backend/services/endpoint_race.py

@@ -65,7 +65,7 @@ _MODE_PREF = "hf_endpoint_mode"          # "auto" | "manual"; absent → default
 _DECISION_PREF = "hf_endpoint_auto"      # cached decision dict (see race())
 
 DECISION_MAX_AGE_S = 7 * 24 * 3600.0     # re-race a decision older than 7 days
-PROBE_TIMEOUT_S = 3.0                    # short: a probe is not a download
+PROBE_TIMEOUT_S = 8.0                    # high-latency / China paths often need >3s
 MIRROR_SPEEDUP_FACTOR = 3.0              # mirror must be ≥3× faster to win
 
 # Small, stable, long-lived public file for the optional ranged-GET

backend/services/engine_disk_usage.py

@@ -33,10 +33,20 @@ _ESTIMATES: dict[str, dict] = {
         "destination": "hf_model_cache",
         "deduplication": None,
     },
+    "audiocpp": {
+        "package_download_bytes": None,
+        "unique_installed_bytes": None,
+        "potentially_shared_bytes": None,
+        "temporary_free_bytes": None,
+        "confidence": "estimated",
+        "destination": "hf_model_cache",
+        "deduplication": None,
+    },
 }
 _MODEL_REPOS = {
     "omnivoice": "k2-fsa/OmniVoice",
     "kittentts": "KittenML/kitten-tts-mini-0.8",
+    "audiocpp": "audio-cpp/audio.cpp-gguf",
 }
 
 

backend/services/engine_evidence.py

@@ -88,14 +88,34 @@ def snapshot(
         evidence_state = "loaded"
         if isolated and provider is None and actual_device is None:
             evidence_state = "subprocess_loaded_provider_unreported"
+    from core.scrub import scrub_text
+
+    runtime_device_name = routing.get("runtime_device_name")
+    device_name = (
+        getattr(caps, "device_name", "")
+        if runtime_device_name is None
+        else runtime_device_name
+    )
+    public_device_name = scrub_text(device_name)[:256]
     return {
         "implementation_variant": f"{engine_cls.__module__}.{engine_cls.__name__}",
-        "declared_device_families": list(getattr(engine_cls, "gpu_compat", ("cpu",))),
+        "declared_device_families": list(
+            routing.get("gpu_compat", getattr(engine_cls, "gpu_compat", ("cpu",)))
+        ),
         "evidence_state": evidence_state,
         "actual_execution_provider": provider,
         "actual_execution_device": actual_device,
-        "gpu_name": getattr(caps, "device_name", "") or None,
-        "gpu_architecture": _gpu_architecture(getattr(caps, "family", "cpu")),
+        "gpu_name": public_device_name or None,
+        "gpu_architecture": None
+        if (
+            routing.get("runtime_hardware_family")
+            and not routing.get("runtime_device_verified")
+        )
+        else _gpu_architecture(
+            routing.get("runtime_hardware_family")
+            or getattr(caps, "family", "cpu")
+        ),
+        "runtime_vram_gb": routing.get("runtime_vram_gb"),
         "precision_or_quantization": precision,
         "cpu_fallback_reason": runtime_fallback_reason or (routing.get("routing_reason") if fallback else None),
         "cpu_fallback_stage": runtime_fallback_stage or ("routing_preflight" if fallback else None),

backend/services/engine_routing.py

@@ -14,6 +14,7 @@ carry a home path.
 """
 from __future__ import annotations
 
+import asyncio
 from typing import Literal, TypedDict
 
 from core.device_caps import (
@@ -31,6 +32,93 @@ class RoutingResult(TypedDict):
     routing_reason: str | None     # raw, pre-scrub
 
 
+def runtime_compute_profile(engine_or_cls, caps: HostCaps) -> dict:
+    """Return one engine's runtime-aware compute contract.
+
+    Native executables may discover providers independently of PyTorch.  They
+    override ``runtime_compute_profile``; all existing engines retain the
+    exact static routing contract.
+    """
+    hook = getattr(engine_or_cls, "runtime_compute_profile", None)
+    if callable(hook):
+        return hook(caps)
+    cls = engine_or_cls if isinstance(engine_or_cls, type) else type(engine_or_cls)
+    compat = tuple(getattr(cls, "gpu_compat", ("cpu",)))
+    floor = float(getattr(cls, "min_vram_gb", 0.0) or 0.0)
+    return {
+        "gpu_compat": compat,
+        "min_vram_gb": floor,
+        **resolve_routing(compat, caps, floor),
+        "runtime_backend": None,
+        "runtime_device_index": None,
+        "runtime_device_name": None,
+        "runtime_hardware_family": None,
+        "runtime_vram_gb": None,
+        "runtime_device_verified": None,
+    }
+
+
+async def runtime_compute_profile_async(engine_or_cls, caps: HostCaps) -> dict:
+    """Resolve runtime compute metadata without blocking the event loop."""
+    return await asyncio.to_thread(runtime_compute_profile, engine_or_cls, caps)
+
+
+def under_provisioned_vram(
+    caps: HostCaps,
+    min_vram_gb: float = 0.0,
+    *,
+    family: str | None = None,
+    vram_gb: float | None = None,
+) -> bool:
+    """Is this host's DEDICATED VRAM below the engine's declared floor?
+
+    The one definition of "under-provisioned", shared by everything that acts
+    on the verdict: the routing caveat below, the timeout guidance, and — since
+    #1804 — the compute-time budget itself (``model_manager
+    .generate_timeout_s``). It was written out inline in each of them, which is
+    how the budget came to disagree with the warning printed next to it.
+
+    Dedicated-VRAM families ONLY. CUDA, ROCm, XPU, and a native Vulkan device
+    report dedicated memory. On MPS, ``HostCaps.vram_gb`` is a heuristic
+    (system RAM / 2, see device_caps) for a UNIFIED memory pool; comparing it
+    against a floor measured on discrete CUDA hardware would tell every 8 GB Mac
+    its 4 GB "VRAM" is too small for an engine that runs fine there. A VRAM
+    figure of 0 means the probe failed — don't guess from it. A floor of 0 means
+    the engine declares none, and inventing one is worse than staying quiet.
+    """
+    if not min_vram_gb or min_vram_gb <= 0:
+        return False
+    if (family or getattr(caps, "family", None)) not in (
+        "cuda", "rocm", "xpu", "vulkan",
+    ):
+        return False
+    raw_vram_gb = getattr(caps, "vram_gb", 0.0) if vram_gb is None else vram_gb
+    available_vram_gb = float(raw_vram_gb or 0.0)
+    return 0 < available_vram_gb < float(min_vram_gb)
+
+
+def low_vram_caveat(
+    caps: HostCaps,
+    min_vram_gb: float = 0.0,
+    *,
+    family: str | None = None,
+    vram_gb: float | None = None,
+) -> str | None:
+    """User-facing advisory for a known under-provisioned dedicated GPU."""
+    if not under_provisioned_vram(
+        caps, min_vram_gb, family=family, vram_gb=vram_gb,
+    ):
+        return None
+    device = caps.device_name or (family or caps.family).upper()
+    available_vram_gb = caps.vram_gb if vram_gb is None else vram_gb
+    return (
+        f"{device} has {available_vram_gb:.1f} GB VRAM; this engine wants about "
+        f"{min_vram_gb:.0f} GB. It will run, but expect slow generations "
+        f"that may time out. Unload other models before generating, keep "
+        f"the text short, or pick a lighter engine."
+    )
+
+
 def _caveat(caps: HostCaps, min_vram_gb: float = 0.0) -> str | None:
     """A caveat string for an otherwise-accelerated host, or None.
 
@@ -51,24 +139,7 @@ def _caveat(caps: HostCaps, min_vram_gb: float = 0.0) -> str | None:
     for note in caps.notes:
         if KERNEL_RISK_MARKER in note:
             return f"{caps.family.upper()} selected, but: {note}"
-    # Dedicated-VRAM families ONLY. On MPS, HostCaps.vram_gb is a heuristic
-    # (system RAM / 2, see device_caps) for a UNIFIED memory pool — comparing
-    # it against a floor measured on discrete CUDA hardware would tell every
-    # 8 GB Mac its 4 GB "VRAM" is too small for an engine that runs fine there.
-    # Different memory model, different (unmeasured) floor; don't guess.
-    if (
-        caps.family in ("cuda", "rocm")
-        and min_vram_gb > 0
-        and 0 < caps.vram_gb < min_vram_gb
-    ):
-        device = caps.device_name or caps.family.upper()
-        return (
-            f"{device} has {caps.vram_gb:.1f} GB VRAM; this engine wants about "
-            f"{min_vram_gb:.0f} GB. It will run, but expect slow generations "
-            f"that may time out. Unload other models before generating, keep "
-            f"the text short, or pick a lighter engine."
-        )
-    return None
+    return low_vram_caveat(caps, min_vram_gb)
 
 
 def resolve_routing(
@@ -208,5 +279,7 @@ def routing_fields(
 
 __all__ = [
     "RoutingStatus", "RoutingResult", "resolve_routing", "routing_fields",
-    "routing_notice", "header_safe_reason",
+    "routing_notice", "header_safe_reason", "low_vram_caveat",
+    "runtime_compute_profile", "runtime_compute_profile_async",
+    "under_provisioned_vram",
 ]

backend/services/hf_revisions.py

@@ -15,6 +15,7 @@ _SHA = re.compile(r"[0-9a-f]{40}\Z")
 CURATED_REVISIONS: dict[str, str] = {
     "facebook/nllb-200-distilled-600M": "f8d333a098d19b4fd9a8b18f94170487ad3f821d",
     "k2-fsa/OmniVoice": "c5fdb5ccb189668d56333f77ba2629f4cd7535f4",
+    "audio-cpp/audio.cpp-gguf": "dc6fecccc2b0c6bdda0a8b2f38fa61394fee0b9c",
     "Systran/faster-whisper-large-v3": "edaa852ec7e145841d8ffdb056a99866b5f0a478",
     "mlx-community/whisper-large-v3-mlx": "49e6aa286ad60c14352c404340ded53710378a11",
     "mlx-community/whisper-large-v3-turbo": "a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb",

backend/services/model_manager.py

@@ -525,6 +525,8 @@ class GpuPoolBusyError(TimeoutError):
 
 def generate_timeout_s(
     text: "str | None", *, engine: object = None, execution_device: "str | None" = None,
+    min_vram_gb: float = 0.0, hardware_family: "str | None" = None,
+    vram_gb: "float | None" = None,
 ) -> float:
     """THE wall-clock execution budget for one synthesis job, scaled to input.
 
@@ -543,23 +545,36 @@ def generate_timeout_s(
     characters past a 1200-character free allowance — generous enough for
     CPU-class hardware, still bounded (a wedged job is caught in minutes, not
     hours).
+
+    #1804: "accelerated" is not one performance class. A card with less VRAM
+    than the engine declares it needs pages to system RAM over PCIe and renders
+    SLOWER than the same machine's CPU would — yet, judged by device family
+    alone, it was handed HALF the CPU budget. That inversion is what three 4 GB
+    reporters hit (#1226 GTX 1650 Ti, #1222 Quadro P2000, #1804 GTX 1650), all
+    on the engine that declares a 6 GB floor. Every layer already knew: routing
+    raises a caveat, the preflight toast warns, and the timeout message names
+    the card. Only the budget ignored it. So an under-provisioned accelerator
+    now floors at the CPU budget — the class of hardware it actually performs
+    like. ``min_vram_gb`` is the engine's declared floor; callers that pass
+    ``engine`` get it read off the engine automatically. Native runtimes pass
+    an explicit ``vram_gb=0`` when their dedicated-memory probe failed; that
+    unknown capacity gets the same conservative CPU-class budget without
+    claiming the card is under-provisioned in user-facing diagnostics.
     """
     base = GPU_JOB_TIMEOUT_S
     try:
         from core.device_caps import detect_host_caps
-        family = execution_device or detect_host_caps().family
+        caps = detect_host_caps()
+        family = execution_device or caps.family
+        if not min_vram_gb and engine is not None:
+            min_vram_gb = float(getattr(engine, "min_vram_gb", 0.0) or 0.0)
         if execution_device is None and engine is not None:
-            from services.engine_routing import resolve_routing
-            compat = getattr(engine, "gpu_compat", None)
-            if compat is None:
-                compat = getattr(type(engine), "gpu_compat", (family, "cpu"))
-            if tuple(compat) == ("cpu",):
-                family = "cpu"
-            else:
-                family = resolve_routing(
-                    compat, detect_host_caps(),
-                    float(getattr(engine, "min_vram_gb", 0.0) or 0.0),
-                )["effective_device"]
+            from services.engine_routing import runtime_compute_profile
+            profile = runtime_compute_profile(engine, caps)
+            family = profile["effective_device"]
+            min_vram_gb = profile["min_vram_gb"]
+            hardware_family = profile.get("runtime_hardware_family")
+            vram_gb = profile.get("runtime_vram_gb")
         universal_override = (
             _GENERATE_TIMEOUT_EXPLICIT
             or GPU_JOB_TIMEOUT_S != _CONFIGURED_GPU_JOB_TIMEOUT_S
@@ -573,6 +588,24 @@ def generate_timeout_s(
         )
         if family == "cpu" and (cpu_explicit or not universal_override):
             base = CPU_JOB_TIMEOUT_S
+        elif not universal_override and family in (
+            "cuda", "rocm", "vulkan", "xpu",
+        ):
+            from services.engine_routing import under_provisioned_vram
+
+            runtime_family = hardware_family or family
+            unknown_dedicated_vram = (
+                min_vram_gb > 0
+                and runtime_family in ("cuda", "rocm", "xpu", "vulkan")
+                and vram_gb is not None
+                and float(vram_gb or 0.0) <= 0
+            )
+            if unknown_dedicated_vram or under_provisioned_vram(
+                caps, min_vram_gb, family=hardware_family, vram_gb=vram_gb,
+            ):
+                # `max`, never a plain assignment: an operator who raised the
+                # accelerated budget above the CPU one must not have it cut.
+                base = max(base, CPU_JOB_TIMEOUT_S)
     except Exception:
         # Device probing is advisory here; the configured universal bound is
         # still safe when a platform probe is unavailable during startup.
@@ -1076,6 +1109,7 @@ def _timeout_guidance(
     """
     family = "cuda"  # conservative default: GPU wording if the probe fails
     device_name, vram_gb = "", 0.0
+    _caps = None  # a failed probe stays None; under_provisioned_vram() reads it safely
     try:
         from core.device_caps import detect_host_caps
         _caps = detect_host_caps()
@@ -1111,7 +1145,7 @@ def _timeout_guidance(
             "compute-bound. For a durable fix try shorter text or a lighter "
             "engine (OmniVoice GGUF and Supertonic-3 are CPU-tuned). If you "
             "expect very long single generations, raise "
-            "OMNIVOICE_GENERATE_TIMEOUT_S."
+            "the compute-time budget in Settings → Performance & Device."
         )
     # #1226/#1222: two users on 4 GB cards were told, generically, that the GPU
     # "is VRAM-starved" — true, but it read as a transient contention problem
@@ -1123,11 +1157,9 @@ def _timeout_guidance(
     # a threshold applied without knowing whose job it is would confidently
     # misdiagnose most of them. And on MPS `vram_gb` is a unified-memory
     # heuristic (RAM/2), not a dedicated pool to compare against.
-    if (
-        min_vram_gb > 0
-        and family in ("cuda", "rocm")
-        and 0 < vram_gb < min_vram_gb
-    ):
+    from services.engine_routing import under_provisioned_vram
+
+    if under_provisioned_vram(_caps, min_vram_gb):
         return common + (
             f"{device_name or 'this GPU'} has {vram_gb:.1f} GB of VRAM and "
             f"this engine wants about {min_vram_gb:.0f} GB — generations here "
@@ -1136,7 +1168,8 @@ def _timeout_guidance(
             f"Supertonic-3 are tuned for small/no GPU) or shorter text; "
             f"Flush caches / Unload the resident model (top toolbar or "
             f"Model Catalogue → Models) frees what little headroom there is. (Raise "
-            f"OMNIVOICE_GENERATE_TIMEOUT_S if you'd rather let long "
+            f"the compute-time budget in Settings → Performance & Device if "
+            f"you'd rather let long "
             f"generations run.)"
         )
     return common + (
@@ -1144,7 +1177,8 @@ def _timeout_guidance(
         "contend for memory). For a durable fix, Flush caches / Unload the "
         "resident model (top toolbar or Model Catalogue → Models) before retrying, "
         "try shorter text, a lighter engine, or set the engine to CPU in "
-        "Model Catalogue → Models. (Raise OMNIVOICE_GENERATE_TIMEOUT_S for very "
+        "Model Catalogue → Models. (Raise the compute-time budget in "
+        "Settings → Performance & Device for very "
         "long single generations.)"
     )
 
@@ -1506,14 +1540,17 @@ def get_best_device():
     # ── DirectML — universal Windows GPU (probe reports this as "cpu") ─
     # Reached only when no torch family was detected (family == "cpu"), which is
     # exactly the DirectML case — the probe classifies DirectML hosts as cpu.
-    try:
-        import torch_directml
-        if torch_directml.device_count() > 0:
-            logger.info("Using DirectML device (GPU %d)", 0)
-            return str(torch_directml.device(0))
-    except ImportError:
-        pass
+    if family == "cpu":
+        try:
+            import torch_directml
+            if torch_directml.device_count() > 0:
+                logger.info("Using DirectML device (GPU %d)", 0)
+                return str(torch_directml.device(0))
+        except ImportError:
+            # DirectML is optional; an absent package leaves CPU available.
+            pass
 
+    # Other families need an explicitly compatible loader (e.g. NPU sidecars).
     return "cpu"
 
 _COMPILE_ERR_MODULE_PREFIXES = ("torch._dynamo", "torch._inductor", "torch.fx", "triton")

backend/services/pronunciation.py

@@ -222,6 +222,46 @@ def entries_for_language(entries, language: Optional[str]) -> dict[str, str]:
     return merged
 
 
+def inert_entries_for_language(entries, language: str | None) -> list[dict]:
+    """Enabled entries that MATCH the language but cannot be applied yet.
+
+    Settings offers three notations — Respelling, IPA, CMU — and only
+    respelling substitutes text today. IPA and CMU rows save cleanly, are
+    validated, get a badge and can be toggled on, and are then dropped before
+    term matching. Nothing downstream reads them.
+
+    That is Phase 1 behaving as designed; the gap is that it is INVISIBLE.
+    "Test a sentence" reported "No entries match; spoken as written" for a
+    term that does match, which is not a degraded answer but a wrong one, and
+    it sent the user off to re-type an entry that was already correct (#1949).
+
+    docs/specs/01-expressive-tts.md asked for exactly the opposite — such
+    entries "passed through and flagged 'phoneme not honored on this
+    engine' (parity-rule: visible degradation)". This is that flag: the
+    caller can now say WHY nothing happened instead of implying nothing
+    matched.
+    """
+    req_prefix = _lang_prefix(language)
+    out: list[dict] = []
+    for e in entries or []:
+        try:
+            if not int(e["enabled"]):
+                continue
+        except (KeyError, IndexError, TypeError, ValueError):
+            continue
+        term = (e["term"] or "").strip()
+        if not term:
+            continue
+        etype = (e["type"] or "respelling").strip().lower()
+        if etype == "respelling":
+            continue
+        scope = (e["language"] or _ALL_LANG).strip() or _ALL_LANG
+        if scope != _ALL_LANG and (req_prefix is None or scope[:2].lower() != req_prefix):
+            continue
+        out.append({"term": term, "type": etype})
+    return out
+
+
 # ── Inline one-off override:  [[term|replacement]]  /  [[replacement]] ─────────
 #
 # Double brackets are unambiguous against the single-bracket grammar

backend/services/segmented_download.py

@@ -29,6 +29,14 @@ import httpx
 _HF_AUTH_HOSTS = ("huggingface.co", "hf.co")
 _DEFAULT_CONNECTIONS = 8
 _MIN_SEGMENT_BYTES = 4 * 1024 * 1024   # don't split below this — overhead > gain
+# Cap on a single segment. Progress is committed to the manifest only when a
+# whole segment lands, so the segment size is also the MOST bytes a dropped
+# connection can throw away. Sizing segments as size/num_connections made that
+# ~100 MB on an 800 MB blob: on a link that drops every ~50 MB no segment ever
+# completed, the manifest was never written, and every retry restarted from
+# zero (#1224 follow-up). Bounded segments turn the same flaky link into steady
+# forward progress.
+_MAX_SEGMENT_BYTES = 16 * 1024 * 1024
 _READ_CHUNK = 1024 * 1024
 
 
@@ -67,8 +75,16 @@ async def _resolve(client: httpx.AsyncClient, url: str, token: Optional[str], ma
 
 
 def _plan_segments(size: int, num_connections: int) -> list[tuple[int, int]]:
+    """Byte ranges to fetch, each at most ``_MAX_SEGMENT_BYTES``.
+
+    ``num_connections`` controls how many run at once (see the semaphore in
+    :func:`segmented_download`), NOT how many segments exist — a large file is
+    split into many bounded segments so each one commits to the manifest
+    quickly and a dropped connection costs at most one segment.
+    """
     n = max(1, min(num_connections, max(1, size // _MIN_SEGMENT_BYTES)))
     step = -(-size // n)  # ceil
+    step = max(_MIN_SEGMENT_BYTES, min(step, _MAX_SEGMENT_BYTES))
     segs = []
     start = 0
     while start < size:
@@ -143,6 +159,10 @@ async def segmented_download(
             _preallocate(part, size)
             segments = [s for s in _plan_segments(size, num_connections) if s not in done]
             lock = asyncio.Lock()
+            # Segments are bounded, so a big file yields many more of them than
+            # there are connections. The semaphore — not the segment count — is
+            # what keeps concurrency at num_connections.
+            sem = asyncio.Semaphore(max(1, num_connections))
 
             async def _fetch(seg: tuple[int, int]):
                 start, end = seg
@@ -169,8 +189,12 @@ async def segmented_download(
                     done.add(seg)
                     _save_done(part, size, done)
 
+            async def _fetch_limited(seg: tuple[int, int]):
+                async with sem:
+                    await _fetch(seg)
+
             if segments:
-                await asyncio.gather(*(_fetch(s) for s in segments))
+                await asyncio.gather(*(_fetch_limited(s) for s in segments))
 
         # ── verify ──────────────────────────────────────────────────────
         actual = os.path.getsize(part)

backend/services/sidecar_install.py

@@ -126,6 +126,33 @@ class SidecarSpec:
     invalidate: Callable[[], None] = field(default=lambda: None)
     # Cheap "is a healthy install already present?" probe (file existence only).
     installed_probe: Callable[[], bool] = field(default=lambda: False)
+    # Extra `uv venv` arguments — an interpreter pin for an upstream that
+    # declares one, e.g. ("--python", "3.10").
+    venv_args: tuple[str, ...] = ()
+    # `uv pip install` target, "{checkout}" substituted. Each upstream installs
+    # differently (editable, editable with an extra, a requirements file, a
+    # constraints file); the default is the editable install IndexTTS uses.
+    install_args: tuple[str, ...] = ("-e", "{checkout}")
+    # Add PyTorch's CUDA index on a CUDA host. Plain PyPI torch is CPU-only on
+    # Windows, and `+cuNNN` local-version pins exist nowhere else.
+    uses_cuda_index: bool = False
+    # Python that proves the venv works; "{checkout}" / "{checkout_repr}"
+    # substituted. None means `import <probe_module>`.
+    probe_code: Optional[str] = None
+    # The file whose presence proves a fetched checkout is the whole
+    # repository. Most upstreams ship a pyproject.toml; Confucius4 ships
+    # only requirements.txt and setup.py.
+    source_manifest: str = "pyproject.toml"
+    # False for an engine that is a PyPI package, not a repository: nothing
+    # is fetched, and the managed root holds only the engine's own venv.
+    has_source: bool = True
+    # Add PyTorch's CPU index on every host, for an engine that only ever
+    # runs torch on the CPU (see core.torch_indexes).
+    cpu_torch_index: bool = False
+    # Can the one-click install work on THIS machine? (ok, reason). Consulted
+    # before an Install button is offered and again when an install starts, so
+    # a host the upstream does not support never gets a job that can only fail.
+    host_supported: Callable[[], tuple[bool, str]] = field(default=lambda: (True, ""))
 
 
 def _indextts_invalidate() -> None:
@@ -138,6 +165,86 @@ def _indextts_installed() -> bool:
     return is_indextts_installed()
 
 
+def _moss_invalidate() -> None:
+    from engines.moss_tts_v15 import bootstrap
+    bootstrap.invalidate()
+
+
+def _moss_installed() -> bool:
+    from engines.moss_tts_v15.bootstrap import is_moss_tts_v15_installed
+    return is_moss_tts_v15_installed()
+
+
+def _confucius4_invalidate() -> None:
+    from engines.confucius4 import bootstrap
+    bootstrap.invalidate()
+
+
+def _confucius4_installed() -> bool:
+    from engines.confucius4.bootstrap import is_confucius4_installed
+    return is_confucius4_installed()
+
+
+def _dots_invalidate() -> None:
+    from engines.dots_tts import bootstrap
+    bootstrap.invalidate()
+
+
+def _dots_installed() -> bool:
+    from engines.dots_tts.bootstrap import is_dots_tts_installed
+    return is_dots_tts_installed()
+
+
+def _host_family() -> str:
+    """The accelerator family this host runs, or "cpu" when it cannot tell."""
+    try:
+        from core.device_caps import detect_host_caps
+        return str(detect_host_caps().family)
+    except Exception:  # noqa: BLE001 — a probe failure must not break installs
+        return "cpu"
+
+
+def _moss_host() -> tuple[bool, str]:
+    if _host_family() == "cuda":
+        return True, ""
+    return False, (
+        "MOSS-TTS-v1.5's one-click install uses its CUDA build of PyTorch, and "
+        "this machine has no NVIDIA GPU available. Its guide covers a manual "
+        "CPU install."
+    )
+
+
+def _dots_host() -> tuple[bool, str]:
+    if sys.platform != "win32":
+        return True, ""
+    return False, (
+        "dots.tts publishes no Windows install. Run VoiceStudio on Linux or "
+        "macOS, or under WSL2, to use it."
+    )
+
+
+def _pockettts_host() -> tuple[bool, str]:
+    import platform
+    if sys.platform == "darwin" and platform.machine().lower() == "x86_64":
+        return False, (
+            "PocketTTS needs a PyTorch version that has no Intel Mac build."
+        )
+    return True, ""
+
+
+def _in_app_env(module: str) -> Callable[[], bool]:
+    """An install made with ``uv sync --extra`` lives in the app's own
+    environment. It counts as installed, so the installer never provisions a
+    second copy over one that works."""
+    def probe() -> bool:
+        import importlib.util
+        try:
+            return importlib.util.find_spec(module) is not None
+        except (ImportError, ValueError):
+            return False
+    return probe
+
+
 SPECS: dict[str, SidecarSpec] = {
     "indextts2": SidecarSpec(
         engine_id="indextts2",
@@ -170,9 +277,165 @@ SPECS: dict[str, SidecarSpec] = {
         invalidate=_indextts_invalidate,
         installed_probe=_indextts_installed,
     ),
+    # Pinned to the upstream commits current on 2026-09-10. Weights are not
+    # fetched here: each engine downloads them into the shared HF cache on its
+    # first synthesis, as its manual install always has.
+    "moss-tts-v15": SidecarSpec(
+        engine_id="moss-tts-v15",
+        display_name="MOSS-TTS-v1.5",
+        repo_url="https://github.com/OpenMOSS/MOSS-TTS.git",
+        tarball_url=(
+            "https://github.com/OpenMOSS/MOSS-TTS/archive/"
+            "934d6826b084c46a0d033402174d5f8ac4ed2519.tar.gz"
+        ),
+        checkout_dirname="MOSS-TTS",
+        env_var="OMNIVOICE_MOSS_TTS_V15_DIR",
+        probe_module="transformers",
+        probe_code="import transformers, torch",
+        source_revision="934d6826b084c46a0d033402174d5f8ac4ed2519",
+        source_required_path="pyproject.toml",
+        venv_args=("--python", "3.11"),
+        install_args=("-e", "{checkout}[torch-runtime]"),
+        uses_cuda_index=True,
+        host_supported=_moss_host,
+        docs_path="docs/engines/moss-tts-v15.md",
+        # ~7 GB CUDA torch venv now, ~16 GB of weights on first synthesis.
+        required_bytes=24 * _GIB,
+        dependency_bytes=8 * _GIB,
+        temporary_free_bytes=8 * _GIB,
+        disk_confidence="estimated",
+        invalidate=_moss_invalidate,
+        installed_probe=_moss_installed,
+    ),
+    "confucius4-tts": SidecarSpec(
+        engine_id="confucius4-tts",
+        display_name="Confucius4-TTS",
+        repo_url="https://github.com/netease-youdao/Confucius4-TTS.git",
+        tarball_url=(
+            "https://github.com/netease-youdao/Confucius4-TTS/archive/"
+            "4fb32c481302d8858c3aec6a1c2a8b4cea8894c0.tar.gz"
+        ),
+        checkout_dirname="Confucius4-TTS",
+        env_var="OMNIVOICE_CONFUCIUS4_TTS_DIR",
+        probe_module="confuciustts",
+        # Upstream is not pip-installable; the package resolves from the
+        # checkout on sys.path, exactly as the engine's sidecar imports it.
+        probe_code="import sys; sys.path.insert(0, {checkout_repr}); import confuciustts",
+        source_revision="4fb32c481302d8858c3aec6a1c2a8b4cea8894c0",
+        # No pyproject.toml upstream: requirements.txt is its manifest.
+        source_manifest="requirements.txt",
+        source_required_path="setup.py",
+        venv_args=("--python", "3.10"),
+        install_args=("-r", "{checkout}/requirements.txt"),
+        # torch==2.7.0: CPU-only from PyPI on Windows; the CUDA index supplies
+        # 2.7.0+cu128, which satisfies the same pin.
+        uses_cuda_index=True,
+        docs_path="docs/engines/confucius4-tts.md",
+        # ~7 GB venv now, ~5 GB of weights on first synthesis.
+        required_bytes=14 * _GIB,
+        dependency_bytes=8 * _GIB,
+        temporary_free_bytes=8 * _GIB,
+        disk_confidence="estimated",
+        invalidate=_confucius4_invalidate,
+        installed_probe=_confucius4_installed,
+    ),
+    "dots-tts": SidecarSpec(
+        engine_id="dots-tts",
+        display_name="dots.tts",
+        repo_url="https://github.com/rednote-hilab/dots.tts.git",
+        tarball_url=(
+            "https://github.com/rednote-hilab/dots.tts/archive/"
+            "32407a55228630475c48ecdb2c4e2c0f9c09e030.tar.gz"
+        ),
+        checkout_dirname="dots.tts",
+        env_var="OMNIVOICE_DOTS_TTS_DIR",
+        probe_module="dots_tts.runtime",
+        source_revision="32407a55228630475c48ecdb2c4e2c0f9c09e030",
+        source_required_path="constraints/recommended.txt",
+        # Upstream requires-python is >=3.10,<3.13.
+        venv_args=("--python", "3.11"),
+        install_args=("-e", "{checkout}", "-c", "{checkout}/constraints/recommended.txt"),
+        host_supported=_dots_host,
+        docs_path="docs/engines/dots-tts.md",
+        # ~7 GB venv now, ~9 GB checkpoint on first synthesis.
+        required_bytes=18 * _GIB,
+        dependency_bytes=8 * _GIB,
+        temporary_free_bytes=8 * _GIB,
+        disk_confidence="estimated",
+        invalidate=_dots_invalidate,
+        installed_probe=_dots_installed,
+    ),
+    # PyPI packages rather than repositories: nothing to clone, and the managed
+    # root holds only the engine's own venv. The pins are the app's own
+    # optional extras (a test ties the two together), so the engine runs the
+    # same wheel whichever way it was installed.
+    "supertonic3": SidecarSpec(
+        engine_id="supertonic3",
+        display_name="Supertonic-3",
+        repo_url="",
+        tarball_url="",
+        checkout_dirname="supertonic3",
+        env_var="OMNIVOICE_SUPERTONIC3_DIR",
+        probe_module="supertonic",
+        has_source=False,
+        venv_args=("--python", "3.11"),
+        install_args=("supertonic==1.3.1",),
+        docs_path="docs/engines/supertonic3.md",
+        # onnxruntime + numpy + huggingface_hub, no torch. The ~400 MB of
+        # weights download on first synthesis into the shared HF cache.
+        required_bytes=1 * _GIB,
+        installed_probe=_in_app_env("supertonic"),
+    ),
+    "pockettts": SidecarSpec(
+        engine_id="pockettts",
+        display_name="PocketTTS",
+        repo_url="",
+        tarball_url="",
+        checkout_dirname="pockettts",
+        env_var="OMNIVOICE_POCKETTTS_DIR",
+        probe_module="pocket_tts",
+        has_source=False,
+        venv_args=("--python", "3.11"),
+        install_args=("pocket-tts==2.1.0",),
+        cpu_torch_index=True,
+        docs_path="docs/engines/pockettts.md",
+        # CPU torch + scipy. The gated weights download on first use.
+        required_bytes=3 * _GIB,
+        installed_probe=_in_app_env("pocket_tts"),
+        host_supported=_pockettts_host,
+    ),
 }
 
 
+class HostUnsupported(RuntimeError):
+    """The one-click install cannot work on this machine. The message is a
+    VoiceStudio-owned sentence from the spec, safe to show the user."""
+
+
+def host_support(spec: SidecarSpec) -> tuple[bool, str]:
+    """Whether *spec*'s install can work here. A probe that raises counts as
+    unsupported: offering a button that fails is worse than not offering it."""
+    try:
+        ok, why = spec.host_supported()
+    except Exception:  # noqa: BLE001
+        return False, (
+            f"Could not check whether {spec.display_name} can be installed on "
+            f"this machine. Its guide ({spec.docs_path}) has the manual steps."
+        )
+    return bool(ok), (why or "")
+
+
+def installable_engine_ids() -> frozenset[str]:
+    """Engines that get an Install button on THIS host."""
+    return frozenset(eid for eid, spec in SPECS.items() if host_support(spec)[0])
+
+
+def _expand(value: str, checkout: Path) -> str:
+    return value.replace("{checkout_repr}", repr(str(checkout))).replace(
+        "{checkout}", str(checkout)
+    )
+
+
 def get_spec(engine_id: str) -> Optional[SidecarSpec]:
     return SPECS.get(engine_id)
 
@@ -196,6 +459,20 @@ def managed_checkout(spec: SidecarSpec) -> Path:
     return managed_root(spec) / spec.checkout_dirname
 
 
+def engine_venv_python(env_var: str) -> Optional[Path]:
+    """The interpreter of the install *env_var* points at, if it has one.
+
+    For engines that can live in the app's environment or in a venv of their
+    own (PocketTTS, Supertonic-3): they prefer their own, and fall back to the
+    app's interpreter for an install made with ``uv sync --extra``.
+    """
+    env_dir = os.environ.get(env_var)
+    if not env_dir:
+        return None
+    py = _venv_python(Path(env_dir) / ".venv")
+    return py if py.is_file() else None
+
+
 def _legacy_managed_checkouts(spec: SidecarSpec) -> tuple[Path, ...]:
     """App-owned predecessor checkouts retained during in-place upgrades."""
     if spec.engine_id == "indextts2":
@@ -272,13 +549,20 @@ def _default_uv_cache_root() -> Path:
     return Path(os.environ.get("XDG_CACHE_HOME") or Path.home() / ".cache") / "uv"
 
 
-def uv_subprocess_env(cache_parent: Path) -> "dict[str, str] | None":
+def uv_subprocess_env(cache_parent: Path) -> "dict[str, str]":
     """Environment for ``uv`` subprocesses that install into *cache_parent*'s volume.
 
-    Returns ``None`` (inherit the parent environment untouched) when uv's
-    default cache already shares a volume with *cache_parent* or the user
-    pinned both variables themselves. Otherwise returns a copy of
-    ``os.environ`` with the *unset* one(s) of ``UV_CACHE_DIR`` /
+    Always a copy of ``os.environ`` with ``UV_NO_CONFIG=1``: an engine's
+    install resolves its own requirements, never VoiceStudio's. The backend
+    runs inside the app's tree, so uv would otherwise discover the app's
+    ``pyproject.toml`` and apply its ``[tool.uv] constraint-dependencies``
+    (``torch==2.8.0``) to the engine's venv. An engine pinning another torch
+    (MOSS-TTS-v1.5, Confucius4) could then never resolve, and one that pins
+    none got the app's torch instead of its own. Mirrors still apply: they
+    arrive as ``UV_INDEX_URL``, an environment variable, not a config file.
+
+    When uv's default cache is on another volume than *cache_parent*, the
+    copy also places the *unset* one(s) of ``UV_CACHE_DIR`` /
     ``UV_PYTHON_INSTALL_DIR`` placed inside *cache_parent*, so downloads, the
     unpacked wheel cache, managed Pythons, and the venv all stay on the
     target volume — and same-volume hardlink installs work again. The two
@@ -291,17 +575,15 @@ def uv_subprocess_env(cache_parent: Path) -> "dict[str, str] | None":
     pass the directory that should hold the shared ``.uv-cache`` — typically
     the common parent of the engine venvs on that volume.
     """
-    if _same_volume(cache_parent, _default_uv_cache_root()):
-        return None
     env = dict(os.environ)
-    overrode = False
+    env["UV_NO_CONFIG"] = "1"
+    if _same_volume(cache_parent, _default_uv_cache_root()):
+        return env
     if not env.get("UV_CACHE_DIR"):  # explicit user choice always wins
         env["UV_CACHE_DIR"] = str(Path(cache_parent) / ".uv-cache")
-        overrode = True
     if not env.get("UV_PYTHON_INSTALL_DIR"):
         env["UV_PYTHON_INSTALL_DIR"] = str(Path(cache_parent) / ".uv-python")
-        overrode = True
-    return env if overrode else None
+    return env
 
 
 # ── Disk preflight ─────────────────────────────────────────────────────────
@@ -520,9 +802,13 @@ def _healthy(spec: SidecarSpec) -> bool:
         return False
     if not _venv_python(checkout / ".venv").is_file():
         return False
-    if spec.weights_repo_id and not _weights_present(spec):
-        return False
-    return True
+    if spec.weights_repo_id:
+        return _weights_present(spec)
+    # Nothing downloaded after the dependencies proves they finished; only
+    # the marker the import probe writes does. IndexTTS (weights) predates
+    # the marker and keeps its own check, so no existing install is asked
+    # to reinstall.
+    return (checkout / _INSTALL_COMPLETE_MARKER).is_file()
 
 
 def _persist(spec: SidecarSpec) -> None:
@@ -548,6 +834,9 @@ def start_install(engine_id: str) -> dict:
     spec = get_spec(engine_id)
     if spec is None:
         raise KeyError(engine_id)
+    ok, why = host_support(spec)
+    if not ok:
+        raise HostUnsupported(why)
     with _jobs_lock:
         existing = _jobs.get(engine_id)
         if existing and existing["state"] == "running":
@@ -679,6 +968,11 @@ def _step_preflight(spec: SidecarSpec, job: dict) -> None:
 def _step_fetch_source(spec: SidecarSpec, job: dict) -> None:
     step = _job_step(job, "fetch_source")
     checkout = managed_checkout(spec)
+    if not spec.has_source:
+        checkout.mkdir(parents=True, exist_ok=True)
+        step["state"] = "done"
+        step["detail"] = "PyPI package, no source to fetch"
+        return
     if _source_present(spec, checkout):
         step["state"] = "done"
         step["detail"] = "source already present"
@@ -717,7 +1011,7 @@ def _step_fetch_source(spec: SidecarSpec, job: dict) -> None:
     _fetch_tarball(spec, job, checkout)
     if not _source_layout_ok(spec, checkout):
         raise _StepError(
-            f"Fetched source at {checkout} has no pyproject.toml — the download "
+            f"Fetched source at {checkout} has no {spec.source_manifest} — the download "
             "appears incomplete or the upstream layout changed.",
             "Re-run the install; if it keeps failing, clone the repository "
             f"manually and set {spec.env_var} to the clone (see the engine docs).",
@@ -727,10 +1021,14 @@ def _step_fetch_source(spec: SidecarSpec, job: dict) -> None:
 
 
 _SOURCE_REVISION_MARKER = ".voicestudio_source_revision"
+# Written once the import probe passes. For an engine with no weights
+# download, the venv interpreter existing proves nothing: a dependency
+# install that died halfway leaves one behind.
+_INSTALL_COMPLETE_MARKER = ".voicestudio_install_complete"
 
 
 def _source_layout_ok(spec: SidecarSpec, checkout: Path) -> bool:
-    if not (checkout / "pyproject.toml").is_file():
+    if not (checkout / spec.source_manifest).is_file():
         return False
     return not spec.source_required_path or (checkout / spec.source_required_path).is_file()
 
@@ -743,6 +1041,8 @@ def _write_source_marker(spec: SidecarSpec, checkout: Path) -> None:
 
 
 def _source_present(spec: SidecarSpec, checkout: Path) -> bool:
+    if not spec.has_source:
+        return checkout.is_dir()
     if not _source_layout_ok(spec, checkout):
         return False
     if not spec.source_revision:
@@ -836,8 +1136,8 @@ def _step_create_venv(spec: SidecarSpec, job: dict) -> None:
     # uv_subprocess_env. The cache parent is the shared engines root, so
     # every sidecar engine reuses one cache.
     uv_env = uv_subprocess_env(Path(DATA_DIR) / "engines")
-    rc = _run_logged(job, [uv, "venv", str(venv_dir)], timeout=_UV_VENV_TIMEOUT_S,
-                     env=uv_env)
+    rc = _run_logged(job, [uv, "venv", str(venv_dir), *spec.venv_args],
+                     timeout=_UV_VENV_TIMEOUT_S, env=uv_env)
     if rc != 0 or not py.is_file():
         raise _StepError(
             f"uv venv failed (exit {rc}) at {venv_dir}.",
@@ -857,17 +1157,28 @@ def _step_install_deps(spec: SidecarSpec, job: dict) -> None:
     """
     checkout = managed_checkout(spec)
     py = _venv_python(checkout / ".venv")
+    # A reinstall that fails must not leave the previous run's marker.
+    (checkout / _INSTALL_COMPLETE_MARKER).unlink(missing_ok=True)
     uv = _locate_uv()
     _log(job, f"Installing {spec.display_name} into its venv (this can take several minutes) …")
+    target = [_expand(arg, checkout) for arg in spec.install_args]
+    if spec.cpu_torch_index:
+        from core.torch_indexes import UV_PIP_CPU_ARGS
+        target += list(UV_PIP_CPU_ARGS)
+    elif spec.uses_cuda_index and _host_family() == "cuda":
+        from core.torch_indexes import UV_PIP_CU128_ARGS
+        target += list(UV_PIP_CU128_ARGS)
+    # Always `--python <this engine's venv>`: the install can only ever land in
+    # the venv this engine owns, never the app's interpreter.
     rc = _run_logged(
         job,
-        [uv, "pip", "install", "--python", str(py), "-e", str(checkout)],
+        [uv, "pip", "install", "--python", str(py), *target],
         timeout=_UV_PIP_INSTALL_TIMEOUT_S,
         env=uv_subprocess_env(Path(DATA_DIR) / "engines"),
     )
     if rc != 0:
         raise _StepError(
-            f"uv pip install -e failed (exit {rc}).",
+            f"uv pip install failed (exit {rc}).",
             "Usually a network hiccup — re-run the install to resume. Behind a "
             "proxy, set HTTPS_PROXY in Settings → Environment first.",
         )
@@ -879,8 +1190,13 @@ def _step_verify(spec: SidecarSpec, job: dict) -> None:
     py = _venv_python(checkout / ".venv")
     _log(job, f"Verifying `import {spec.probe_module}` inside the venv …")
     try:
+        probe = (
+            _expand(spec.probe_code, checkout)
+            if spec.probe_code
+            else f"import {spec.probe_module}"
+        )
         proc = subprocess.run(
-            [str(py), "-c", f"import {spec.probe_module}"],
+            [str(py), "-c", probe],
             capture_output=True, timeout=_IMPORT_PROBE_TIMEOUT_S,
         )
     except (subprocess.TimeoutExpired, OSError) as exc:
@@ -898,6 +1214,7 @@ def _step_verify(spec: SidecarSpec, job: dict) -> None:
             "the engine docs.",
         )
     _job_step(job, "verify")["detail"] = f"import {spec.probe_module} OK"
+    (checkout / _INSTALL_COMPLETE_MARKER).write_text(f"{spec.probe_module}\n", encoding="utf-8")
     _log(job, "Venv verified.")
 
 

backend/services/subprocess_backend.py

@@ -385,6 +385,10 @@ class SubprocessBackend(TTSBackend):
 
     def __init__(self) -> None:
         self._proc: Optional[subprocess.Popen] = None
+        # A failed bounded reap must retain ownership and forbid reuse. This
+        # lock is separate from _lock: the receive owner joins its watchdog.
+        self._timeout_quarantine: list[subprocess.Popen] = []
+        self._timeout_quarantine_lock = threading.Lock()
         # Single lock serialises spawn + every send/recv pair so two threads
         # can't interleave half-frames on the same pipe.
         self._lock = threading.Lock()
@@ -451,6 +455,10 @@ class SubprocessBackend(TTSBackend):
     def _spawn(self) -> None:
         """Launch the sidecar if not already running. Blocks on the ready
         handshake. Caller must hold self._lock."""
+        if not self._retry_timeout_cleanup():
+            raise RuntimeError(
+                f"{self.id} sidecar is still stopping after a timeout; retry once it exits"
+            )
         if self._proc is not None and self._proc.poll() is None:
             return  # already up
 
@@ -542,6 +550,7 @@ class SubprocessBackend(TTSBackend):
         """Idempotent. Sends {op:shutdown}; falls back to terminate/kill."""
         proc = self._proc
         if proc is None:
+            self._retry_timeout_cleanup()
             return
         try:
             try:
@@ -577,6 +586,7 @@ class SubprocessBackend(TTSBackend):
                         pass
         finally:
             self._proc = None
+            self._retry_timeout_cleanup()
 
     def _force_kill(self) -> None:
         """Internal: kill a sidecar that never reached the ready state."""
@@ -777,38 +787,59 @@ class SubprocessBackend(TTSBackend):
         return msg
 
     def _recv_with_timeout(self, timeout_s: float) -> Optional[dict]:
-        """Recv that aborts if the sidecar goes silent.
+        """Read one frame, finishing timeout cleanup before the caller can retry.
 
-        Implemented by polling the proc for liveness with a deadline. We
-        don't block on a `select` of the pipe because Windows can't select
-        on subprocess pipes — keeping the implementation cross-platform
-        means a simpler polling loop here.
+        A watchdog closes the pipe on timeout; Windows cannot select on pipes.
+        EOF alone does not prove the owned process/supervisor has exited.
         """
         # On Unix we could use selectors; on Windows the pipe is not
         # selectable. Use a watchdog thread that kills the sidecar on
         # timeout — that triggers EOF on stdout, so _recv returns None
         # and the caller raises.
-        watchdog = threading.Timer(timeout_s, self._timeout_kill)
+        proc = self._proc
+        watchdog = threading.Timer(timeout_s, self._timeout_kill, args=(proc,))
         watchdog.daemon = True
         watchdog.start()
         try:
             return self._recv()
         finally:
             watchdog.cancel()
+            # cancel() cannot stop an already-running callback. Finish its
+            # bounded reap before another receive or generation starts.
+            watchdog.join()
             self._touch()  # any reply (or attempt) counts as recent activity
 
-    def _timeout_kill(self) -> None:
-        proc = self._proc
+    def _timeout_kill(self, proc: Optional[subprocess.Popen]) -> None:
+        """Kill only the child this receive captured, then reap its owner."""
         if proc is None:
             return
+        logger.error("[%s] sidecar exceeded recv timeout; killing", self.id)
         try:
-            logger.error(
-                "[%s] sidecar exceeded recv timeout; killing",
-                self.id,
-            )
             proc.kill()
         except Exception:
+            # A raced exit can make kill fail, but its owner still needs reaping.
             pass
+        try:
+            proc.wait(timeout=2)
+        except Exception:
+            # Do not discard a possibly live owner, or replace a newer _proc.
+            with self._timeout_quarantine_lock:
+                if not any(item is proc for item in self._timeout_quarantine):
+                    self._timeout_quarantine.append(proc)
+        else:
+            with self._timeout_quarantine_lock:
+                self._timeout_quarantine = [
+                    item for item in self._timeout_quarantine if item is not proc
+                ]
+
+    def _retry_timeout_cleanup(self) -> bool:
+        """Retry bounded cleanup, retaining every owner that could still be live."""
+        with self._timeout_quarantine_lock:
+            pending = tuple(self._timeout_quarantine)
+        for proc in pending:
+            self._timeout_kill(proc)
+        with self._timeout_quarantine_lock:
+            return not self._timeout_quarantine
 
     # ── stderr drain ───────────────────────────────────────────────────────
 

backend/services/text_normalization.py

@@ -112,6 +112,13 @@ _FULL_NAME_TO_CODE = {
     "vietnamese": "vi",
     "kazakh": "kz",
     "standard arabic": "ar",
+    # Below: inert for num2words (absent from _NUM2WORDS_LANGS, which reads
+    # digits natively for these scripts), present so _plain_lang_code can
+    # resolve them for the digit-range rule.
+    "korean": "ko",
+    "japanese": "ja",
+    "chinese": "zh",
+    "mandarin chinese": "zh",
 }
 
 # ISO codes whose num2words locale name differs.
@@ -178,6 +185,82 @@ def _num2words_lang(language: Optional[str]) -> Optional[str]:
     return None
 
 
+def _plain_lang_code(language: Optional[str]) -> Optional[str]:
+    """Resolve a request language to a bare ISO code, with no num2words gate.
+
+    :func:`_num2words_lang` answers "may I call num2words for this?" and so
+    returns ``None`` for ko/ja/zh/th/vi. Rules that are not num2words-backed
+    need the code itself, which is what this returns.
+    """
+    if not language:
+        return None
+    s = str(language).strip().lower()
+    if not s or s == "auto":
+        return None
+    code = _FULL_NAME_TO_CODE.get(s)
+    if code:
+        return code
+    m = _ISO_CODE_RE.match(s)
+    if m:
+        return _ISO_ALIASES.get(m.group(1), m.group(1))
+    return None
+
+
+# ── Digit ranges ─────────────────────────────────────────────────────────────
+# "20~30" loses its separator at the engine and reads as ONE number: OmniVoice
+# says "이십삼" (23) for "20~30초". Speak the separator instead. Verified by
+# rendering each form and transcribing it back (ko, OmniVoice):
+#     "20~30초"        → heard "23초"          ✗
+#     "20에서 30초"     → heard "20에서 30초"    ✓
+# Only the tilde family is rewritten — those are unambiguously range marks
+# between digits. An ASCII hyphen is left alone on purpose: it also spells
+# dates, phone numbers and product codes, where "to" would be wrong.
+#: Spacing is part of the form, not decoration: a Korean postposition binds to
+#: the numeral ("20에서 30"), Japanese and Chinese set no spaces at all, and
+#: English needs them on both sides.
+_RANGE_FORM = {
+    "ko": "{a}에서 {b}",
+    "ja": "{a}から{b}",
+    "zh": "{a}到{b}",
+    "en": "{a} to {b}",
+}
+
+#: ASCII tilde, wave dash, fullwidth tilde — Japanese and Korean IMEs emit the
+#: latter two, so all three have to match.
+#:
+#: Match complete signed/decimal endpoints; reject partial numbers and product
+#: codes while allowing adjacent CJK units. Guard all tilde forms so malformed
+#: chains cannot be partially rewritten, including when their separators have
+#: whitespace around them.
+_RANGE_MARKS = "~\u301c\uff5e"
+_RANGE_ENDPOINT = r"[+-]?(?:\d{1,6}(?:\.\d{1,6})?|\.\d{1,6})"
+_NUM_RANGE_RE = re.compile(
+    rf"(?<![\d.,A-Za-z+{_RANGE_MARKS}-])({_RANGE_ENDPOINT})"
+    rf"\s*[{_RANGE_MARKS}]\s*({_RANGE_ENDPOINT})"
+    rf"(?![\d.,A-Za-z+{_RANGE_MARKS}-])"
+)
+
+
+def _speak_number_ranges(text: str, lang: str) -> str:
+    """Speak complete tilde ranges only for languages with a verified form."""
+    form = _RANGE_FORM.get(lang)
+    if not form:
+        return text
+
+    def replace(match: re.Match) -> str:
+        before, after = match.start() - 1, match.end()
+        while before >= 0 and text[before].isspace():
+            before -= 1
+        while after < len(text) and text[after].isspace():
+            after += 1
+        if ((before >= 0 and text[before] in _RANGE_MARKS)
+                or (after < len(text) and text[after] in _RANGE_MARKS)):
+            return match.group(0)
+        return form.format(a=match.group(1), b=match.group(2))
+
+    return _NUM_RANGE_RE.sub(replace, text)
+
+
 # ── Universal safety filters (all languages) ─────────────────────────────────
 
 # Zero-width & bidi controls, C0/C1 controls (except \t \n \r), BOM, U+FFFD.
@@ -487,6 +570,11 @@ def normalize_text(text: str, language: Optional[str] = None) -> str:
     if not text:
         return text or ""
     out = _safety_filters(text)
+    # Runs outside the num2words gate below: ko/ja/zh keep their digits (that
+    # gate returns None for them) but still need the range mark spoken.
+    plain = _plain_lang_code(language)
+    if plain:
+        out = _outside_brackets(out, lambda t: _speak_number_ranges(t, plain))
     lang = _num2words_lang(language)
     if lang:
         if lang in _ABBREV_COMPILED:

backend/services/token_resolver.py

@@ -4,7 +4,7 @@ Resolution priority (highest → lowest):
 
   1. app    — `settings_store.get_hf_token()` (encrypted in SQLite)
   2. env    — `HF_TOKEN` or the legacy `HUGGING_FACE_HUB_TOKEN` env var
-  3. hf-cli — `huggingface_hub.get_token()` (canonical ~/.cache/huggingface/token)
+  3. hf-cli — the selected local Hub token file (`HF_TOKEN_PATH`)
 
 For each candidate, the resolver calls `huggingface_hub.whoami(token=...)`
 to verify the token is live; any HTTP error (401, 403, network) skips to
@@ -20,6 +20,7 @@ from __future__ import annotations
 import hashlib
 import logging
 import os
+from pathlib import Path
 import threading
 import time
 from dataclasses import dataclass
@@ -46,7 +47,7 @@ class SourceState:
     set: bool
     masked: Optional[str]
     whoami_user: Optional[str]
-    whoami_ok: bool
+    whoami_ok: Optional[bool]
 
 
 # ── module-level cache ────────────────────────────────────────────────────
@@ -79,19 +80,26 @@ def _read_app() -> Optional[str]:
         return None
 
 
+def _clean_token(value: Optional[str]) -> Optional[str]:
+    if not value:
+        return None
+    return value.replace("\r", "").replace("\n", "").strip() or None
+
+
 def _read_env() -> Optional[str]:
     # HF docs explicitly accept either name; user may have either exported.
     val = os.environ.get("HF_TOKEN") or os.environ.get("HUGGING_FACE_HUB_TOKEN")
-    return val or None
+    return _clean_token(val)
 
 
 def _read_hf_cli() -> Optional[str]:
     try:
-        import huggingface_hub
-        tok = huggingface_hub.get_token()
-        return tok or None
+        from huggingface_hub import constants
+        return _clean_token(Path(constants.HF_TOKEN_PATH).read_text(encoding="utf-8"))
+    except FileNotFoundError:
+        return None
     except Exception:
-        logger.exception("huggingface_hub.get_token failed")
+        logger.warning("Could not read the local Hugging Face token file")
         return None
 
 
@@ -184,17 +192,17 @@ def on_401(active_source: Source) -> Optional[ResolvedToken]:
     return resolve(skip=frozenset({active_source}))
 
 
-def state() -> dict:
+def state(*, validate: bool = False) -> dict:
     """Return one SourceState per priority position so the Settings UI can
     render the cascade table. Includes a masked token + whoami result;
-    never includes the raw token."""
+    never includes the raw token. Reads are local unless validation is explicitly requested."""
     rows: list[SourceState] = []
     active: Optional[Source] = None
     for source in _PRIORITY:
         token = _READERS[source]()
         if token:
-            username = _validate(source, token)
-            ok = username is not None
+            username = _validate(source, token) if validate else None
+            ok = (username is not None) if validate else None
             rows.append(SourceState(
                 source=source,
                 set=True,
@@ -249,15 +257,30 @@ def save_app_token(token: str) -> None:
     invalidate_cache()
 
 
+def clear_hf_cli_tokens() -> None:
+    """Remove recognized Hub token files without refreshing or revoking tokens."""
+    from core.config import HF_CLI_TOKEN_PATHS
+    from huggingface_hub import constants
+
+    # Hub's active path remains authoritative if imported before app config.
+    paths = set(HF_CLI_TOKEN_PATHS) | {constants.HF_TOKEN_PATH}
+    failed = False
+    for token_path in paths:
+        path = Path(token_path)
+        for target in (path, path.parent / "stored_tokens"):
+            try:
+                target.unlink(missing_ok=True)
+            except OSError:
+                failed = True
+    invalidate_cache()
+    if failed:
+        raise OSError("Could not clear all local Hugging Face token files")
+
+
 def clear_app_token(also_clear_hf_cli: bool = False) -> None:
-    """Remove from the encrypted settings store; optionally also call
-    `huggingface_hub.logout()` to clear the canonical HF file."""
+    """Clear the encrypted app token, optionally recognized local Hub files."""
     from services import settings_store
     settings_store.clear_hf_token()
     if also_clear_hf_cli:
-        try:
-            import huggingface_hub
-            huggingface_hub.logout()
-        except Exception:
-            logger.exception("huggingface_hub.logout failed (non-fatal)")
+        clear_hf_cli_tokens()
     invalidate_cache()

backend/services/translation_engines.py

@@ -17,6 +17,8 @@ from __future__ import annotations
 import asyncio
 import importlib
 import logging
+import functools
+import re
 import os
 import shutil
 import subprocess
@@ -91,6 +93,8 @@ REGISTRY: dict[str, dict] = {
         "probe_module": "openai",
         "category": "llm",
         "needs_key": True,
+        # A core dependency: Settings → LLM Providers uses it too.
+        "builtin": True,
         "notes": (
             "Uses the LLM provider you configure in Settings → LLM Providers "
             "(route it via the 'Dub translation' skill in Settings → LLM Skills): "
@@ -181,6 +185,65 @@ def list_engines() -> list[dict]:
     return out
 
 
+def _normalize(name: str) -> str:
+    """A distribution name in PEP 503 form (deep_translator == deep-translator)."""
+    return re.sub(r"[-_.]+", "-", name).lower()
+
+
+@functools.lru_cache(maxsize=1)
+def _app_dependency_names() -> frozenset[str]:
+    """Distribution names VoiceStudio itself requires, normalized.
+
+    Read from the installed package metadata, so it follows the lockfile with
+    no second list to keep in step. Without metadata this guards nothing
+    rather than failing.
+    """
+    try:
+        from importlib.metadata import requires
+
+        reqs = requires("omnivoice") or []
+    except Exception:  # noqa: BLE001
+        return frozenset()
+    names = set()
+    for req in reqs:
+        if "extra ==" in req:
+            continue
+        names.add(_normalize(re.split(r"[\s;<>=!~\[@(]", req, maxsplit=1)[0]))
+    return frozenset(names)
+
+
+def uninstall_blocker(engine_id: str) -> "tuple[int, str] | None":
+    """Why removing this engine's package would break something, or None.
+
+    `pip uninstall` acts on the app's own environment. A package VoiceStudio
+    depends on (openai, argostranslate) would break the app, and a package
+    other translation engines share (deep_translator backs four) would break
+    those engines too.
+    """
+    entry = REGISTRY.get(engine_id)
+    pkg = entry.get("pip_package") if entry else None
+    if not pkg:
+        return None
+    if _normalize(pkg) in _app_dependency_names():
+        return 400, (
+            f"{entry['display_name']} uses {pkg}, which VoiceStudio itself "
+            "depends on. Uninstalling it would break the app."
+        )
+    sharing = [
+        other["display_name"]
+        for other_id, other in REGISTRY.items()
+        if other_id != engine_id
+        and other.get("pip_package")
+        and _normalize(other["pip_package"]) == _normalize(pkg)
+    ]
+    if sharing:
+        return 409, (
+            f"{entry['display_name']} shares {pkg} with {', '.join(sharing)}. "
+            "Uninstalling it would stop those working too."
+        )
+    return None
+
+
 def get_engine(engine_id: str) -> dict | None:
     return REGISTRY.get(engine_id)
 

backend/services/tts_backend.py

@@ -300,6 +300,31 @@ class TTSBackend(ABC):
     #: 0 means "no meaningful floor" (CPU-class engines) and never warns.
     min_vram_gb: float = 0.0
 
+    @classmethod
+    def runtime_compute_profile(cls, caps) -> dict:
+        """Resolved compute metadata for this engine on the current host.
+
+        Most engines have one implementation whose static declarations are
+        sufficient. Native adapters may override this single hook when the
+        installed executable determines both the available runtimes and the
+        device actually selected.
+        """
+        from services.engine_routing import resolve_routing
+
+        gpu_compat = tuple(getattr(cls, "gpu_compat", ("cpu",)))
+        min_vram_gb = float(getattr(cls, "min_vram_gb", 0.0) or 0.0)
+        return {
+            "gpu_compat": gpu_compat,
+            "min_vram_gb": min_vram_gb,
+            **resolve_routing(gpu_compat, caps, min_vram_gb),
+            "runtime_backend": None,
+            "runtime_device_index": None,
+            "runtime_device_name": None,
+            "runtime_hardware_family": None,
+            "runtime_vram_gb": None,
+            "runtime_device_verified": None,
+        }
+
     #: True when generation allocates in ANOTHER process — a dedicated-venv
     #: sidecar (SubprocessBackend) or a spawned binary (omnivoice-gguf).
     #: Parent-process accelerator counters cannot see those allocations, so
@@ -566,7 +591,12 @@ def _get_clone_prompt(
 ):
     """Return a cached/precomputed ``VoiceClonePrompt`` for
     (ref_audio, ref_text, preprocess_prompt), or ``None`` to fall back to the
-    inline ref path. Never raises.
+    inline ref path.
+
+    Raises only on a device OOM that survives a cache-drop retry (#1790): the
+    inline path is the same allocation on the same device, so falling back to
+    it after an OOM cannot succeed and has been observed taking the whole
+    process down instead. Every other failure still falls back silently.
 
     ``store=False`` still *reads* the cache (a hit is free) but never inserts:
     it exists for single-use references — a dub's per-segment ref clips are each
@@ -596,10 +626,43 @@ def _get_clone_prompt(
                 ref_audio, ref_text=ref_text, preprocess_prompt=preprocess_prompt
             )
         except Exception as e:  # noqa: BLE001 — fall back, never break synthesis
-            logger.warning(
-                "voice-clone prompt precompute failed; using inline ref: %s", e
-            )
-            return None
+            # #1790/#1777: a GPU OOM is the one failure this fallback cannot
+            # absorb. `generate()`'s inline ref path runs the SAME encode on the
+            # SAME device — the docstring above says so, because producing
+            # identical output is the point — so returning None after an OOM
+            # guarantees a second OOM moments later, on a device with even less
+            # headroom than the first attempt found. Both reporters' backends
+            # then died with a Windows access violation (exit code
+            # -1073741819) seconds after this exact log line, mid-generation on
+            # a GPU that had just refused an 86 MiB allocation.
+            #
+            # An OOM here is also the most recoverable kind: the allocator is
+            # typically holding reserved-but-unallocated blocks (#1790's own
+            # log reports 90 MiB reserved against an 86 MiB request). Drop them
+            # and try once more. If it still will not fit, raise — the failure
+            # layer turns a device OOM into the actionable GPU_OOM message
+            # ("close other GPU-heavy apps or unload models…"), which is a far
+            # better answer than walking into a native fault.
+            from core.failure import is_gpu_oom
+
+            if is_gpu_oom(e):
+                logger.warning(
+                    "voice-clone prompt precompute hit a device OOM (%s) — "
+                    "releasing allocator caches and retrying once", e,
+                )
+                try:
+                    from services.model_manager import free_vram
+                    free_vram()
+                except Exception:  # noqa: BLE001 — reclaim is best-effort
+                    logger.debug("VRAM reclaim before OOM retry failed", exc_info=True)
+                prompt = model.create_voice_clone_prompt(
+                    ref_audio, ref_text=ref_text, preprocess_prompt=preprocess_prompt
+                )
+            else:
+                logger.warning(
+                    "voice-clone prompt precompute failed; using inline ref: %s", e
+                )
+                return None
         if store:
             _prompt_disk_save(key, prompt)
     if not store:
@@ -2197,6 +2260,13 @@ _LAZY_REGISTRY: dict[str, tuple[str, str]] = {
     # 2026-07-02 (CPU, Apple Silicon; 22.05 kHz output). Gated behind
     # OMNIVOICE_CONFUCIUS4_TTS_DIR so it's inert until enabled.
     "confucius4-tts": ("engines.confucius4", "Confucius4Backend"),
+    # audio.cpp (0xShug0/audio.cpp) — pure-C++ ggml runtime, no Python venv.
+    # v1 serves Breeze-TTS-2 (en+zh, clone+design) through a parent-managed
+    # audiocpp_server over loopback HTTP. Gated behind a server binary
+    # (OMNIVOICE_AUDIOCPP_BIN) so it's inert until enabled. Lazy for the
+    # same import-cycle reason as the entries above (engines.audiocpp
+    # imports services.tts_backend for TTSBackend).
+    "audiocpp": ("engines.audiocpp", "AudioCPPBackend"),
 }
 
 
@@ -2298,9 +2368,10 @@ _INSTALL_HINTS: dict[str, str] = {
     "omnivoice-gguf":"Bundled — runs the C++ omnivoice-tts binary in bin/. Quants download lazily from Serveurperso/OmniVoice-GGUF on first generate.",
     "supertonic3":   "uv sync --extra supertonic  (CPU-only ONNX, 31 langs, ~400 MB model on first use; OpenRAIL-M model license)",
     "pockettts":     "uv sync --extra pockettts  (Kyutai, CPU-only, ~100 MB model on first use; MIT code + CC-BY-4.0 weights; HF-gated, review terms and set HF_TOKEN)",
-    "moss-tts-v15":  "git clone OpenMOSS/MOSS-TTS + set OMNIVOICE_MOSS_TTS_V15_DIR  (own venv, transformers==5.0; 8B, ~16 GB weights; CUDA/CPU, no MPS; Apache-2.0)",
+    "moss-tts-v15":  "git clone OpenMOSS/MOSS-TTS + set OMNIVOICE_MOSS_TTS_V15_DIR  (own venv, transformers==5.0; 8B, ~16 GB weights; CUDA/ROCm/XPU/NPU/CPU, no MPS; Apache-2.0)",
     "dots-tts":      "git clone rednote-hilab/dots.tts + set OMNIVOICE_DOTS_TTS_DIR  (own venv, transformers==4.57; 2B, ~9 GB weights; CUDA/CPU, Linux/macOS only — no Windows; Apache-2.0)",
-    "confucius4-tts":"git clone netease-youdao/Confucius4-TTS + set OMNIVOICE_CONFUCIUS4_TTS_DIR  (own Python 3.10 venv; 14-lang cross-lingual zero-shot clone; ~5 GB weights auto-download; CUDA/CPU, no MPS; Apache-2.0)",
+    "confucius4-tts":"git clone netease-youdao/Confucius4-TTS + set OMNIVOICE_CONFUCIUS4_TTS_DIR  (own Python 3.10 venv; 14-lang cross-lingual zero-shot clone; ~5 GB weights auto-download; CUDA/ROCm/XPU/NPU/CPU, no MPS; Apache-2.0)",
+    "audiocpp":     "download the matching audio.cpp v0.7.2 prebuilt + set OMNIVOICE_AUDIOCPP_BIN, then explicitly install Breeze-TTS-2 in Model Catalogue → Models  (native CPU/Vulkan/CUDA/Metal GGUF server, no Python; en+zh clone+design; ~4.73 GiB; weights research/non-commercial only)",
 }
 
 
@@ -2321,6 +2392,51 @@ _SETUP_SNIPPETS: dict[str, str] = {
 }
 
 
+# Per-engine documentation page, as a repo-relative path (#1866). Every one of
+# these docs already exists and several are CI-guarded against the code they
+# describe (e.g. tests/test_cosyvoice_install_docs.py), but nothing in the app
+# linked to them, so the point of failure — an unavailable engine row — was a
+# dead end. Paths rather than URLs so tests/test_engine_docs.py can assert the
+# file is really there; the URL is built once, at read time, from core.links.
+#
+# Keyed on the engine id, so it stays correct when the doc filename does not
+# match the id (indextts2 → indextts.md).
+_ENGINE_DOCS: dict[str, str] = {
+    "omnivoice":            "docs/engines/omnivoice.md",
+    "omnivoice-subprocess": "docs/engines/omnivoice-subprocess.md",
+    "omnivoice-gguf":       "docs/engines/omnivoice-gguf.md",
+    "cosyvoice":            "docs/engines/cosyvoice.md",
+    "kittentts":            "docs/engines/kittentts.md",
+    "mlx-audio":            "docs/engines/mlx-audio.md",
+    "voxcpm2":              "docs/engines/voxcpm2.md",
+    "moss-tts-nano":        "docs/engines/moss-tts-nano.md",
+    "moss-tts-v15":         "docs/engines/moss-tts-v15.md",
+    "dots-tts":             "docs/engines/dots-tts.md",
+    "confucius4-tts":       "docs/engines/confucius4-tts.md",
+    "indextts2":            "docs/engines/indextts.md",
+    "gpt-sovits":           "docs/engines/gpt-sovits.md",
+    "sherpa-onnx":          "docs/engines/sherpa-onnx.md",
+    "supertonic3":          "docs/engines/supertonic3.md",
+    "pockettts":            "docs/engines/pockettts.md",
+    "audiocpp":             "docs/engines/audio-cpp.md",
+}
+
+
+def _engine_docs_url(bid: str) -> str | None:
+    """Public URL of this engine's doc page, or None when it has none.
+
+    VoiceStudio-owned constant either way: the path comes from the registry
+    above and the base from :mod:`core.links`, so no part of it is derived
+    from an engine probe. That is what lets it cross the public boundary
+    intact (see api.public_engine_metadata).
+    """
+    path = _ENGINE_DOCS.get(bid)
+    if not path:
+        return None
+    from core import links
+    return f"{links.PROJECT_REPO_BLOB_MAIN}/{path}"
+
+
 # Short, readable labels for mlx-audio's curated models (#981) — surfaced in
 # the Model Catalogue → Engines model picker so users see more than a bare key.
 # Single-sourced here rather than on MLXAudioBackend.CURATED_MODELS itself so
@@ -2347,14 +2463,24 @@ def _sidecar_installable_ids() -> frozenset[str]:
     button into their matrix rows.
     """
     try:
-        from services.sidecar_install import SPECS
-        return frozenset(SPECS)
+        # Host-aware: an engine whose installer cannot work on THIS machine
+        # (dots.tts on Windows, a CUDA-only install on a CPU host) must not get
+        # an Install button that can only fail.
+        from services.sidecar_install import installable_engine_ids
+        return installable_engine_ids()
     except Exception:  # pragma: no cover — defensive only
         return frozenset()
 
 
-def list_backends() -> list[dict]:
-    """Enumerate every registered backend with its availability state.
+def list_backends(*, include_hidden: bool = False) -> list[dict]:
+    """Enumerate the engine catalogue with each backend's availability state.
+
+    On MPS, the canonical ``omnivoice`` id already resolves to the killable
+    OmniVoice sidecar. The explicit ``omnivoice-subprocess`` compatibility id
+    is therefore omitted from the normal catalogue so the picker does not
+    advertise two choices with the same runtime behavior. Internal callers
+    that must validate or preserve a stored compatibility id can pass
+    ``include_hidden=True``.
 
     Per-entry shape (ENGINE-05 + ENGINE-06):
 
@@ -2368,10 +2494,11 @@ def list_backends() -> list[dict]:
                                                     #   e.g. VoxCPM2's >=2.0.3 upgrade hint)
           "install_hint":   Optional[str],
           "setup_snippet":  Optional[str],          # exact `export VAR=...` for path-gated opt-in engines
+          "docs_url":       Optional[str],          # this engine's doc page (registry-authored constant)
           "one_click_install": bool,                # services.sidecar_install can provision it in-app
           "last_error":     Optional[str],          # cached most-recent failure
           "isolation_mode": "in-process" | "subprocess",
-          "gpu_compat":     list[str],              # subset of {cuda, rocm, mps, xpu, cpu}
+          "gpu_compat":     list[str],              # subset of {cuda, rocm, mps, vulkan, xpu, npu, cpu}
           "supports_cloning": Optional[bool],       # True/False from the class attr; None when
                                                     #   model-dependent (property, e.g. mlx-audio)
           "effective_device": str,                  # device this engine uses on THIS host
@@ -2401,12 +2528,17 @@ def list_backends() -> list[dict]:
     from core.device_caps import detect_host_caps
     from services.engine_disk_usage import disk_summary_for
     from services.engine_evidence import snapshot as execution_snapshot
-    from services.engine_routing import routing_fields
     caps = detect_host_caps()
     installable = _sidecar_installable_ids()
 
     out: list[dict] = []
     for bid, cls in _REGISTRY.items():
+        if (
+            not include_hidden
+            and caps.family == "mps"
+            and bid == "omnivoice-subprocess"
+        ):
+            continue
         cls = _effective_backend_class(bid, cls, caps.family)
         try:
             ok, msg = cls.is_available()
@@ -2427,14 +2559,40 @@ def list_backends() -> list[dict]:
             isolation = "subprocess"
         else:
             isolation = "in-process"
-        gpu_compat = getattr(cls, "gpu_compat", ("cpu",))
+        from services.engine_routing import resolve_routing, runtime_compute_profile
+        try:
+            profile = runtime_compute_profile(cls, caps)
+        except Exception:
+            # Runtime-aware native probes remain optional metadata. A broken
+            # provider probe must not take down the engine picker, especially
+            # when availability already explains a missing binary or model.
+            compat = tuple(getattr(cls, "gpu_compat", ("cpu",)))
+            floor = float(getattr(cls, "min_vram_gb", 0.0) or 0.0)
+            profile = {
+                "gpu_compat": compat,
+                "min_vram_gb": floor,
+                **resolve_routing(compat, caps, floor),
+                "runtime_backend": None,
+                "runtime_device_index": None,
+                "runtime_device_name": None,
+                "runtime_hardware_family": None,
+                "runtime_vram_gb": None,
+                "runtime_device_verified": None,
+            }
+        gpu_compat = profile["gpu_compat"]
         # Cloning capability: same descriptor guard as
         # cloning_capable_engine_ids() — a class-level getattr on a *property*
         # (mlx-audio: capability depends on the picked model) returns the
         # descriptor, not a bool, so report None (= model-dependent) there
         # instead of an always-truthy false positive.
         _clone = getattr(cls, "supports_cloning", True)
-        routing = routing_fields(gpu_compat, caps, getattr(cls, "min_vram_gb", 0.0))
+        from core.scrub import scrub_text
+        routing = {
+            "effective_device": profile["effective_device"],
+            "routing_status": profile["routing_status"],
+            "routing_reason": scrub_text(profile["routing_reason"])
+            if profile["routing_reason"] else None,
+        }
         loaded_instance = None
         if _active_instance_id == bid:
             loaded_instance = _active_instance
@@ -2455,6 +2613,10 @@ def list_backends() -> list[dict]:
             "install_hint": _INSTALL_HINTS.get(bid),
             # Exact `export VAR=...` line for path-gated opt-in engines, or None.
             "setup_snippet": _SETUP_SNIPPETS.get(bid),
+            # This engine's doc page (#1866). Registry-authored constant, so it
+            # survives api.public_engine_metadata and gives an unavailable row
+            # somewhere to send the user.
+            "docs_url": _engine_docs_url(bid),
             # True when services.sidecar_install can provision this engine
             # in-app (Settings renders an Install button instead of leading
             # with the manual setup snippet).
@@ -2464,7 +2626,7 @@ def list_backends() -> list[dict]:
             "isolation_mode": isolation,
             "gpu_compat": list(gpu_compat),
             # effective_device / routing_status / routing_reason (scrubbed):
-            "min_vram_gb": getattr(cls, "min_vram_gb", 0.0) or None,
+            "min_vram_gb": profile["min_vram_gb"] or None,
             # effective_device / routing_status / routing_reason (scrubbed);
             # the reason now also carries the under-provisioned-GPU caveat.
             **routing,
@@ -2472,7 +2634,7 @@ def list_backends() -> list[dict]:
                 engine_id=bid,
                 engine_cls=cls,
                 instance=loaded_instance,
-                routing=routing,
+                routing={**profile, **routing},
                 caps=caps,
             ),
         })
@@ -2554,7 +2716,9 @@ def active_routing() -> dict | None:
     """
     try:
         active = active_backend_id()
-        for b in list_backends():
+        # The MPS picker intentionally hides the redundant compatibility id,
+        # but routing must still describe a saved or environment-pinned id.
+        for b in list_backends(include_hidden=True):
             if b.get("id") == active:
                 return {
                     "engine": active,
@@ -2894,10 +3058,9 @@ async def resolve_generation_backend(
         raise ValueError(f"TTS engine '{engine_id}' is not available: {_mask_hf_tokens(msg)}")
 
     from core.device_caps import detect_host_caps
-    from services.engine_routing import resolve_routing
-    routing = resolve_routing(
-        getattr(backend_cls, "gpu_compat", ("cpu",)), detect_host_caps(),
-        getattr(backend_cls, "min_vram_gb", 0.0),
+    from services.engine_routing import runtime_compute_profile_async
+    routing = await runtime_compute_profile_async(
+        backend_cls, detect_host_caps()
     )
     if routing["routing_status"] == "unavailable":
         raise ValueError(routing["routing_reason"])
@@ -2935,4 +3098,6 @@ def __getattr__(name: str):  # pragma: no cover - exercised via tests
         return _REGISTRY[name if name in _REGISTRY else None]
     if name == "IndexTTS2Backend":
         return _REGISTRY["indextts2"]
+    if name == "AudioCPPBackend":
+        return _REGISTRY["audiocpp"]
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

backend/tests/test_archetype_preview_quality.py

@@ -7,13 +7,15 @@ old silence-only guard missed it (the buzz is loud, not silent) so the garbage
 was cached and served.
 
 These tests cover the fix *without the 5 GB model / a GPU*: they drive the pure
-``_spectral_flatness`` / ``_is_unusable_audio`` helpers with synthetic signals,
-and assert the render constants didn't regress. The real end-to-end render is
-verified manually (spectral flatness back in the speech range + Whisper ASR).
+``_spectral_flatness`` / ``_is_unusable_audio`` helpers with synthetic tones and
+tracked speech demo renders, and assert the render constants did not regress.
 """
 from __future__ import annotations
 
 import math
+from pathlib import Path
+
+import soundfile as sf
 
 import pytest
 
@@ -39,6 +41,13 @@ def _white_noise() -> "torch.Tensor":
     return 0.5 * (torch.rand(N, generator=g) * 2 - 1)
 
 
+def _two_tone_buzz() -> "torch.Tensor":
+    """Two inharmonic partials — the other shape a collapsed render takes."""
+    t = torch.arange(N, dtype=torch.float32) / SR
+    s = torch.sin(2 * math.pi * 180.0 * t) + 0.6 * torch.sin(2 * math.pi * 361.0 * t)
+    return 0.8 * s / s.abs().max()
+
+
 def _speech_like() -> "torch.Tensor":
     """Broadband + harmonic + amplitude-modulated — a coarse stand-in for voiced
     speech: several harmonics (formant-ish), additive noise (consonants), and a
@@ -85,6 +94,71 @@ def test_speech_like_is_usable():
     assert arch._is_unusable_audio(_speech_like()) is False
 
 
+# ── Threshold stays between the two things it has to separate ───────────────
+# Synthetic broadband speech has much higher flatness than real voiced audio.
+# Measure both sides of the threshold against actual inputs, including the
+# existing demo renders that the old thresholds rejected.
+
+
+def test_tonal_ceiling_is_measured_not_assumed():
+    """Derive the tonal side of the margin instead of trusting a literal.
+
+    A bare constant would keep passing if `_spectral_flatness` stopped scoring
+    tones near zero, so measure the degenerate signals here and require the
+    threshold to clear the worst of them tenfold.
+    """
+    tones = [
+        arch._spectral_flatness(_pure_tone(80.0)),
+        arch._spectral_flatness(_pure_tone(220.0)),
+        arch._spectral_flatness(_two_tone_buzz()),
+    ]
+    assert all(t is not None for t in tones)
+    assert max(tones) * 10 < arch._DEGENERATE_FLATNESS
+
+
+_SAMPLES = Path(__file__).resolve().parents[1] / "assets" / "samples"
+_SPEECH_FIXTURES = [
+    "demo_voice.wav",
+    "demo_clone_output.wav",
+    *[f"voice_design/demo_voice_design_{name}.wav" for name in (
+        "audiobook_uk_narrator", "aussie_podcaster", "bedtime_storyteller",
+        "gravelly_villain", "indian_support_agent", "mandarin_sichuan", "us_news_anchor",
+    )],
+    *[f"dictation/{name}.wav" for name in (
+        "en_conversational", "en_technical", "fr_reservation",
+    )],
+    *[f"demo/dubbing/{name}.src.wav" for name in (
+        "source", "dubbed_es", "dubbed_fr", "dubbed_ja", "dubbed_zh",
+    )],
+]
+
+
+@pytest.mark.parametrize("fixture", _SPEECH_FIXTURES)
+def test_real_shipped_speech_clears_quality_floor(fixture):
+    # These are existing tracked demo renders, not synthesized stand-ins or
+    # asserted measurements. Loading PCM needs neither a model nor a network.
+    audio, _sample_rate = sf.read(_SAMPLES / fixture, dtype="float32", always_2d=True)
+    speech = torch.from_numpy(audio.T)
+    flatness = arch._spectral_flatness(speech)
+    assert flatness is not None
+    assert flatness > arch._DEGENERATE_FLATNESS * 10
+    assert arch._is_unusable_audio(speech) is False
+
+
+def test_flatness_is_not_clip_length_dependent():
+    """Repeating a signal must not change what it measures.
+
+    The whole-clip FFT this replaced failed exactly here: its frequency
+    resolution grew with duration, so the same audio measured 0.0229 at 3 s and
+    ~0 at 12 s (100% drift). Framed, the drift is under 0.1%.
+    """
+    short = _speech_like()
+    long = torch.cat([short] * 4)
+    a, b = arch._spectral_flatness(short), arch._spectral_flatness(long)
+    assert a is not None and b is not None
+    assert abs(a - b) / a < 0.02
+
+
 # ── Constants didn't regress ────────────────────────────────────────────────
 def test_preview_render_constants():
     # 16 steps under-converged on the social script; the fix bumped it.

backend/tests/test_contained_subprocess.py

@@ -120,6 +120,7 @@ def test_drain_fd_is_explicitly_inherited_by_wrapper_but_not_operation(monkeypat
             proc.wait(timeout=5)
 
 
+@pytest.mark.skipif(os.name != "posix", reason="Unix drain pipe contract")
 def test_invalid_or_missing_desktop_drain_fd_fails_safe(monkeypatch):
     monkeypatch.setenv("OMNIVOICE_DESKTOP_CONTAINED", "1")
     monkeypatch.setenv("OMNIVOICE_DESKTOP_DRAIN_FD", "not-an-fd")

backend/tests/test_contained_subprocess_waitid_fallback.py

@@ -15,6 +15,15 @@ import pytest
 
 from core import contained_subprocess as owned
 
+# This module simulates macOS by deleting os.waitid, then drives the fallback
+# with os.waitpid/os.WNOHANG and start_new_session — POSIX-only APIs that
+# Windows does not have at all (os.WNOHANG raises AttributeError before the
+# first assertion). CI runs this suite on Linux, so nothing is lost by
+# skipping; what is gained is a Windows contributor whose checkout runs green.
+pytestmark = pytest.mark.skipif(
+    os.name != "posix", reason="simulates a POSIX platform without os.waitid"
+)
+
 
 def _make_owned(argv):
     cr, cw = os.pipe()

backend/tests/test_omnivoice_subprocess.py

@@ -160,6 +160,96 @@ def test_engine_catalogue_reports_effective_mps_isolation(monkeypatch):
     assert row["isolation_mode"] == "subprocess"
 
 
+def test_mps_catalogue_hides_redundant_explicit_omnivoice_sidecar(monkeypatch):
+    """The picker advertises the canonical id, while legacy callers retain both."""
+    from core.device_caps import HostCaps
+    from services import tts_backend
+
+    monkeypatch.setattr(
+        tts_backend,
+        "_REGISTRY",
+        {
+            "omnivoice": OmniVoiceBackend,
+            "omnivoice-subprocess": OmniVoiceSubprocessBackend,
+        },
+    )
+    monkeypatch.setattr(
+        "core.device_caps.detect_host_caps",
+        lambda: HostCaps(family="mps", available_families=("mps", "cpu")),
+    )
+    monkeypatch.setattr(
+        OmniVoiceSubprocessBackend,
+        "is_available",
+        classmethod(lambda cls: (True, "ready")),
+    )
+
+    picker_ids = {item["id"] for item in list_backends()}
+    assert picker_ids == {"omnivoice"}
+    assert get_backend_class("omnivoice") is OmniVoiceMPSSubprocessBackend
+
+    all_ids = {item["id"] for item in list_backends(include_hidden=True)}
+    assert all_ids == {"omnivoice", "omnivoice-subprocess"}
+    assert get_backend_class("omnivoice-subprocess") is OmniVoiceSubprocessBackend
+
+
+def test_mps_active_routing_preserves_hidden_compatibility_id(monkeypatch):
+    from core.device_caps import HostCaps
+    from services import tts_backend
+
+    monkeypatch.setattr(
+        tts_backend,
+        "_REGISTRY",
+        {"omnivoice-subprocess": OmniVoiceSubprocessBackend},
+    )
+    monkeypatch.setattr(tts_backend, "active_backend_id", lambda: "omnivoice-subprocess")
+    monkeypatch.setattr(
+        "core.device_caps.detect_host_caps",
+        lambda: HostCaps(family="mps", available_families=("mps", "cpu")),
+    )
+    monkeypatch.setattr(
+        OmniVoiceSubprocessBackend,
+        "is_available",
+        classmethod(lambda cls: (True, "ready")),
+    )
+
+    assert tts_backend.active_routing() == {
+        "engine": "omnivoice-subprocess",
+        "available": True,
+        "effective_device": "mps",
+        "routing_status": "accelerated",
+        "routing_reason": None,
+    }
+
+
+@pytest.mark.parametrize("family", ("cuda", "cpu"))
+def test_non_mps_catalogue_keeps_explicit_omnivoice_sidecar(monkeypatch, family):
+    from core.device_caps import HostCaps
+    from services import tts_backend
+
+    monkeypatch.setattr(
+        tts_backend,
+        "_REGISTRY",
+        {
+            "omnivoice": OmniVoiceBackend,
+            "omnivoice-subprocess": OmniVoiceSubprocessBackend,
+        },
+    )
+    monkeypatch.setattr(
+        "core.device_caps.detect_host_caps",
+        lambda: HostCaps(family=family, available_families=(family, "cpu")),
+    )
+    monkeypatch.setattr(
+        OmniVoiceSubprocessBackend,
+        "is_available",
+        classmethod(lambda cls: (True, "ready")),
+    )
+
+    assert {item["id"] for item in list_backends()} == {
+        "omnivoice",
+        "omnivoice-subprocess",
+    }
+
+
 def test_mps_startup_does_not_preload_native_model(monkeypatch):
     from core.device_caps import HostCaps
     from services import model_manager
@@ -382,7 +472,17 @@ def test_mps_proxy_survives_fatal_child_exit_and_recovers(stub_sidecar, monkeypa
     try:
         with pytest.raises(RuntimeError, match="backend is still running"):
             b.generate("CRASH")
-        assert b._proc is not None and b._proc.poll() is not None
+        assert b._proc is not None
+        # The child called os._exit; the parent raised the moment its pipe hit
+        # EOF, which is BEFORE the OS has reaped the process. Asserting poll()
+        # on the next line is a race the test happened to win on Linux and lost
+        # every time on Windows. Wait for the death instead of assuming it has
+        # already been observed — the claim is that the child is gone, not that
+        # it is gone within one instruction.
+        deadline = time.monotonic() + 5
+        while b._proc.poll() is None and time.monotonic() < deadline:
+            time.sleep(0.02)
+        assert b._proc.poll() is not None, "the crashed sidecar never died"
         assert b.generate("ok").shape[1] == 24000
     finally:
         b.shutdown()
@@ -523,3 +623,163 @@ def test_generation_proxy_forwards_native_controls_and_seed():
         "class_temperature": 0.8,
         "seed": 321,
     })]
+
+
+def test_timeout_reaps_captured_process_before_recv_returns(monkeypatch):
+    import threading
+
+    class Process:
+        def __init__(self):
+            self.killed = threading.Event()
+            self.reaped = False
+            self.wait_entered = threading.Event()
+            self.release_wait = threading.Event()
+
+        def kill(self):
+            self.killed.set()  # EOF may arrive before the process is reaped.
+
+        def wait(self, timeout):
+            assert timeout is not None
+            self.wait_entered.set()
+            assert self.release_wait.wait(2)
+            self.reaped = True
+            return -9
+
+    proc = Process()
+    backend = OmniVoiceSubprocessBackend()
+    backend._proc = proc
+
+    def recv():
+        assert proc.killed.wait(2)
+        return None
+
+    monkeypatch.setattr(backend, '_recv', recv)
+    returned = threading.Event()
+    results = []
+
+    def receive():
+        results.append(backend._recv_with_timeout(0.01))
+        returned.set()
+
+    reader = threading.Thread(target=receive)
+    reader.start()
+    try:
+        assert proc.wait_entered.wait(2)
+        assert not returned.wait(0.05), "EOF must not release the caller before process cleanup"
+    finally:
+        proc.release_wait.set()
+        reader.join(2)
+        backend._proc = None
+    assert not reader.is_alive()
+    assert returned.is_set()
+    assert results == [None]
+    assert proc.reaped
+
+
+def test_timeout_never_kills_a_replacement_process(monkeypatch):
+    from unittest.mock import Mock
+    import services.subprocess_backend as module
+
+    class ManualTimer:
+        def __init__(self, _timeout, callback, args=()):
+            self.callback = lambda: callback(*args)
+            self.daemon = False
+
+        def start(self):
+            pass
+
+        def cancel(self):
+            pass
+
+        def join(self):
+            pass
+
+    timers = []
+    def timer(*args, **kwargs):
+        result = ManualTimer(*args, **kwargs)
+        timers.append(result)
+        return result
+
+    monkeypatch.setattr(module.threading, 'Timer', timer)
+    backend = OmniVoiceSubprocessBackend()
+    original, replacement = Mock(), Mock()
+    backend._proc = original
+
+    def recv():
+        backend._proc = replacement
+        timers[0].callback()
+        return None
+
+    monkeypatch.setattr(backend, '_recv', recv)
+    try:
+        backend._recv_with_timeout(1)
+        original.kill.assert_called_once()
+        replacement.kill.assert_not_called()
+    finally:
+        backend._proc = None
+
+
+@pytest.mark.parametrize("failure", ["wait", "kill"])
+def test_timeout_quarantine_blocks_reuse_and_retains_cleanup_handle(failure):
+    class StuckProcess:
+        stdin = None
+        def __init__(self):
+            self.exited = False
+            self.kill_calls = 0
+        def poll(self):
+            return 0 if self.exited else None
+        def kill(self):
+            self.kill_calls += 1
+            if failure == "kill" and not self.exited:
+                raise PermissionError("kill failed")
+        def terminate(self):
+            pass
+        def wait(self, timeout):
+            if not self.exited:
+                raise subprocess.TimeoutExpired("stuck-sidecar", timeout)
+            return 0
+
+    backend = OmniVoiceSubprocessBackend()
+    proc = StuckProcess()
+    backend._proc = proc
+    try:
+        backend._timeout_kill(proc)
+        with pytest.raises(RuntimeError, match="still stopping"):
+            backend._spawn()
+        backend.shutdown()
+        # Even after shutdown clears the current slot, ownership survives;
+        # retry must not silently start a second process next to this one.
+        before = proc.kill_calls
+        with pytest.raises(RuntimeError, match="still stopping"):
+            backend._spawn()
+        assert proc.kill_calls > before
+    finally:
+        proc.exited = True
+        backend.shutdown()
+
+
+def test_timeout_quarantine_does_not_clear_or_kill_replacement():
+    from unittest.mock import Mock
+    backend = OmniVoiceSubprocessBackend()
+    original = Mock()
+    original.wait.side_effect = subprocess.TimeoutExpired("old-sidecar", 2)
+    replacement = Mock()
+    replacement.poll.return_value = None
+    backend._proc = replacement
+    try:
+        backend._timeout_kill(original)
+        with pytest.raises(RuntimeError, match="still stopping"):
+            backend._spawn()
+        assert backend._proc is replacement
+        replacement.kill.assert_not_called()
+        # Once the captured owner is reaped, reuse of the healthy replacement
+        # is allowed without starting or terminating another process.
+        original.wait.side_effect = None
+        original.wait.return_value = 0
+        backend._spawn()
+        assert backend._proc is replacement
+        replacement.kill.assert_not_called()
+    finally:
+        original.wait.side_effect = None
+        backend._proc = None
+        backend.shutdown()

backend/tests/test_run_sentinel.py

@@ -113,6 +113,93 @@ def test_unclean_shutdown_yields_crash_record(sentinel_env, monkeypatch):
     assert acked is False, "a fresh crash record must be unacknowledged"
 
 
+def test_lifespan_clears_sentinel_even_if_later_shutdown_raises(monkeypatch, tmp_path):
+    """THE #1895 regression: before this fix, ``clear_sentinel()`` was the
+    LAST statement of ``main.py``'s lifespan shutdown, behind ~50s of bounded
+    waits plus model unload / ``free_vram()`` / ``gc.collect()`` / httpx
+    close. The desktop shell's quit path grants only a 2s grace before
+    SIGKILL (``frontend/src-tauri/src/bootstrap.rs``
+    ``terminate_process_tree``), and Windows grants no graceful phase at all
+    (``tools.rs``) — nowhere near enough, so a deliberate, clean quit
+    routinely got killed before reaching that last line, leaving the
+    sentinel behind for the NEXT startup to misreport as "did not shut down
+    cleanly... likely crashed".
+
+    Simulates that class of interruption without an actual SIGKILL: a later
+    shutdown step (``model_loads_begin_shutdown()``, called unguarded well
+    after the sentinel clear) raises, so nothing past it in the shutdown
+    body ever runs — for this purpose, the same effect as being killed
+    mid-teardown.
+
+    Fail-before/pass-after: with ``clear_sentinel()`` moved to the TOP of
+    the shutdown block (immediately after ``yield``), the sentinel is
+    already gone by the time this raise happens, so the next startup must
+    not fabricate a crash record.
+    """
+    import asyncio
+    from fastapi import FastAPI
+
+    # Fresh `main`/`core`/`api`/`services` import, mirroring
+    # tests/test_model_load_shutdown.py's `_reimported_backend_modules`: a
+    # sibling suite may have purged these names from sys.modules, leaving a
+    # collection-time alias stale. Purging and re-importing here makes this
+    # test self-consistent in isolation, not dependent on suite order.
+    purge_names = ("main", "core", "api", "services")
+    purge_prefixes = ("core.", "api.", "services.")
+    saved = {
+        name: mod for name, mod in sys.modules.items()
+        if name in purge_names or name.startswith(purge_prefixes)
+    }
+
+    def _purge():
+        for name in [
+            n for n in sys.modules
+            if n in purge_names or n.startswith(purge_prefixes)
+        ]:
+            sys.modules.pop(name, None)
+
+    _purge()
+    try:
+        import main as main_mod
+        from core import run_sentinel as fresh_run_sentinel
+
+        monkeypatch.setattr(
+            fresh_run_sentinel, "SENTINEL_PATH", str(tmp_path / "run_sentinel.json")
+        )
+        monkeypatch.setattr(
+            fresh_run_sentinel, "CRASH_RECORD_PATH", str(tmp_path / "last_run_crash.json")
+        )
+        monkeypatch.setattr(
+            fresh_run_sentinel, "LOG_PATH", str(tmp_path / "omnivoice.log")
+        )
+        fresh_run_sentinel._reset_for_tests()
+
+        def _boom():
+            raise RuntimeError("simulated kill: interrupted after the early clear")
+
+        monkeypatch.setattr(main_mod, "model_loads_begin_shutdown", _boom)
+
+        async def scenario():
+            app = FastAPI()
+            async with main_mod.lifespan(app):
+                pass
+
+        with pytest.raises(RuntimeError, match="simulated kill"):
+            asyncio.run(scenario())
+
+        assert not os.path.exists(fresh_run_sentinel.SENTINEL_PATH), (
+            "sentinel must already be cleared even though a later shutdown "
+            "step raised before ever reaching the old clear-sentinel line"
+        )
+        assert fresh_run_sentinel.detect_unclean_shutdown() is None, (
+            "a deliberate quit interrupted after the early clear must never "
+            "be reported as a crash on the next startup"
+        )
+    finally:
+        _purge()
+        sys.modules.update(saved)
+
+
 def test_live_pid_means_second_instance_not_a_crash(sentinel_env):
     """A sentinel owned by a LIVE process is a concurrent second instance
     sharing DATA_DIR — never a crash, and we must not take over or delete

backend/worker/capabilities.py

@@ -23,6 +23,8 @@ from __future__ import annotations
 import logging
 from typing import Optional
 
+from worker.capacity import derive_concurrency
+
 logger = logging.getLogger("omnivoice.worker")
 
 # gpu_compat families that mean "this would run on the CPU here", which is
@@ -74,6 +76,12 @@ def discover(*, include_unavailable: bool = False) -> list[dict]:
         gpu_compat = set(entry.get("gpu_compat") or [])
         repo_ids = repo_ids_for(entry)
         downloaded = _downloaded(repo_ids)
+        runtime_vram_gb = (entry.get("execution_evidence") or {}).get(
+            "runtime_vram_gb"
+        )
+        engine_free_bytes = free_bytes if runtime_vram_gb is None else int(
+            float(runtime_vram_gb or 0.0) * 1024**3
+        )
         discovered.append(
             {
                 "engine": engine_id,
@@ -97,7 +105,17 @@ def discover(*, include_unavailable: bool = False) -> list[dict]:
                 "min_memory_bytes": int(float(entry.get("min_vram_gb") or 0) * 1024**3),
                 "precision": "",
                 "backend": entry.get("effective_device") or family,
-                "free_memory_bytes": free_bytes,
+                "free_memory_bytes": engine_free_bytes,
+                # A native provider that works independently of torch may not
+                # expose memory telemetry. Unknown capacity still gets one
+                # serial slot; zero must not turn a working Vulkan engine into
+                # an unschedulable capability.
+                "derived_concurrency": 1
+                if (
+                    runtime_vram_gb is not None
+                    and float(runtime_vram_gb or 0.0) <= 0
+                    and routing == "accelerated"
+                ) else 0,
                 # Capability is not acceleration: an engine present but routed
                 # to the CPU here should not be preferred for GPU work.
                 "cpu_fallback": routing in ("cpu_fallback", "cpu_only")
@@ -291,9 +309,18 @@ def max_concurrent_tasks(capabilities: Optional[list[dict]] = None) -> int:
     caps = capabilities if capabilities is not None else discover()
     if not caps:
         return 1
-    derived = [int(c.get("derived_concurrency") or 0) for c in caps]
-    positive = [d for d in derived if d > 0]
-    return min(positive) if positive else 1
+    derived: list[int] = []
+    for cap in caps:
+        concurrency = int(cap.get("derived_concurrency") or 0)
+        if concurrency <= 0:
+            concurrency = derive_concurrency(
+                backend=str(cap.get("backend") or ""),
+                free_memory_bytes=int(cap.get("free_memory_bytes") or 0),
+                min_model_bytes=int(cap.get("min_memory_bytes") or 0),
+                compiled=bool(cap.get("compiled")),
+            )
+        derived.append(concurrency)
+    return min(derived)
 
 
 __all__ = [

backend/worker/capacity.py

@@ -97,18 +97,16 @@ def derive_concurrency(
 ) -> int:
     """How many jobs of this model may run at once on this worker.
 
-    Returns 0 when the model cannot run here at all — a capability mismatch,
-    which the scheduler must treat as "send it elsewhere", never as a worker
-    fault.
+    Under-provisioned accelerators remain usable with one serial slot and the
+    longer CPU-class deadline. Zero memory means telemetry is unknown, not
+    that the engine cannot run.
     """
     family = (backend or "").strip().lower()
-    if min_model_bytes and free_memory_bytes < min_model_bytes:
-        return 0
     if compiled:
         # Thread-affinity pinning (#315). One job, always.
         return 1
     if family in _ALWAYS_SERIAL:
-        return 1 if (not min_model_bytes or free_memory_bytes >= min_model_bytes) else 0
+        return 1
     budget = max(min_model_bytes, _VRAM_PER_JOB_BYTES)
     if budget <= 0:
         return 1

backend/worker/deadlines.py

@@ -127,16 +127,31 @@ class Deadlines:
 
 
 def _base_execution_seconds(
-    text: Optional[str], *, execution_device: Optional[str] = None
+    text: Optional[str], *, execution_device: Optional[str] = None,
+    under_provisioned: bool = False,
 ) -> float:
     """Delegate to model_manager's budget; fall back to its formula.
 
     The lazy import keeps this module usable in a process that has no torch —
     the control plane schedules work it never executes.
+
+    ``under_provisioned`` is the worker's own verdict that its card sits below
+    the engine's declared VRAM floor (``ConnectedWorker.under_provisioned``).
+    It floors the budget at what the same job would get on a CPU, because that
+    is what a card paging to system RAM performs like (#1804). Derived by asking
+    for the CPU budget rather than by probing VRAM here: this process is the
+    control plane, and its hardware is not the worker's.
     """
     target_device = str(execution_device or "cpu").lower()
-    if target_device not in {"cpu", "cuda", "mps", "mlx", "directml", "rocm", "xpu"}:
+    if target_device not in {
+        "cpu", "cuda", "mps", "mlx", "directml", "rocm", "vulkan", "xpu",
+    }:
         target_device = "cpu"
+    if under_provisioned and target_device != "cpu":
+        return max(
+            _base_execution_seconds(text, execution_device=target_device),
+            _base_execution_seconds(text, execution_device="cpu"),
+        )
     try:
         from services import model_manager  # noqa: PLC0415 — intentionally lazy
 
@@ -171,6 +186,7 @@ def for_task(
     model_downloaded: bool = True,
     input_seconds: float = 0.0,
     execution_device: Optional[str] = None,
+    under_provisioned: bool = False,
 ) -> Deadlines:
     """Compute the deadlines for one attempt.
 
@@ -183,7 +199,8 @@ def for_task(
     multiplier, grace = _PROFILE[op]
 
     execution = _base_execution_seconds(
-        text, execution_device=execution_device
+        text, execution_device=execution_device,
+        under_provisioned=under_provisioned,
     ) * multiplier
     # Media-length operations scale on duration, not characters.
     if input_seconds > 0:

backend/worker/lifecycle.py

@@ -32,6 +32,7 @@ import uuid
 from dataclasses import dataclass, field
 from typing import Iterable, Optional
 
+from worker.deadlines import Deadlines
 from worker.clock import resolve
 from worker.errors import ErrorClass, WorkerError
 
@@ -224,6 +225,9 @@ class Attempt:
     stage: str = ""
     error: Optional[WorkerError] = None
 
+    # Snapshot the lease policy granted at dispatch, including after restart.
+    deadlines: Optional[Deadlines] = None
+
     def matches(self, *, session_epoch: Optional[int] = None) -> bool:
         """Fence check: reject messages from a superseded session."""
         if session_epoch is None:

backend/worker/pool.py

@@ -34,7 +34,7 @@ _HEARTBEAT_MISS_SECONDS = 90.0
 # enough that one slow answer cannot move it.
 _LATENCY_WINDOW = 5
 _KNOWN_EXECUTION_DEVICES = frozenset(
-    {"cpu", "cuda", "mps", "mlx", "directml", "rocm", "xpu"}
+    {"cpu", "cuda", "mps", "mlx", "directml", "rocm", "vulkan", "xpu"}
 )
 
 
@@ -93,12 +93,11 @@ class ConnectedWorker:
             return "busy"
         return "ready"
 
-    def supports(self, engine: str, model_id: str, operation: str) -> bool:
-        """Can this worker run this work at all?
+    def _capability_for(self, engine: str, model_id: str, operation: str):
+        """The advertised capability this task would actually be run by.
 
-        ``supported`` alone is not enough — an engine whose weights are not on
-        disk cannot start without a download, and one that is not installed
-        cannot start at all. Both are capability mismatches, not failures.
+        One selection rule, so the answers below cannot describe different
+        capabilities of the same worker.
         """
         for cap in self.record.capabilities:
             if cap.get("engine") != engine:
@@ -107,23 +106,57 @@ class ConnectedWorker:
                 continue
             if operation and operation not in (cap.get("operations") or [operation]):
                 continue
-            return bool(cap.get("supported")) and bool(cap.get("installed", True))
-        return False
+            return cap
+        return None
+
+    def supports(self, engine: str, model_id: str, operation: str) -> bool:
+        """Can this worker run this work at all?
+
+        ``supported`` alone is not enough — an engine whose weights are not on
+        disk cannot start without a download, and one that is not installed
+        cannot start at all. Both are capability mismatches, not failures.
+        """
+        cap = self._capability_for(engine, model_id, operation)
+        if cap is None:
+            return False
+        return bool(cap.get("supported")) and bool(cap.get("installed", True))
 
     def execution_device(self, engine: str, model_id: str, operation: str) -> str:
         """Device used by the exact capability selected for this task."""
-        for cap in self.record.capabilities:
-            if cap.get("engine") != engine:
-                continue
-            if model_id and cap.get("model_id") not in (model_id, "", None):
-                continue
-            if operation and operation not in (cap.get("operations") or [operation]):
-                continue
-            if cap.get("cpu_fallback"):
-                return "cpu"
-            backend = str(cap.get("backend") or "").lower()
-            return backend if backend in _KNOWN_EXECUTION_DEVICES else "cpu"
-        return "cpu"
+        cap = self._capability_for(engine, model_id, operation)
+        if cap is None:
+            return "cpu"
+        if cap.get("cpu_fallback"):
+            return "cpu"
+        backend = str(cap.get("backend") or "").lower()
+        return backend if backend in _KNOWN_EXECUTION_DEVICES else "cpu"
+
+    def under_provisioned(self, engine: str, model_id: str, operation: str) -> bool:
+        """Is this worker's GPU below the engine's declared VRAM floor?
+
+        The remote half of #1804. A card under the floor pages to system RAM and
+        renders slower than a CPU, so it must not be given the shorter
+        accelerated deadline. Decided from the two figures the WORKER itself
+        advertises (``free_memory_bytes`` / ``min_memory_bytes``, both set in
+        ``worker/capabilities.py``): the control plane's own VRAM says nothing
+        about the machine that will run the job, so
+        ``engine_routing.under_provisioned_vram`` — which probes THIS host —
+        cannot answer for a remote worker.
+
+        Same rules as that predicate otherwise: dedicated-VRAM devices only
+        (unified memory is not a comparable pool), and a zero on either side
+        means "unknown", never "too small".
+        """
+        cap = self._capability_for(engine, model_id, operation)
+        if cap is None or cap.get("cpu_fallback"):
+            return False
+        if str(cap.get("backend") or "").lower() not in (
+            "cuda", "rocm", "vulkan",
+        ):
+            return False
+        floor = int(cap.get("min_memory_bytes") or 0)
+        have = int(cap.get("free_memory_bytes") or 0)
+        return floor > 0 and 0 < have < floor
 
     def is_warm(self, engine: str, model_id: str) -> bool:
         return self.capacity.is_resident(engine, model_id)

backend/worker/protocol/gen/worker_v1_pb2.pyi

@@ -93,7 +93,7 @@ class HostInfo(_message.Message):
     def __init__(self, hostname: _Optional[str] = ..., os: _Optional[str] = ..., arch: _Optional[str] = ..., worker_version: _Optional[str] = ..., cpu_count: _Optional[int] = ..., system_memory_bytes: _Optional[int] = ..., gpus: _Optional[_Iterable[_Union[GpuInfo, _Mapping]]] = ...) -> None: ...
 
 class ModelCapability(_message.Message):
-    __slots__ = ("engine", "model_id", "operations", "supported", "installed", "downloaded", "resident", "min_memory_bytes", "precision", "derived_concurrency", "cpu_fallback", "repo_ids", "display_name")
+    __slots__ = ("engine", "model_id", "operations", "supported", "installed", "downloaded", "resident", "min_memory_bytes", "precision", "derived_concurrency", "cpu_fallback", "repo_ids", "display_name", "backend", "free_memory_bytes")
     ENGINE_FIELD_NUMBER: _ClassVar[int]
     MODEL_ID_FIELD_NUMBER: _ClassVar[int]
     OPERATIONS_FIELD_NUMBER: _ClassVar[int]
@@ -107,6 +107,8 @@ class ModelCapability(_message.Message):
     CPU_FALLBACK_FIELD_NUMBER: _ClassVar[int]
     REPO_IDS_FIELD_NUMBER: _ClassVar[int]
     DISPLAY_NAME_FIELD_NUMBER: _ClassVar[int]
+    BACKEND_FIELD_NUMBER: _ClassVar[int]
+    FREE_MEMORY_BYTES_FIELD_NUMBER: _ClassVar[int]
     engine: str
     model_id: str
     operations: _containers.RepeatedScalarFieldContainer[str]
@@ -120,7 +122,9 @@ class ModelCapability(_message.Message):
     cpu_fallback: bool
     repo_ids: _containers.RepeatedScalarFieldContainer[str]
     display_name: str
-    def __init__(self, engine: _Optional[str] = ..., model_id: _Optional[str] = ..., operations: _Optional[_Iterable[str]] = ..., supported: _Optional[bool] = ..., installed: _Optional[bool] = ..., downloaded: _Optional[bool] = ..., resident: _Optional[bool] = ..., min_memory_bytes: _Optional[int] = ..., precision: _Optional[str] = ..., derived_concurrency: _Optional[int] = ..., cpu_fallback: _Optional[bool] = ..., repo_ids: _Optional[_Iterable[str]] = ..., display_name: _Optional[str] = ...) -> None: ...
+    backend: str
+    free_memory_bytes: int
+    def __init__(self, engine: _Optional[str] = ..., model_id: _Optional[str] = ..., operations: _Optional[_Iterable[str]] = ..., supported: _Optional[bool] = ..., installed: _Optional[bool] = ..., downloaded: _Optional[bool] = ..., resident: _Optional[bool] = ..., min_memory_bytes: _Optional[int] = ..., precision: _Optional[str] = ..., derived_concurrency: _Optional[int] = ..., cpu_fallback: _Optional[bool] = ..., repo_ids: _Optional[_Iterable[str]] = ..., display_name: _Optional[str] = ..., backend: _Optional[str] = ..., free_memory_bytes: _Optional[int] = ...) -> None: ...
 
 class RegisterRequest(_message.Message):
     __slots__ = ("envelope", "protocol_version_min", "protocol_version_max", "enrollment_token", "worker_id", "public_key", "challenge_signature", "challenge", "host", "capabilities", "max_concurrent_tasks", "in_flight", "completed_unacked", "key_id", "nonce", "labels", "features")

backend/worker/protocol/worker_v1.proto

@@ -168,6 +168,11 @@ message ModelCapability {
   // Human-readable UI label. Never use this as a scheduling or residency key;
   // unlike model_id it may change with ordinary copy edits.
   string display_name = 13;
+  // Per-engine runtime routing. Native engines can select a provider that is
+  // independent of the worker's global torch device.
+  string backend = 14;
+  // Memory measured for that exact selected provider/device; zero is unknown.
+  uint64 free_memory_bytes = 15;
 }
 
 message RegisterRequest {

backend/worker/scheduler.py

@@ -652,7 +652,11 @@ class Scheduler:
             execution_device=worker.execution_device(
                 task.engine, task.model_id, task.operation
             ),
+            under_provisioned=worker.under_provisioned(
+                task.engine, task.model_id, task.operation
+            ),
         )
+        attempt.deadlines = budget
         attempt.renew_lease(budget.accept_seconds, now=now)
         self._save(task, now=now)
         self._emit("assigned", task)
@@ -1295,7 +1299,13 @@ class Scheduler:
 
     def _budget_for(self, task: Task) -> deadline_policy.Deadlines:
         attempt = task.active_attempt
+        if attempt is not None and attempt.deadlines is not None:
+            return attempt.deadlines
+        # A legacy attempt has no recorded device once its worker is absent.
+        # Cover both configured device classes instead of assuming the shorter
+        # CPU budget; for_task floors an under-provisioned GPU at max(CPU, GPU).
         worker = self.pool.get(attempt.worker_id) if attempt else None
+        unknown_legacy_worker = attempt is not None and worker is None
         return deadline_policy.for_task(
             task.operation,
             text=task.params.get("text"),
@@ -1303,7 +1313,13 @@ class Scheduler:
             input_seconds=float(task.params.get("input_seconds") or 0.0),
             execution_device=(
                 worker.execution_device(task.engine, task.model_id, task.operation)
-                if worker else None
+                if worker else ("cuda" if unknown_legacy_worker else None)
+            ),
+            under_provisioned=unknown_legacy_worker or bool(
+                worker
+                and worker.under_provisioned(
+                    task.engine, task.model_id, task.operation
+                )
             ),
         )
 

backend/worker/task_store.py

@@ -33,6 +33,7 @@ from core.db import db_conn
 from core.path_security import UnsafePath, resolve_within, safe_filename
 from worker.clock import resolve
 from worker.errors import ErrorClass, WorkerError
+from worker.deadlines import Deadlines
 from worker.lifecycle import Attempt, AttemptState, PriorityClass, Task, TaskState
 
 logger = logging.getLogger("omnivoice.worker")
@@ -67,6 +68,8 @@ def _row_to_attempt(row) -> Attempt:
         state=AttemptState(row["state"]),
         created_at=float(row["created_at"]),
     )
+    if row["deadlines_json"]:
+        attempt.deadlines = Deadlines(**json.loads(row["deadlines_json"]))
     attempt.accepted_at = row["accepted_at"]
     attempt.started_at = row["started_at"]
     attempt.finished_at = row["finished_at"]
@@ -131,6 +134,32 @@ INPUT_PARAM_KEYS: tuple[str, ...] = (
 # task records what was staged for it. The record is what makes the purge
 # exact: an input is deletable only when no surviving task still refers to it.
 INPUTS_DIRNAME = "inputs"
+
+
+def artifact_id_for(name: str) -> str:
+    """The id a staged input is known by, everywhere.
+
+    This is a PROTOCOL identifier, not a local path: it is persisted in
+    ``params_json``, handed to remote workers over gRPC, and matched against
+    what a later sweep finds on disk. ``os.path.join`` made it OS-specific, so
+    a Windows control plane stored and shipped ``inputs\\<sha>.wav`` — which a
+    Linux worker cannot resolve, and which stops matching the moment the same
+    data directory is opened on another OS. Always ``/``; ``resolve_within``
+    already treats both separators as structural, so resolution is unaffected.
+    """
+    return f"{INPUTS_DIRNAME}/{name}"
+
+
+def normalize_artifact_id(artifact_id: str) -> str:
+    """Compare ids written by any host on equal terms.
+
+    Rows staged by a Windows control plane before this was canonicalised carry
+    a backslash. The sweeper decides whether a file on disk is still
+    referenced by comparing ids, so without this an upgraded install would
+    read every legacy row as unreferenced and delete inputs that surviving
+    tasks still point at.
+    """
+    return (artifact_id or "").replace("\\", "/")
 INPUTS_PARAM_KEY = "inputs"
 
 _HASH_CHUNK_BYTES = 1024 * 1024
@@ -289,7 +318,7 @@ def stage_input(
             f"Could not read the task input {source!r}: {exc}"
         ) from exc
 
-    artifact_id = os.path.join(INPUTS_DIRNAME, f"{digest}{_extension(source)}")
+    artifact_id = artifact_id_for(f"{digest}{_extension(source)}")
     try:
         destination = resolve_within(base, artifact_id)
     except UnsafePath as exc:  # pragma: no cover — the id is ours, hex only
@@ -470,7 +499,7 @@ def _referenced_artifacts(conn) -> set[str]:
             continue
         for entry in entries:
             if isinstance(entry, dict) and entry.get("artifact_id"):
-                referenced.add(str(entry["artifact_id"]))
+                referenced.add(normalize_artifact_id(str(entry["artifact_id"])))
     return referenced
 
 
@@ -557,8 +586,7 @@ def purge_artifacts(
     except OSError:
         return removed
     for name in names:
-        artifact_id = os.path.join(INPUTS_DIRNAME, name)
-        if artifact_id in referenced:
+        if artifact_id_for(name) in referenced:
             continue
         path = os.path.join(inputs_dir, name)
         try:
@@ -635,12 +663,13 @@ def _upsert_attempts(conn, task: Task) -> None:
             "INSERT INTO remote_task_attempts "
             "(id, task_id, worker_id, session_epoch, attempt_number, state, progress, stage, "
             " error_json, created_at, accepted_at, started_at, finished_at, lease_expires_at, "
-            " grace_expires_at) "
-            "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?) "
+            " grace_expires_at, deadlines_json) "
+            "VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?) "
             "ON CONFLICT(id) DO UPDATE SET state=excluded.state, progress=excluded.progress, "
             " stage=excluded.stage, error_json=excluded.error_json, accepted_at=excluded.accepted_at, "
             " started_at=excluded.started_at, finished_at=excluded.finished_at, "
-            " lease_expires_at=excluded.lease_expires_at, grace_expires_at=excluded.grace_expires_at",
+            " lease_expires_at=excluded.lease_expires_at, grace_expires_at=excluded.grace_expires_at, "
+            " deadlines_json=excluded.deadlines_json",
             (
                 attempt.attempt_id,
                 attempt.task_id,
@@ -657,6 +686,7 @@ def _upsert_attempts(conn, task: Task) -> None:
                 attempt.finished_at,
                 attempt.lease_expires_at,
                 attempt.grace_expires_at,
+                json.dumps(attempt.deadlines.to_dict()) if attempt.deadlines else None,
             ),
         )
 
@@ -888,6 +918,8 @@ def purge_finished(
 
 __all__ = [
     "INPUTS_DIRNAME",
+    "artifact_id_for",
+    "normalize_artifact_id",
     "INPUTS_PARAM_KEY",
     "INPUT_PARAM_KEYS",
     "InputStagingError",

backend/worker/transport/codec.py

@@ -229,10 +229,27 @@ def capability_to_pb(cap: dict) -> pb.ModelCapability:
         cpu_fallback=bool(cap.get("cpu_fallback")),
         repo_ids=list(cap.get("repo_ids") or []),
         display_name=str(cap.get("display_name") or ""),
+        backend=str(cap.get("backend") or ""),
+        free_memory_bytes=int(cap.get("free_memory_bytes") or 0),
     )
 
 
-def capability_from_pb(message: pb.ModelCapability) -> dict:
+def capability_from_pb(
+    message: pb.ModelCapability, *, fallback_backend: str = ""
+) -> dict:
+    """Decode a capability, including protocol-v2 peers from before backend.
+
+    ``backend`` was added to the existing protocol-v2 message, so an older
+    peer legitimately sends its protobuf default (the empty string).  The
+    host-level GPU backend is the only compatible execution-device signal in
+    that payload.  A capability explicitly marked as a CPU fallback must stay
+    on CPU even when its host also has a GPU.
+    """
+    backend = str(message.backend or "").strip().lower()
+    if message.cpu_fallback:
+        backend = "cpu"
+    elif not backend:
+        backend = str(fallback_backend or "").strip().lower()
     return {
         "engine": message.engine,
         "model_id": message.model_id,
@@ -249,6 +266,8 @@ def capability_from_pb(message: pb.ModelCapability) -> dict:
         "cpu_fallback": message.cpu_fallback,
         "repo_ids": list(message.repo_ids),
         "display_name": message.display_name,
+        "backend": backend,
+        "free_memory_bytes": message.free_memory_bytes,
     }
 
 

backend/worker/transport/server.py

@@ -851,10 +851,12 @@ class WorkerServicer(pb_grpc.WorkerServiceServicer):
         epoch: int,
     ) -> pb.RegisterResponse:
         session = identity.issue_session(worker_id=worker.id, key_id=worker.key_id, epoch=epoch)
-        capabilities = [codec.capability_from_pb(c) for c in request.capabilities]
         host = codec.host_from_pb(request.host)
-
         backend = host["gpus"][0].get("backend", "") if host.get("gpus") else ""
+        capabilities = [
+            codec.capability_from_pb(c, fallback_backend=backend)
+            for c in request.capabilities
+        ]
         claimed_refs = {
             ref.attempt_id: codec.task_ref(
                 ref.task_id, ref.attempt_id, ref.session_epoch
@@ -2044,7 +2046,14 @@ class WorkerServicer(pb_grpc.WorkerServiceServicer):
                     session.worker_id,
                 )
                 return
-            caps = [codec.capability_from_pb(c) for c in update.capabilities]
+            worker = self.pool.get(session.worker_id)
+            fallback_backend = (
+                worker.capacity.backend if worker is not None else ""
+            )
+            caps = [
+                codec.capability_from_pb(c, fallback_backend=fallback_backend)
+                for c in update.capabilities
+            ]
             self._queue_capability_update(session, caps)
             return
 

deploy/dockerhub-overview.md

@@ -14,12 +14,21 @@ cloning, and cinematic video dubbing — fully local, with no cloud API keys or
 
 ![VoiceStudio — the open-source ElevenLabs alternative](https://raw.githubusercontent.com/debpalash/VoiceStudio/main/.github/assets/social-preview.png)
 
-VoiceStudio runs entirely on your own hardware (CUDA / MPS / ROCm / CPU
+VoiceStudio runs entirely on your own hardware (CUDA / ROCm / CPU
 auto-detect) — nothing is sent to the cloud. This image is the **headless
 web-server build**: a FastAPI backend serving a pre-built React UI over HTTP, so
-you can run it on a homelab box, a GPU server, or anywhere Docker runs and open
+you can run it on an AMD64 homelab box or GPU server and open
 the UI in a browser.
 
+**Architecture:** published images are **`linux/amd64` only**; there is no
+native ARM64 image. On Apple Silicon, use the
+[native macOS app](https://github.com/debpalash/VoiceStudio/blob/main/docs/install/macos.md)
+for Apple GPU acceleration; the Linux container cannot access the Mac's Apple
+GPU through MPS or MLX. Other ARM64 hosts need an AMD64 server or CPU emulation,
+which can be much slower. See the
+[architecture requirements](https://github.com/debpalash/VoiceStudio/blob/main/docs/install/docker.md#architecture)
+before pulling an image.
+
 > The Tauri desktop app's auto-updater and update-channel toggle are
 > **desktop-only** and do not apply to this image — to update, pull a newer tag
 > and recreate the container.
@@ -136,7 +145,7 @@ are mirrored on GHCR at
 - **📦 Batch Queue** — drop 50 videos and walk away; per-job progress.
 - **🤖 MCP Server** — drive VoiceStudio from Claude, Cursor, or any MCP client.
 - **🛡️ AI Watermark** — invisible AudioSeal (Meta) marking that survives compression.
-- **⚡ GPU Auto-Detect** — CUDA · MPS · ROCm · CPU, with auto-offload on ≤8 GB cards.
+- **⚡ GPU Auto-Detect** — CUDA · ROCm · CPU, with auto-offload on ≤8 GB cards.
 - **🧩 Extensible** — subclass `TTSBackend` to add any engine in ~50 lines.
 
 Multiple TTS engines ship out of the box (IndexTTS, CosyVoice, Supertonic-3, and

docs/RELEASING.md

@@ -147,3 +147,12 @@ There's no "revert update" flow for clients — they'll only see a *newer* versi
 3. Clients auto-update to the "new" v0.2.1 which is actually the old code.
 
 Ugly but it works. Better plan: test with Option B above before publishing the draft.
+
+## Retrying a partially published build
+
+Use GitHub Actions **Re-run failed jobs** for the same release run. On retries,
+the workflow removes only the current version's installers for that job's target
+before Tauri uploads them again. A macOS retry also replaces that architecture's
+versionless updater archive. Other versions, sibling platforms, and updater
+manifests remain intact. Inventory or deletion permission/network failures stop
+the job instead of hiding an upload collision.

docs/STRUCTURE.md

@@ -7,39 +7,71 @@ Every folder has a single job. Every file at the root earns its place.
 ```
 VoiceStudio/
 │
-├── README.md                    ⟵ user-facing overview
-├── CHANGELOG.md                 ⟵ release history
-├── LICENSE
+├── README.md / README_CN.md     ⟵ user-facing overview (English / Chinese)
+├── CHANGELOG.md                 ⟵ release history; release.yml extracts the tag's section verbatim
+├── CLAUDE.md / AGENTS.md        ⟵ the working contract for AI agents — keep the two in sync
+├── LICENSE, LICENSE-NOTICE.md, SPONSORS.md
 │
-├── pyproject.toml               ⟵ Python project manifest
+├── pyproject.toml               ⟵ Python project manifest (+ pytest / lint config)
 ├── uv.lock                      ⟵ Python lockfile
 ├── package.json                 ⟵ monorepo manifest (Bun workspaces + Turborepo)
-├── bun.lock                     ⟵ JS lockfile
+├── bun.lock                     ⟵ JS lockfile — repo-root, covers frontend/ too
 ├── turbo.json                   ⟵ turborepo pipeline
 │
+├── .coderabbit.yaml             ⟵ CodeRabbit PR review config (fed CLAUDE.md)
+├── greptile.json                ⟵ Greptile PR review config (fed CLAUDE.md)
+├── skills-lock.json             ⟵ pins the sources + hashes of .agents/skills/
+├── .gitleaks.toml               ⟵ secret-scan config
+├── .gitmodules                  ⟵ omnivoice-gallery submodule
+├── .python-version
 ├── .dockerignore                ⟵ Docker build context filter
 ├── backend.spec                 ⟵ pyinstaller spec (stays at root by pyinstaller convention)
 ├── alembic.ini                  ⟵ DB migration config (stays at root by alembic convention)
 │
-├── .env                         ⟵ user config; gitignored, .env.example is the template
-├── .gitignore
+├── .gitignore                   ⟵ a repo-local .env stays ignored, but user config is NOT
+│                                   kept here: the durable env file is ~/.config/omnivoice/env
+│                                   (backend/core/user_env.py), written by the Settings panel
 │
 ├── backend/                     ⟵ FastAPI server
-│   ├── main.py
-│   ├── api/routers/             HTTP endpoints (thin)
-│   ├── core/                    config, db, task queue, metrics
-│   ├── services/                business logic
-│   └── schemas/                 pydantic request/response shapes
+│   ├── main.py                  the one entry point; its boot order is load-bearing —
+│   │                            read the comments before reordering anything
+│   ├── api/routers/             39 routers, auto-included; thin HTTP/WS surface
+│   │   └── setup/               first-run wizard, model download
+│   ├── core/                    config, db, job queue, event bus, auth/CSRF, path security,
+│   │                            opt-in analytics, version, diagnostics
+│   ├── services/                78 modules of business logic — TTS, dubbing pipeline,
+│   │                            audio DSP, GPU gateway, engine routing, model lifecycle
+│   ├── engines/                 per-engine adapters: indextts, supertonic3, confucius4,
+│   │                            dots_tts, moss_tts_v15, pockettts, audiocpp,
+│   │                            omnivoice_gguf, omnivoice_subprocess, _asr_sidecar, _echo
+│   ├── worker/                  remote / distributed workers — scheduler, pool, routing,
+│   │                            breaker, capacity, plus protocol/ and inbound/
+│   ├── mcp_shim/                MCP server entry point (docs/mcp.md)
+│   ├── speech_client/           speech sidecar client entry point
+│   ├── schemas/                 pydantic request/response shapes
+│   ├── migrations/versions/     alembic revisions — every schema change goes through here
+│   ├── plugins/                 plugin drop-in point (see services/plugin_sdk.py)
+│   ├── hooks/                   pyinstaller runtime hooks
+│   ├── config/models.yaml       model catalogue
+│   └── tests/                   the isolated pytest session — see "Where tests live"
 │
 ├── frontend/                    ⟵ React 19 + Vite + Tauri desktop
+│   ├── package.json             THE app version — every other version file mirrors it
 │   ├── src/
 │   │   ├── pages/               one file per top-level view
-│   │   ├── components/          reusable UI
-│   │   ├── api/                 typed API clients
-│   │   ├── store/               Zustand slices
+│   │   ├── components/          reusable UI (+ audiobook/ clone/ dub/ gallery/ settings/ …)
+│   │   ├── ui/, lib/            shared primitives and helpers
+│   │   ├── api/                 typed API clients, one per router group
+│   │   ├── store/               Zustand slices (+ persisted-state migrations)
 │   │   ├── hooks/               custom React hooks
-│   │   └── utils/
-│   ├── src-tauri/               Rust desktop shell
+│   │   ├── i18n/locales/        the ONLY home for user-facing strings
+│   │   ├── config/, data/, assets/, utils/
+│   │   └── test/                vitest setup + visual-test helpers
+│   ├── e2e/, e2e-perf/, e2e-prod/   Playwright suites: functional, perf, packaged bundle
+│   ├── src-tauri/               Rust desktop shell — backend spawn/bootstrap, updater
+│   │   │                        channel, dictation shortcut, crash/reset/uninstall
+│   │   ├── capabilities/, icons/, wix/, debian/, appimage/   packaging inputs
+│   │   └── tests/
 │   └── public/
 │
 ├── omnivoice/                   ⟵ the underlying TTS model package
@@ -51,46 +83,59 @@ VoiceStudio/
 │   ├── training/
 │   └── utils/
 │
-├── tests/                       ⟵ all tests live here, no exceptions
-│   ├── conftest.py
-│   ├── test_api.py
-│   ├── test_dub_*.py
-│   ├── test_job_queue.py
-│   ├── test_segmentation.py
-│   └── frontend/                Node-based frontend tests
+├── tests/                       ⟵ the main pytest session (testpaths in pyproject.toml)
+│   ├── conftest.py              hermetic OMNIVOICE_DATA_DIR — never touches real app state
+│   ├── backend/, scripts/       mirrors of the source trees they cover
+│   ├── smoke/                   fast end-to-end checks (own CI job, HF_HUB_OFFLINE=1)
+│   ├── evals/                   quality evals (evals.yml)
+│   ├── probe/, fixtures/
+│   └── frontend/                Node-based frontend tests (legacy; vitest is the default)
 │
-├── scripts/                     ⟵ dev / build / release shell + python scripts
-│   ├── install.sh               universal installer (macOS/Linux/WSL)
-│   ├── install.ps1              universal installer (Windows)
-│   ├── run.sh                   universal launcher
+├── scripts/                     ⟵ dev / build / release scripts (shell, python, mjs)
+│   ├── install.sh / install.ps1 universal installers
+│   ├── desktop-*.mjs            dev, prod and fresh desktop launchers
 │   ├── smoke-test.sh            end-to-end validation
-│   └── desktop-prod.sh          production desktop build
-
+│   ├── check-docs-drift.py      the docs-drift.yml checker (docs/features.yaml is canonical)
+│   └── build-omnivoice-tts.sh   builds the bin/ sidecars
+│
+├── bin/                         ⟵ prebuilt omnivoice-tts sidecars, one per platform
+│
+├── .agents/skills/              ⟵ canonical skill copies (vite, fastapi-python), pinned by
+│                                   skills-lock.json — followed by path, never symlinked
+├── skills/                      ⟵ skills this repo publishes (omnivoice, oss-maintainer)
+│
 ├── infra/                       ⟵ edge/deploy workers (not the Docker deploy path)
 │   └── install-redirect/        voicestudio.sh/install — UA-sniffing installer worker
 │
 ├── deploy/                      ⟵ Docker deployment configs
-│   ├── Dockerfile               single-stage CUDA image
-│   └── docker-compose.yml       one-click local deployment
+│   ├── Dockerfile               CUDA by default; CI builds the ROCm variant from the same
+│   │                            file via BASE_IMAGE / GPU_FLAVOR overrides
+│   ├── docker-compose.yml       one-click local deployment
+│   ├── torch-constraints.txt    pinned torch resolution for the image
+│   └── dockerhub-overview.md    synced to the Docker Hub overview page at release
 │
 ├── docs/                        ⟵ developer docs, screenshots, branding
 │   ├── ROADMAP.md               where this project is going
 │   ├── STRUCTURE.md             you are here
-│   ├── mcp.json                 MCP config template
-│   ├── preview.png              README hero image
-│   ├── logo.png, logo.svg       branding assets
-│   ├── screenshot-*.png         feature screenshots
-│   ├── languages.md
-│   ├── training.md
-│   ├── data_preparation.md
-│   ├── evaluation.md
-│   └── voice-design.md
+│   ├── RELEASING.md             the release checklist — every deployment channel
+│   ├── features.yaml            canonical feature inventory (drives docs-drift.yml)
+│   ├── adr/                     architecture decision records
+│   ├── agents/                  agent-facing docs (issue tracker, triage labels, domain)
+│   ├── engines/, dubbing/, install/, setup/, migration/, features/, playbooks/, specs/
+│   ├── media/, screenshot-*.png, preview.png, logo.*
+│   └── languages.md, training.md, data_preparation.md, evaluation.md, voice-design.md
 │
-├── examples/                    ⟵ runnable demos + sample inputs
+├── examples/                    ⟵ runnable demos + sample inputs (agentic/, speech-platform/)
+│
+├── notebooks/                   ⟵ OmniVoice_Studio_Colab.ipynb
+│
+├── omnivoice-gallery/           ⟵ git submodule — the published voice gallery
 │
 ├── omnivoice_data/              ⟵ Docker bind-mount target (gitignored)
 │                                   DB + HF cache live here when running via compose
 │
+├── .github/workflows/           ⟵ ci, docker, release, security, docs-drift, evals,
+│                                   install-smoke, build-omnivoice-tts
 └── .git/
 ```
 
@@ -98,11 +143,22 @@ VoiceStudio/
 
 1. **Nothing at the root is a runtime artifact.** Outputs, temp files, local DBs, crash logs — all go to `~/Library/Application Support/OmniVoice/` (or the OS equivalent), *never* into the repo. The one exception is `omnivoice_data/`, which exists as a bind-mount anchor for Docker.
 
-2. **No ad-hoc scripts at the root.** One-off debug scripts live in `scripts/`. Tests live in `tests/`. Benchmarks live in `scripts/benchmarks/` (when we create them).
+2. **No ad-hoc scripts at the root.** One-off debug scripts live in `scripts/`. Tests live in one of the three homes below, never at the root.
 
 3. **Each subdirectory owns one concern.** If you can't describe what goes in a directory in one sentence, it's wrong.
 
-4. **Every package has a manifest.** `backend/`, `frontend/`, `omnivoice/` each have their own deps declared via `pyproject.toml` / `package.json` — they are independently testable.
+4. **Every package has a manifest.** `backend/`, `frontend/`, `omnivoice/` each have their own deps declared via `pyproject.toml` / `package.json` — they are independently testable. The JS lockfile is the **repo-root** `bun.lock` (Bun workspace), and `deploy/Dockerfile` installs from it with `--frozen-lockfile`.
+
+## Where tests live
+
+Three homes, each with its own runner. CI runs all three inside the single `test` job in
+`ci.yml`, as separate steps. The split is deliberate, not drift:
+
+| Home | Runner | Why it's separate |
+|---|---|---|
+| `tests/` | `pytest tests/` — the `testpaths` default | The main suite. Its `conftest.py` points `OMNIVOICE_DATA_DIR` at a throwaway dir so a run can never touch the developer's real app state (#878). |
+| `backend/tests/` | `pytest backend/tests/` — its own pytest session (the `Run pytest (backend/tests, isolated)` step) | Runs as an isolated session against `backend/`'s bare imports. Its `conftest.py` sets the same hermetic data dir; **never** reintroduce module-level `sys.modules` stubs there — they leak process-wide at collection time and poison mixed runs. |
+| `frontend/src/**/*.test.{js,jsx,ts,tsx}` | `bun run test` (vitest, jsdom) | Co-located with the component under test. `frontend/e2e*/` hold the Playwright suites; `tests/frontend/` is the older `node:test` set. |
 
 ## What lives where
 
@@ -110,10 +166,14 @@ VoiceStudio/
 |---|---|
 | User-facing product code | `backend/`, `frontend/` |
 | The TTS model (independent of the studio) | `omnivoice/` |
+| A new TTS/ASR engine adapter | `backend/engines/<engine>/` |
 | Everything executable but not user-facing | `scripts/` |
-| Tests | `tests/` |
+| Prebuilt platform sidecars | `bin/` |
+| Python tests | `tests/` (or `backend/tests/` when the isolated session is required) |
+| Frontend unit tests | next to the component, as `*.test.jsx` |
 | Developer + user docs (Markdown) | `docs/` |
 | Architecture decision records (ADRs) | `docs/adr/` |
+| Agent-facing docs | `docs/agents/` |
 | Runnable demos and sample data | `examples/` |
 | Runtime data (never committed) | `~/Library/Application Support/OmniVoice/` on Mac |
 
@@ -137,10 +197,10 @@ Removed in the 2026-07-12 cleanup pass (all preserved in git history):
 | Dir | Why it was there | Where it went |
 |---|---|---|
 | `.planning/` (74 files) | GSD-era planning archive: phases, quick plans, issue clusters. The GSD workflow was retired 2026-07-08. | Deleted; the four load-bearing decision docs moved to `docs/adr/`. |
-| `specs/` | spec-kit specs for features 001–007 — all shipped. | Deleted. |
+| `specs/` | spec-kit specs for features 001–007 — all shipped. | Deleted; `docs/specs/` is the current home. |
 | `design/` | ASCII mockups of the pre-React target UX, superseded by the shipped app. | Deleted. |
 | `research/` | Archived legacy Gradio UI + April-2026 competitor notes. | Deleted. |
-| `.agents/` | Rules for a third-party agent tool no longer in use. | Deleted. |
+| `.agents/` | Rules for a third-party agent tool no longer in use. | Deleted — then reintroduced with a different job: `.agents/skills/` now holds the canonical skill copies pinned by `skills-lock.json`. |
 
 ## Scaling path (proposed, not yet executed)
 
@@ -169,12 +229,13 @@ VoiceStudio/
 - `backend.spec` (`['backend/main.py']`, `pathex=['.']`)
 - `frontend/src-tauri/tauri.*.conf.json` sidecar paths
 - every import that reads `from backend.main import …` (tests, scripts)
+- `frontend/package.json` as the version source of truth, and the mirrors that track it
 
 Migrate when adding the second `apps/*` or the second `packages/*`. Not before.
 
 ## Conventions
 
 - **Filenames:** snake_case for Python, kebab-case or PascalCase for JS/TS components, lowercase for Markdown.
-- **Tests mirror source paths.** `backend/services/dub_pipeline.py` → `tests/services/test_dub_pipeline.py`.
+- **Tests mirror source paths** where a mirror exists: `tests/backend/` mirrors `api/ core/ engines/ services/`, so `backend/services/ffmpeg_utils.py` → `tests/backend/services/test_ffmpeg_utils.py`. Everything else stays flat — `tests/backend/test_*.py` for backend-wide cases, `tests/test_*.py` for cross-cutting ones. A React component's test sits next to the component.
 - **One-off scripts** go into `scripts/` with a descriptive name, not `test_*.py` at the root.
 - **New top-level directories** require a PR that updates *this file*.

docs/downloading-models.md

@@ -38,7 +38,19 @@ To keep that path **fast** despite Xet being off, the app runs a built-in
 **multi-connection (segmented) downloader on by default** — it fetches each file
 over parallel byte-ranges (IDM/uGet style), so the legacy-LFS path is no longer
 single-stream. It reports real live speed/ETA and **falls back to the normal
-download on any error**, so it can never compromise a correct install. Adding a
+download**, so it can never compromise a correct install.
+
+Ranges are capped at 16 MB each and eight of them are in flight at a time, so a
+completed range is committed to a resume manifest every few seconds. On a
+connection that drops mid-transfer, only the ranges in flight are refetched: the
+attempt is retried and the accelerator resumes from its manifest rather than
+starting the file over. An origin that does not serve ranges at all is handled
+inside the accelerator as a single stream, not as a failure.
+
+The accelerator is disabled for the rest of the install — handing over to the
+plain `snapshot_download` path — when it fails for a reason that is not
+transient network trouble, and on the install's final attempt, so it can never
+be the reason an install fails outright. Adding a
 free Hugging Face token (first-run setup, or Settings → Credentials) makes this
 faster still — authenticated downloads get higher rate limits and fewer stalls.
 To force the old single-stream path, set `OMNIVOICE_SEGMENTED_DOWNLOAD=0`.

docs/engine-acceptance.md

@@ -19,6 +19,7 @@ Every engine in the tree owns at least one job. A job has exactly one holder.
 | Fastest CPU render / lowest latency | *open — see #1306* |
 | Best Chinese/Japanese expressiveness | `cosyvoice`, `indextts2` |
 | CPU-realtime English, tiny footprint | `kittentts`, `supertonic3` |
+| Bilingual reference-free voice design and direction | `audiocpp` |
 | Best transcription accuracy | `whisperx`, `faster-whisper` |
 | Fastest Apple-Silicon transcription | `parakeet-mlx`, `mlx-whisper` |
 | Crash isolation for transcription | `faster-whisper-isolated` |
@@ -35,8 +36,11 @@ which is a property of the bar, not a judgement of the contributor.
    does not cover it. Latency, language, hardware envelope or quality tier —
    something a user would choose it *for*.
 2. **Licence clean for commercial use.** Model weights *and* code. No
-   research-only weights, no ambiguous provenance. This is the one that most
-   often ends a proposal, so check it first.
+   research-only weights or ambiguous provenance. The single approved
+   exception is `audiocpp` with Breeze-TTS-2 research/non-commercial weights,
+   approved by the owner on 2026-09-08 for its requested bilingual voice-design
+   and direction workflow. It remains opt-in and discloses the restriction
+   before selection and download; `@debpalash` is its named steward.
 3. **Every platform, or explicitly opt-in.** macOS (Apple Silicon and Intel),
    Windows, Linux. A CPU path is required — an engine that only runs on one
    accelerator is fine, but it must degrade rather than break, and a

docs/engines/README.md

@@ -5,6 +5,12 @@ quirks. Select engines in **Model Catalogue → Engines** (or quick-switch with
 <kbd>Ctrl</kbd>/<kbd>Cmd</kbd>+<kbd>E</kbd>), or pin one with
 `OMNIVOICE_TTS_BACKEND` / `OMNIVOICE_ASR_BACKEND`.
 
+When an engine reports itself unavailable, expand its row's **Why?** panel and
+use **Learn more** to jump straight to that engine's page here. The row's own
+message stays deliberately generic — an availability probe can carry local
+paths or credentials, so it is never shown verbatim — and the page below is
+where the actual requirements and setup steps live.
+
 The compute device (CUDA/ROCm/MPS/CPU) is auto-detected; pin it under
 **Settings → Performance & Device** (or `OMNIVOICE_DEVICE`) if auto-detect
 picks wrong — see [performance](../performance.md).
@@ -36,9 +42,10 @@ approval), [Windows](../install/windows.md), [Linux](../install/linux.md),
 | Supertonic-3 | [supertonic3](supertonic3.md) | CPU | — (7 preset voices) | `uv sync --extra supertonic` + license |
 | MOSS-TTS-v1.5 (8B) | [moss-tts-v15](moss-tts-v15.md) | CUDA · CPU | ✅ | clone + env var |
 | dots.tts (2B) | [dots-tts](dots-tts.md) | CUDA · CPU (not Windows) | ✅ | clone + env var |
-| OmniVoice (subprocess) | [omnivoice-subprocess](omnivoice-subprocess.md) | CUDA · MPS · CPU | ✅ | opt-in pick, no install |
+| OmniVoice (subprocess) | [omnivoice-subprocess](omnivoice-subprocess.md) | CUDA · MPS · CPU | ✅ | opt-in pick off MPS; automatic via default OmniVoice on MPS |
 | PocketTTS (Kyutai) | [pockettts](pockettts.md) | CPU (not Intel Mac) | ✅ | `uv sync --extra pockettts` + license |
 | Confucius4-TTS | [confucius4-tts](confucius4-tts.md) | CUDA · CPU | ✅ | clone + env var |
+| audio.cpp (Breeze-TTS-2) | [audio-cpp](audio-cpp.md) | CPU + Vulkan/Metal/CUDA/HIP/ROCm where compiled | ✅ + voice design | prebuilt binary + env var (weights research/non-commercial) |
 
 ## Speech-to-text
 

docs/engines/audio-cpp.md

@@ -0,0 +1,142 @@
+# VoiceStudio — audio.cpp Engine (Breeze-TTS-2)
+
+[audio.cpp](https://github.com/0xShug0/audio.cpp) is a pure-C++ ggml audio
+inference framework with prebuilt binaries for Windows, macOS, and Linux and
+no Python dependency. VoiceStudio discovers the compute providers compiled
+into the installed binary and selects the best available device. VoiceStudio drives its
+`audiocpp_server` over loopback HTTP — v1 serves the **`breeze_tts`**
+family: **Breeze-TTS-2** (BreezeBlue, 3B params, English + Chinese, voice
+clone + voice design + voice direction, 24 kHz).
+
+> **Opt-in, and never a default.** Select `audiocpp` explicitly in **Model
+> Catalogue → Engines** (or `OMNIVOICE_TTS_BACKEND=audiocpp`).
+
+## License — read before enabling
+
+- **audio.cpp code:** Apache-2.0.
+- **Breeze-TTS-2 weights** (upstream `BreezeBlue/Breeze-TTS-2` and the
+  `audio-cpp/audio.cpp-gguf` GGUF repack): **research and non-commercial
+  use only** under the [BreezeBlue Research and Non-Commercial License](
+  https://huggingface.co/BreezeBlue/Breeze-TTS-2/blob/main/LICENSE).
+  Self-hosted outputs inherit the restriction; a BreezeBlue paid
+  subscription covers hosted-platform outputs only, not this engine.
+
+## Platform support
+
+| Host | Binary | Compute |
+|---|---|---|
+| Windows x64 | CPU, Vulkan, CUDA 12.4/13.3 prebuilts | CPU, Vulkan, CUDA |
+| Linux x64 | CPU and Vulkan prebuilts | CPU, Vulkan; CUDA/ROCm from a self-build |
+| macOS arm64 | upstream Metal prebuilt | Metal + CPU |
+| macOS x64 | upstream `metal` archive (Metal disabled by upstream) | CPU |
+| Linux aarch64 | none upstream | unavailable in v1 |
+
+VoiceStudio runs `audiocpp_server --list-devices` once per installed binary,
+prefers CUDA, ROCm, Metal, then a discrete Vulkan GPU, and keeps CPU as a safe
+fallback. Device numbers are local to each backend registry. The Q8_0 GGUF is
+approximately 4.73 GiB, plus runtime memory.
+
+Allow about 6 GB of dedicated VRAM for CUDA, HIP, or a discrete GPU exposed
+through Vulkan. Metal and Vulkan integrated GPUs use unified-memory handling
+instead of that dedicated-VRAM floor.
+
+## Install
+
+1. Download the v0.7.2 prebuilt for your platform from
+   [audio.cpp releases](https://github.com/0xShug0/audio.cpp/releases/tag/v0.7.2)
+   and extract it. Use the Vulkan archive on Windows or Linux for broad GPU
+   support, the CPU archive when Vulkan is unavailable, or the matching CUDA
+   archive on Windows for NVIDIA. A Windows CUDA install needs both the
+   `bin-…-cuda…` and matching `cudart-…-cuda…` archives extracted into the
+   same directory, as required by upstream. VoiceStudio does not download
+   executable code for this engine. Linux archives do not preserve the
+   executable bit, so run `chmod +x audiocpp_server` after extracting one.
+
+   Verify the archive before extracting it. The pinned SHA-256 checksums are:
+
+   | Archive | SHA-256 |
+   |---|---|
+   | `audio-v0.7.2-bin-windows-x64-cpu-portable.zip` | `0b1f4bd78c5226ee3fa0eb24d95d603a429439cdf5dab45872d44a87412dd8c1` |
+   | `audio-v0.7.2-bin-windows-x64-vulkan.zip` | `15b8232eae740e21e507d87f827a89966de9451b085a45932d9e214e032962c1` |
+   | `audio-v0.7.2-bin-windows-x64-cuda12.4.zip` | `06c426095008022a2984ff1c75de4c9fab463c4201c0ff0a5dc4e14043f52326` |
+   | `audio-v0.7.2-cudart-windows-x64-cuda12.4.zip` | `7115be4d462817ad293f7932a8ac436d51023128e6728af09bba92a85593f393` |
+   | `audio-v0.7.2-bin-windows-x64-cuda13.3.zip` | `f975fec52745807b8c787e826c110acc3424a45156092455e1635604b68ec832` |
+   | `audio-v0.7.2-cudart-windows-x64-cuda13.3.zip` | `9b508f702636a9cdf3bf4dd8e75a86c20a0b87bdc39ca07e714c82f748efc1fa` |
+   | `audio-v0.7.2-bin-ubuntu-x64-cpu.tar.gz` | `6f5e43dd7b80e8ddf688ef84b411fadcd1f934d2c83963178bc4e2d9c4f07736` |
+   | `audio-v0.7.2-bin-ubuntu-x64-vulkan.tar.gz` | `fee1f978cee76453cf17f00196554bc2ee294645739538af0726a143b6a69a23` |
+   | `audio-v0.7.2-bin-macos-arm64-metal.tar.gz` | `c01e4f82971bedbe341697e63a9cebd5a5d1f72d5a9bcb51a3191f95ddab7a95` |
+   | `audio-v0.7.2-bin-macos-x64-metal.tar.gz` | `3862270f33439077225324169313f727064f727305b54d8ce920244d75ddcc24` |
+
+   Run `sha256sum <archive>` on Linux, `shasum -a 256 <archive>` on macOS,
+   or `Get-FileHash <archive> -Algorithm SHA256` in PowerShell and compare the
+   complete result with the table.
+2. Set `OMNIVOICE_AUDIOCPP_BIN` to the `audiocpp_server` binary
+   (`audiocpp_server.exe` on Windows):
+
+   ```bash
+   # macOS / Linux
+   echo 'export OMNIVOICE_AUDIOCPP_BIN=$HOME/apps/audio.cpp/audiocpp_server' >> ~/.zshrc
+   source ~/.zshrc
+   ```
+
+   Alternatively set `OMNIVOICE_AUDIOCPP_DIR` to the directory containing it.
+3. Restart VoiceStudio, open **Model Catalogue → Models**, find
+   **Breeze-TTS-2 Q8_0 for audio.cpp**, review its research/non-commercial
+   license note, and click **Install**. Generation never starts this ~4.73 GiB
+   download automatically.
+4. Pick `audiocpp` in **Model Catalogue → Engines**. The server starts
+   lazily on first generate (`server.json` + `server.log` live under the app
+   data `audiocpp/` directory).
+
+## Voice modes
+
+All three go through the one speech endpoint — reference presence selects:
+
+- **Clone:** `ref_audio` + `ref_text` (exact transcript, as upstream).
+- **Direction:** `ref_audio` + `ref_text` + `instruct`
+  (e.g. "Speak slowly with a restrained, serious tone").
+- **Design:** `description` (or `instruct`) with no `ref_audio`
+  (e.g. "A warm, thoughtful young woman…"). Upstream strengthens
+  instruction-following with `guidance_scale` ≈ 4.
+
+## Optional env knobs
+
+| Variable | Default | Purpose |
+|----------|---------|---------|
+| `OMNIVOICE_AUDIOCPP_BIN` | — | Absolute path to `audiocpp_server`. |
+| `OMNIVOICE_AUDIOCPP_DIR` | — | Directory containing `audiocpp_server`. |
+| `OMNIVOICE_AUDIOCPP_MODEL` | Model Catalogue cache | GGUF file or directory override. |
+| `OMNIVOICE_AUDIOCPP_PACKAGE` | `breeze-tts-2-q8_0.gguf` | Package filename (`…-bf16.gguf` for full precision). |
+| `OMNIVOICE_AUDIOCPP_PORT` | `17860` | Loopback port. |
+| `OMNIVOICE_AUDIOCPP_BACKEND` | Settings, then auto | Exact runtime: `cuda`, `hip`/`rocm`, `vulkan`, `metal`, or `cpu`. |
+| `OMNIVOICE_AUDIOCPP_DEVICE` | best device | Backend-local non-negative device index; requires `OMNIVOICE_AUDIOCPP_BACKEND`. |
+
+The audio.cpp overrides take precedence over the global Settings compute
+choice. A global CUDA/ROCm choice can match an NVIDIA/AMD GPU exposed through
+Vulkan. An unavailable explicit audio.cpp override is an error; an unavailable
+global preference falls back to CPU and is shown as a routing fallback.
+
+## Common errors
+
+### `audiocpp_server not found ...`
+
+The binary isn't installed. Follow **Install** — the message carries the
+exact release URL and SHA for your platform.
+
+### `audiocpp_server exited during startup ...`
+
+The managed loopback port may be taken. Check `server.log` next to
+`server.json` in the app data `audiocpp/` directory, or set a different
+`OMNIVOICE_AUDIOCPP_PORT` and restart VoiceStudio.
+
+### `Breeze-TTS-2 ... not installed` or `package ... not completely installed`
+
+Install the model from **Model Catalogue → Models**. If an interrupted install
+left it incomplete, use **Reinstall** there. If the error persists after a
+complete reinstall, file an issue with the package listing.
+
+---
+
+audio.cpp runs as a managed native server (no Python venv, no
+`transformers` conflict). Only the downloaded GGUF counts toward
+[sidecar disk usage](disk-usage.md).

docs/engines/confucius4-tts.md

@@ -25,6 +25,18 @@ is an LLM-based multilingual / cross-lingual zero-shot voice-cloning TTS.
 Like IndexTTS-2 / MOSS-TTS-v1.5 / dots.tts, it runs in its **own subprocess venv**
 so its dependency stack never touches the default VoiceStudio interpreter.
 
+## One-click install
+
+**Model Catalogue → Engines → Confucius4-TTS → Install** does the steps below
+for you, on Windows, Linux and macOS. It installs into its own folder under VoiceStudio's data directory, with its own Python environment. Nothing it installs touches VoiceStudio itself or any other engine, so you can switch to it and back without breaking what already worked. **Uninstall** in the same row removes only that folder. On an NVIDIA machine it installs
+the CUDA build of PyTorch; elsewhere it installs the CPU build. The ~5 GB of
+weights still download on first synthesis.
+
+The first synthesis downloads the weights, which takes a while on a slow
+connection. The generation stays alive while the download makes progress;
+if a stalled download runs out of time, raise the compute-time budget in
+**Settings → Performance & Device** and try again.
+
 ## Install
 
 ```bash
@@ -88,3 +100,13 @@ sr = model.sample_rate  # 22050
 - ✅ **Sidecar logic unit-tested** (`tests/test_confucius4_sidecar.py`):
   language normalization, tensor→PCM (mono/stereo/clip), config-path
   resolution, clone sys.path injection, wire framing, synthesize dispatch.
+
+## Accelerator routing
+
+The sidecar passes a runtime-available CUDA/ROCm, XPU, or registered NPU
+through upstream's device-aware model loading. The engine venv needs a matching
+PyTorch/vendor runtime, also noted in the catalogue install hint. XPU/NPU selection is covered by mocked loader and routing
+tests; this change does not certify synthesis on physical XPU/NPU hardware.
+MPS keeps the existing CPU fallback described in the validation record above.
+If modern accelerator detection or the legacy CUDA probe raises, loading falls
+back to CPU instead of aborting before model construction.

docs/engines/dots-tts.md

@@ -26,6 +26,17 @@ pins `transformers>=5.3` — the same isolation primitive used by
   which VoiceStudio does not auto-wire.
 - **VRAM:** ~9 GB checkpoint; a 12–16 GB CUDA GPU is the realistic target.
 
+## One-click install
+
+On Linux and macOS, **Model Catalogue → Engines → dots.tts → Install** does the
+steps below for you. It installs into its own folder under VoiceStudio's data directory, with its own Python environment. Nothing it installs touches VoiceStudio itself or any other engine, so you can switch to it and back without breaking what already worked. **Uninstall** in the same row removes only that folder. It is not offered on Windows, where upstream
+publishes no install. The ~9 GB checkpoint still downloads on first synthesis.
+
+The first synthesis downloads the weights, which takes a while on a slow
+connection. The generation stays alive while the download makes progress;
+if a stalled download runs out of time, raise the compute-time budget in
+**Settings → Performance & Device** and try again.
+
 ## Install
 
 dots.tts is **not** bundled (large checkpoint + conflicting `transformers`).
@@ -115,3 +126,9 @@ dots.tts runs in a dedicated sidecar venv (it pins `transformers==4.57`,
 which conflicts with the parent's `transformers>=5.3`). For why that adds
 disk and how uv keeps the cost down, see
 [Engine venvs & disk usage](disk-usage.md).
+
+The upstream runtime selects CUDA or CPU internally. Automatic precision follows
+that selection: bfloat16 on CUDA, float32 otherwise, including XPU/NPU/MPS hosts
+where this runtime executes on CPU. `OMNIVOICE_DOTS_TTS_PRECISION` remains an
+explicit override. If the CUDA availability probe raises, the automatic precision
+default stays float32; upstream remains responsible for its device selection.

docs/engines/moss-tts-v15.md

@@ -23,10 +23,24 @@ interpreter, so MOSS runs behind
   8 GB GPUs when quantized; the bf16 Transformers path used here is ~16 GB
   of weights, so a 16 GB+ GPU is the realistic CUDA target. It also runs on
   **CPU** (fp32) — correct but slow.
-- **Device:** CUDA when present, else CPU. **There is no MPS path** —
-  upstream documents only CUDA/CPU and the custom modelling code is
-  untested on Apple Silicon, so VoiceStudio never routes MOSS to MPS. On a
-  Mac it runs on CPU.
+- **Device:** the sidecar uses a runtime-available PyTorch CUDA/ROCm, XPU,
+  or registered NPU backend, otherwise CPU. The isolated engine venv needs the
+  matching torch/vendor integration. A failed accelerator probe falls back to
+  CPU, including in older venvs without the unified accelerator API. MPS still uses CPU. XPU/NPU routing is
+  covered by mocked loader tests; physical-device synthesis has not been
+  validated by this change.
+
+## One-click install
+
+On a machine with an NVIDIA GPU, **Model Catalogue → Engines → MOSS-TTS-v1.5 →
+Install** does every step below for you. It installs into its own folder under VoiceStudio's data directory, with its own Python environment. Nothing it installs touches VoiceStudio itself or any other engine, so you can switch to it and back without breaking what already worked. **Uninstall** in the same row removes only that folder. The ~16 GB of weights still
+download on first synthesis. On a CPU-only host the button is not offered; use
+the manual install.
+
+The first synthesis downloads the weights, which takes a while on a slow
+connection. The generation stays alive while the download makes progress;
+if a stalled download runs out of time, raise the compute-time budget in
+**Settings → Performance & Device** and try again.
 
 ## Install
 
@@ -48,9 +62,13 @@ into an isolated venv on demand.
    ```bash
    cd MOSS-TTS
    uv venv .venv
-   uv pip install -e ".[torch-runtime]"
+   uv pip install -e ".[torch-runtime]"      --extra-index-url https://download.pytorch.org/whl/cu128      --index-strategy unsafe-best-match
    ```
 
+   The extra pins `torch==2.9.1+cu128`, which is published only on PyTorch's
+   own index, so the `--extra-index-url` is required — without it uv reports
+   the requirements as unsatisfiable on every host.
+
    On a **non-CUDA / CPU host** (e.g. Apple Silicon), install plain
    `torch`/`torchaudio`/`transformers==5.0.0` into the venv instead of the
    `+cu128` extra (the auto-bootstrap below only targets CUDA hosts).

docs/engines/omnivoice-subprocess.md

@@ -32,12 +32,14 @@ lower call overhead.
 
 ## Selecting it
 
-- **Model Catalogue → Engines**, or
+- **Model Catalogue → Engines** on CUDA, ROCm, or CPU, or
 - `OMNIVOICE_TTS_BACKEND=omnivoice-subprocess`
 
-The explicit engine is opt-in on CUDA, ROCm, and CPU. Apple Silicon gets the
-same isolation automatically while keeping the default `omnivoice` id in APIs,
-Settings, and saved projects.
+The explicit engine is opt-in on CUDA, ROCm, and CPU. On Apple Silicon it is
+not listed separately: the canonical `omnivoice` choice automatically uses the
+same isolation while keeping that default id in APIs, Settings, and saved
+projects. Existing explicit `omnivoice-subprocess` configuration remains
+accepted for compatibility.
 
 ## Platform support
 

docs/engines/omnivoice.md

@@ -26,7 +26,11 @@ quantized native binary with a much smaller memory footprint.
   [#1222](https://github.com/debpalash/VoiceStudio/issues/1222)) the driver
   pages to system RAM and a render that should take seconds runs for minutes
   until the compute budget kills it. The UI warns before you wait; nothing
-  hard-blocks, since short inputs can still fit.
+  hard-blocks, since short inputs can still fit. A CUDA or ROCm card below the
+  floor is also budgeted as the CPU-class hardware it performs like — the longer
+  `OMNIVOICE_CPU_GENERATE_TIMEOUT_S` (600 s), not the accelerated 300 s
+  ([#1804](https://github.com/debpalash/VoiceStudio/issues/1804)). Apple Silicon
+  is excluded: unified memory has no dedicated pool to compare against.
 - No extra install — the model ships with the app and downloads its weights
   on first use (see [downloading-models.md](../downloading-models.md)).
 
@@ -88,9 +92,10 @@ The env var overrides the persisted UI choice.
   is ignored. Design is trained on English and Chinese and can be unstable
   in low-resource languages; for description-driven design in other cases
   try [VoxCPM2](voxcpm2.md).
-- Below the 6 GB VRAM floor, expect very slow renders or budget timeouts;
-  prefer [OmniVoice GGUF](omnivoice-gguf.md) or a CPU engine such as
-  [PocketTTS](pockettts.md).
+- Below the 6 GB VRAM floor on a CUDA or ROCm card, expect very slow renders;
+  they get the longer CPU compute-time budget rather than the accelerated one,
+  but can still time out. Prefer [OmniVoice GGUF](omnivoice-gguf.md) or a CPU
+  engine such as [PocketTTS](pockettts.md).
 
 ## Troubleshooting
 
@@ -106,3 +111,10 @@ See also: [benchmarks.md](../benchmarks.md),
 [performance.md](../performance.md),
 [expressive-speech.md](../expressive-speech.md),
 [disk usage](disk-usage.md).
+
+A timed-out subprocess is killed and given a bounded wait to exit before the
+request returns, so retrying cannot reuse its closing process. Timeout cleanup
+remains tied to the original child and cannot kill a replacement sidecar.
+If that wait cannot confirm exit, VoiceStudio retains the process for cleanup
+and blocks another attempt until it can be reaped, rather than starting a second
+engine alongside it. A later retry or shutdown retries the bounded cleanup.

docs/engines/pockettts.md

@@ -24,7 +24,13 @@ for this model.
    uv sync --extra pockettts
    ```
 
-   (Or enable it from **Model Catalogue → Engines**.)
+   Or click **Install** in **Model Catalogue → Engines → PocketTTS**. That
+   installs the same pinned package into the engine's own Python environment
+   under VoiceStudio's data directory, with the CPU build of PyTorch, because
+   PocketTTS never uses a GPU. Nothing it installs touches VoiceStudio itself
+   or any other engine, and **Uninstall** in the same row removes only that
+   folder. An install made with `uv sync` keeps working as it is. The button
+   is not offered on Intel Macs (see Platform notes).
 
 2. **Accept the license in-app**
    ([#1306](https://github.com/debpalash/VoiceStudio/issues/1306)). The code
@@ -50,8 +56,9 @@ for this model.
 - Output is 24 kHz mono.
 - Six languages, one model per language, chosen by the `language` you
   request; cloning takes a short reference clip.
-- Runs in a crash-isolated sidecar process (parent Python environment): a
-  wedged generation is hard-killed by a watchdog and its memory reclaimed —
+- Runs in a crash-isolated sidecar process: from its own environment after
+  a one-click install, otherwise from VoiceStudio's (where `uv sync --extra
+  pockettts` puts it). A wedged generation is hard-killed by a watchdog and its memory reclaimed —
   something an in-process engine cannot do.
 - The first use downloads the gated weights; the sidecar heartbeats
   progress during the download so the watchdog doesn't fire.

docs/engines/supertonic3.md

@@ -19,8 +19,11 @@ crashes and cold init never block the rest of VoiceStudio.
    uv sync --extra supertonic
    ```
 
-   (Or enable it from **Model Catalogue → Engines**, which installs the
-   pinned `supertonic` wheel for you.)
+   Or click **Install** in **Model Catalogue → Engines → Supertonic-3**. That
+   installs the same pinned wheel into the engine's own Python environment
+   under VoiceStudio's data directory. Nothing it installs touches VoiceStudio
+   itself or any other engine, and **Uninstall** in the same row removes only
+   that folder. An install made with `uv sync` keeps working as it is.
 
 2. **Accept the license in-app.** First use is gated behind an explicit
    acceptance dialog: the inference SDK is MIT, but the model weights are
@@ -45,9 +48,9 @@ log line.
 ## Behaviour notes
 
 - Output is 44.1 kHz mono.
-- Runs as a long-lived sidecar in the parent Python environment (its
-  dependencies — onnxruntime, numpy, soundfile — already match
-  VoiceStudio's pins); subsequent calls reuse the warm ONNX session.
+- Runs as a long-lived sidecar: from its own environment after a one-click
+  install, otherwise from VoiceStudio's (where `uv sync --extra supertonic`
+  puts it). Subsequent calls reuse the warm ONNX session.
 - `speed` is clamped to 0.7–2.0; quality steps clamp to 5–12.
 - Language is an ISO 639-1 code; Auto engages the SDK's multilingual
   fallback.

docs/expressive-speech.md

@@ -18,11 +18,18 @@ ignores.
 | Emotion ("excited", "sad", graded intensity) | IndexTTS2's emotion controls — Audiobook tab's Production Overrides, or the `/ws/tts` API — or CosyVoice 3 instruct | Opt-in engines only |
 | The same take again | Pin the seed / lock the profile | Default engine |
 
+## Recovering an interrupted audiobook
+
+Switching away from the Audiobook tab explicitly interrupts synthesis at a chapter boundary. The Audiobook recovery card lets you resume with its cached chapters, and **Open chapter cache** reveals the chapter audio cache.
+
 ## Why bracket tags work at all (and when they don't)
 
 Everything you type in the text box reaches the active engine **verbatim** —
 the pipeline goes out of its way not to break tags:
 
+- Tilde-separated integer, signed, and decimal ranges get a spoken separator in
+  English, Korean, Japanese, and Chinese; malformed chains and product codes
+  are left unchanged.
 - The text-normalization pass (numbers, abbreviations) skips every `[…]` span
   (`backend/services/text_normalization.py`).
 - The long-text chunker never cuts inside a bracket tag

docs/features.yaml

@@ -60,6 +60,8 @@ tts_engines:
     readme: "**Confucius4-TTS**"
     doc: docs/engines/confucius4-tts.md
   - id: pockettts
+  - id: audiocpp
+    doc: docs/engines/audio-cpp.md
 
 # Same contract against backend/services/asr_backend.py _REGISTRY.
 asr_engines:

docs/features/dictation.md

@@ -12,11 +12,19 @@ own microphone audio to the versioned WebSocket API. See the
 
 ## Use it
 
-1. Choose an installed dictation model in the Model Catalogue.
+1. Install a dictation model from the Model Catalogue, then pick it in **Settings → Voice** or from the top-bar Engines menu (**Transcription → Sherpa-ONNX dictation → Dictation model**). The same choice is what the Sherpa-ONNX engine loads for dubbing and batch transcription.
 2. Set the shortcut and hold/toggle behavior in **Settings → Hotkey**.
 3. Put the cursor in a text field, press the shortcut, speak, then release or
    press again.
 
+The **Transcriptions** page offers the same recorder as one contextual
+**Start dictation** action: it appears in the empty state before the first
+transcript and moves to the page header once history exists. A desktop start
+wakes the recorder window before dispatch, so a hidden WebView cannot silently
+miss the request. The in-app action confirms listener receipt, then resolves
+only after microphone startup is accepted. Disabled, rejected, timed-out, or
+failed starts are reported back on the page.
+
 Whisper Tiny is the recommended default on macOS, Windows, and Linux. It
 auto-detects more than 90 languages. Parakeet TDT v3 remains available for its
 25 supported European languages, but it is not selected automatically.
@@ -67,3 +75,13 @@ changes. `dotool` needs direct write access to `/dev/uinput`; `ydotool` 1.0+
 needs a running `ydotoold` with that access and a user-readable socket.
 VoiceStudio checks these prerequisites before selection. Tray-started Wayland
 dictation always stays copy-only.
+
+### Transcriptions model setup
+
+Transcriptions checks the active dictation model before enabling **Start dictation**. If weights are missing, the page lists every dictation model, grouped by the trade-off you are choosing between — best accuracy (offline, transcribes after you stop) versus lowest latency (streaming, live text while you speak) — with languages and download size on each row, so you install the one that fits your work rather than only the recommended default. A model already on disk can be switched to without a download. The chosen model's name, download size, and installation progress are shown. Downloads require an explicit click. Failed downloads can be retried, and model state refreshes when returning from Settings. Once installation is verified, Start dictation becomes available; recording never starts automatically. Existing transcription history remains accessible during setup.
+
+### Floating recording controls
+
+The recording bubble includes **Pause / Resume**, **Stop**, and **Close**. Pause disables microphone tracks, stops sending new audio, and freezes the elapsed recording timer while preserving the session. Resume continues the same session. Stop (or the recording shortcut) finishes and transcribes, including when paused. Close cancels pending recording/transcription and releases the microphone; text already delivered to another app cannot be retracted. Pausing is manual, not triggered by silence or desktop inactivity.
+
+The Transcriptions page displays the effective recording shortcut and the platform paste shortcut. The floating bubble places its live transcript below the controls in a multiline preview, so controls cannot squeeze the text into a few characters.

docs/hardware-notes-tesla-t4.md

@@ -31,7 +31,9 @@ request succeeds in ~1s (reproduced 5x: 1.574s / 1.034s / 1.065s / 0.995s / 0.91
   (`repo_id` is required — `InstallModelRequest` in `backend/api/schemas.py` rejects a bare/empty
   body — and must match one of the entries in `KNOWN_MODELS`, e.g. the default engine's
   `k2-fsa/OmniVoice`.) Progress streams over the existing `/setup/download-stream` SSE feed.
-- Or raise `OMNIVOICE_GENERATE_TIMEOUT_S` for the first request.
+- Or raise the compute-time budget in **Settings → Performance & Device** for the first
+  request (`OMNIVOICE_GENERATE_TIMEOUT_S` does the same thing from the environment, and
+  takes precedence over the setting when both are present).
 
 ## OpenAI-compatible endpoint doesn't expose `num_step` / `guidance_scale`
 

docs/install/docker.md

@@ -7,6 +7,25 @@ it in a normal browser.
 **Official images:** [`ghcr.io/debpalash/omnivoice-studio`](https://github.com/debpalash/VoiceStudio/pkgs/container/omnivoice-studio)
 and [`palashdeb/omnivoice-studio` on Docker Hub](https://hub.docker.com/r/palashdeb/omnivoice-studio) — same images, same tags.
 
+## Architecture
+
+The published images are **`linux/amd64` (x86-64) only**, including `:stable`,
+`:latest`, and the ROCm variants. There is no native `linux/arm64` image.
+On an ARM64 host, pulling without an explicit platform can fail with
+`no matching manifest for linux/arm64/v8 in the manifest list entries`.
+
+- **Apple Silicon (M-series Macs):** use the [native macOS app](macos.md),
+  which supports Apple GPU acceleration. The Linux container cannot access
+  the Mac's Apple GPU through MPS or MLX.
+- **AMD64 emulation on ARM64 (including Apple Silicon):** if your Docker
+  installation supports it, place `--platform linux/amd64` **before the image
+  name** in both `docker pull` and `docker run` from the CPU instructions
+  below. This is an emulated CPU option, not native
+  ARM64 support; inference can be much slower and is not a GPU workaround.
+  Without emulation, use an AMD64 server for this Docker deployment.
+
+## Image tags
+
 > **Image ↔ version mapping**
 >
 > | Tag | What you get |
@@ -237,6 +256,45 @@ docker compose -f deploy/docker-compose.yml --profile gpu up -d
 docker compose -f deploy/docker-compose.yml --profile rocm up -d
 ```
 
+> **ARM64 hosts:** Compose has no per-command `--platform` flag, so the
+> override that works for `docker pull` and `docker run` does not reach it.
+> Set `DOCKER_DEFAULT_PLATFORM=linux/amd64` in the shell you run Compose from,
+> or the image resolves to the ARM64 manifest that does not exist and fails
+> with `no matching manifest for linux/arm64/v8`. Only the CPU profile makes
+> sense under emulation — it is not a GPU workaround.
+>
+> ```bash
+> export DOCKER_DEFAULT_PLATFORM=linux/amd64
+> docker compose -f deploy/docker-compose.yml --profile cpu pull
+> docker compose -f deploy/docker-compose.yml --profile cpu up -d
+> ```
+>
+> In PowerShell, set both the administrator key and the platform for the
+> session before running the same two commands:
+>
+> ```powershell
+> $rng = [System.Security.Cryptography.RandomNumberGenerator]::Create()
+> try {
+>     $keyBytes = New-Object byte[] 32
+>     $rng.GetBytes($keyBytes)
+>     # URL-safe, like the `secrets.token_urlsafe` the Bash line uses: the
+>     # key is also accepted as an `?api_key=` query parameter, where a
+>     # raw Base64 `+` decodes to a space and silently mismatches.
+>     $env:OMNIVOICE_API_KEY =
+>         [Convert]::ToBase64String($keyBytes).TrimEnd('=').Replace('+', '-').Replace('/', '_')
+> }
+> finally {
+>     $rng.Dispose()
+> }
+> $env:DOCKER_DEFAULT_PLATFORM = 'linux/amd64'
+> docker compose -f deploy/docker-compose.yml --profile cpu pull
+> docker compose -f deploy/docker-compose.yml --profile cpu up -d
+> ```
+>
+> Either way the setting lives only in that shell and the processes it starts.
+> The [architecture limits](#architecture) still apply: this is emulated CPU
+> inference, not native ARM64 support.
+
 The `docker-compose.yml` shipped in `deploy/` defaults to `127.0.0.1:3900`
 on the host. The backend inside the container binds to `0.0.0.0` so the
 host port mapping can forward — the host-side `127.0.0.1` binding is what

docs/install/macos.md

@@ -190,3 +190,20 @@ Hit a wall? See [docs/install/troubleshooting.md](troubleshooting.md).
 The in-app error UI (the React error boundary that fires on backend errors)
 includes an **"Open docs for this error"** button — that button deeplinks
 back into this docs tree at the right section for the error class.
+
+### Desktop window chrome
+
+The main window uses native macOS traffic lights with an overlay title bar;
+window sizing, resize limits, and application file-drop behavior match the
+other desktop platforms. The platform configuration repeats the complete window
+list because Tauri replaces arrays when merging it with the shared config.
+The capture widget remains a separate borderless window created at runtime.
+Its window-scoped Tauri capability permits hiding after recording or idle
+reconciliation on every desktop platform.
+
+### Fast process shutdown
+
+A process that exits while shutdown is signalling it can report a macOS
+permission error. VoiceStudio accepts this only after confirming the original
+process exited without being reaped, then still waits for nested operations to
+drain. Live-process permission errors and lost process ownership remain failures.