fix(release): strip+filter bundle + report size (#13)

* ci: opt JavaScript actions into Node 24 runtime

GH deprecates Node 20 for JavaScript actions on 2026-09-16. The
deprecation warning surfaces on every run right now. Setting
FORCE_JAVASCRIPT_ACTIONS_TO_NODE24=true at workflow level makes
actions/checkout, actions/setup-*, astral-sh/setup-uv, and
oven-sh/setup-bun all run on Node 24 without bumping action versions.

This is a runtime override only — our own test script still pins
Node 22 via actions/setup-node@v4 (required for
--experimental-strip-types).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(release): strip symbols + filter CUDA/CUDA-provider binaries, report size

Previous slim pass (PR #11, CPU-only torch + module excludes) still left
the frozen backend above GH Releases' 2 GB per-asset cap. Two more levers:

1. strip=True on EXE + COLLECT. Strips debug symbols from ELF/Mach-O
   native libraries. libtorch_cpu.so and friends drop ~25-30%. No-op on
   Windows (MSVC stores symbols in separate .pdb files).

2. optimize=2 in Analysis. Compiles embedded bytecode with -OO:
   docstrings + assertions removed. ~50-80 MB off the PYZ archive.

3. Post-hoc binary filter after collect_all. Even with nvidia wheels
   excluded as Python modules, collect_all('torch')/('onnxruntime') can
   still pull the CUDA-runtime shared libs via their linker hints.
   Pattern-match them out of a.binaries before PYZ.

4. Log bundle size after freeze so CI runs can be compared without
   downloading artifacts.

If this round still overshoots 2 GB, the next step is splitting the
payload (thin installer + post-install download of the Python bundle).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/release.yml

@@ -192,6 +192,13 @@ jobs:
           uv run pyinstaller backend.spec --noconfirm --clean
           echo "Backend frozen at dist/omnivoice-backend/"
           ls dist/omnivoice-backend/ | head
+          # Report total bundle size so we can eyeball CI against GH's 2 GB
+          # per-asset cap on Releases uploads.
+          if [[ "$RUNNER_OS" == "Windows" ]]; then
+            powershell -Command "(Get-ChildItem -Recurse dist/omnivoice-backend | Measure-Object -Property Length -Sum).Sum / 1MB"
+          else
+            du -sh dist/omnivoice-backend
+          fi
 
       # ── Frontend build ─────────────────────────────────────────────────
       - name: Install frontend deps

backend.spec

@@ -179,8 +179,37 @@ a = Analysis(
         'numpy.tests', 'numpy.testing.tests',
     ],
     noarchive=False,
-    optimize=0,
+    # optimize=2 compiles the embedded stdlib + site-packages with -OO,
+    # stripping assert statements + docstrings. Saves ~50-80 MB on a bundle
+    # this size. Runtime impact is negligible because we never inspect
+    # docstrings at runtime.
+    optimize=2,
 )
+
+# Post-hoc binary filter: even on CPU-only torch wheels, collect_all pulls
+# a few large libraries the desktop runtime never touches. Drop them by
+# substring match on the archive name — PyInstaller re-runs when any of
+# these return False, so be precise (no matching "cuda" would eat too much).
+_DROP_BINARY_PATTERNS = (
+    # nvidia wheels (already in excludes, but collect_all can still pull their
+    # .so/.dll via torch's linker hints)
+    'libcudart.', 'libcublas.', 'libcublasLt.', 'libcudnn',
+    'libcurand.', 'libcufft.', 'libcusolver.', 'libcusparse.',
+    'libnccl.', 'libnvToolsExt.', 'libnvrtc.', 'libnvjitlink.',
+    'cudart64_', 'cublas64_', 'cublasLt64_', 'cudnn64_',
+    'curand64_', 'cufft64_', 'cusolver64_', 'cusparse64_',
+    # torch training / JIT runtimes not used by inference.
+    'libtorch_cuda', 'torch_cuda.', 'torch_cuda_linalg.',
+    # onnxruntime CUDA provider (we use CPU provider only).
+    'onnxruntime_providers_cuda', 'onnxruntime_providers_tensorrt',
+)
+
+def _keep_binary(entry):
+    name = entry[0].lower()
+    return not any(pat.lower() in name for pat in _DROP_BINARY_PATTERNS)
+
+a.binaries = [b for b in a.binaries if _keep_binary(b)]
+
 pyz = PYZ(a.pure)
 
 exe = EXE(
@@ -191,7 +220,10 @@ exe = EXE(
     name='omnivoice-backend',
     debug=False,
     bootloader_ignore_signals=False,
-    strip=False,
+    # strip=True removes debug symbols from ELF/Mach-O binaries (no-op on
+    # Windows since MSVC doesn't emit symbols in the same way). Saves
+    # 10-30% on native libraries like libtorch_cpu.so (~300 MB → ~220 MB).
+    strip=True,
     upx=False,              # UPX often corrupts ML native libs — disabled.
     console=True,
     disable_windowed_traceback=False,
@@ -204,7 +236,7 @@ coll = COLLECT(
     exe,
     a.binaries,
     a.datas,
-    strip=False,
+    strip=True,
     upx=False,
     upx_exclude=[],
     name='omnivoice-backend',