test(ci): root-cause and neutralize the flaky trio — a leaked torch fp16 default dtype (#1021)

test_effects_chain / test_generation_audio_guard / test_persona_bundle
failed intermittently on CI (never locally) with identical signatures
across three unrelated PRs today (#1002, #1019, #1016) — costing a
full CI cycle per occurrence and repeatedly muddying merge decisions.

Root cause, confirmed by local reproduction: a leaked
torch.set_default_dtype(torch.float16) from some earlier test in the
CI-Linux ordering. The smoking gun was test_generation_audio_guard's
observed 0.0999755859375 — exactly float16(0.1), i.e.
torch.tensor([0.1, …]) built under a leaked fp16 default. Reproducing
with a simulated polluter locally produced the trio's exact failures:
Pedalboard refuses fp16 audio outright ("only supports 32-bit and
64-bit floating point") and silently returns unmodified audio for
every preset, so test_effects_chain's preset outputs compare
identical; and the fp16 tensor value breaks the sanitize approx-check.

Fix: an autouse conftest guard (same philosophy as the existing
LLM-state isolation guard, #878) that checks torch's default dtype
after every test, resets any leak to float32, and emits a UserWarning
naming the offending test's nodeid — so the actual CI-only polluter
identifies itself in the next CI log instead of being chased blind.
Regression test: a deliberate-leak pair proving reset-between-tests.

Fail-before/pass-after verified: with the guard stashed, a simulated
polluter + the trio reproduced 2/3 failures locally with the exact CI
signatures; with the guard active, 73/73 pass and the warning names
the polluter.

Co-authored-by: mergetest <test@local>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

tests/conftest.py

@@ -37,6 +37,38 @@ if not os.environ.get("OMNIVOICE_ENV_FILE"):
 
 
 import pytest
+import warnings as _warnings
+
+
+# ── torch default-dtype isolation (CI flaky trio) ───────────────────────────
+# Three tests (test_effects_chain / test_generation_audio_guard /
+# test_persona_bundle) fail intermittently on CI — never locally — with
+# signatures that all trace to one cause: a leaked
+# `torch.set_default_dtype(torch.float16)` from some earlier test. The
+# smoking gun is test_generation_audio_guard's observed value
+# 0.0999755859375, which is exactly float16(0.1): `torch.tensor([0.1, …])`
+# built under a leaked fp16 default. The same leak collapses
+# test_effects_chain's preset differences into identical quantized outputs,
+# and hands test_persona_bundle's soundfile writer fp16 data libsndfile
+# can't encode. The polluter only executes on CI-Linux (it never reproduces
+# on macOS), so rather than chase it blind, this guard makes the whole leak
+# class impossible — same philosophy as the LLM-state guard below — and
+# names the offender in CI output when it fires, so it CAN be chased.
+@pytest.fixture(autouse=True)
+def _torch_default_dtype_guard(request):
+    yield
+    torch = sys.modules.get("torch")
+    if torch is None:
+        return
+    if torch.get_default_dtype() is not torch.float32:
+        _warnings.warn(
+            f"{request.node.nodeid} leaked torch default dtype "
+            f"{torch.get_default_dtype()} — resetting to float32. This is "
+            f"the polluter behind the CI flaky trio; fix it at the source.",
+            stacklevel=1,
+        )
+        torch.set_default_dtype(torch.float32)
+
 
 # ── LLM-provider state isolation (issue #878) ──────────────────────────────
 # LLM provider selection is process-global three ways: env vars (the

tests/test_torch_dtype_guard.py

@@ -0,0 +1,31 @@
+"""The conftest torch-dtype guard resets a leaked default dtype between tests.
+
+The CI "flaky trio" (test_effects_chain / test_generation_audio_guard /
+test_persona_bundle) failed intermittently on CI-Linux with signatures that
+all trace to one leak: some earlier test leaves
+``torch.set_default_dtype(torch.float16)`` behind. Reproduced locally with a
+simulated polluter — ``torch.tensor([0.1, …])`` under fp16 yields exactly the
+0.0999755859375 CI observed, and Pedalboard refuses fp16 audio outright
+("only supports 32-bit and 64-bit floating point"), silently returning
+unmodified audio for every preset so their outputs compare identical.
+
+These two tests are order-dependent BY DESIGN (pytest runs tests within a
+file in definition order): the first leaks, the second proves the autouse
+guard in conftest.py reset the leak before the next test began.
+"""
+import torch
+
+
+def test_a_deliberate_dtype_leak():
+    # Simulates the CI polluter. The conftest guard must clean this up (and
+    # emit a UserWarning naming this exact test as the offender).
+    torch.set_default_dtype(torch.float16)
+    assert torch.get_default_dtype() is torch.float16
+
+
+def test_b_next_test_starts_back_at_float32():
+    # If the guard is ever removed/broken, this fails — and so, eventually,
+    # does the flaky trio on CI, much less legibly.
+    assert torch.get_default_dtype() is torch.float32
+    # The exact fp16 signature the trio's CI failures showed, as documentation:
+    assert torch.tensor([0.1]).item() != 0.0999755859375