fix(audio): remove subsampling from the zero-fill detector and share it across both paths

[evinjohnn]

Stack 7 of 7. Base: pr/windows-build-tooling — merge that first.

Change summary

Commit	Change
289553d8	Remove subsampling from the zero-fill detector and share it across both paths

Files: electron/audio/systemAudioHealthClassifier.mjs, systemAudioHealthClassifier.d.mts, electron/main.ts, plus SystemAudioHealthClassifier.test.mjs and ZerofillDetectorPeakToPeak.test.mjs.

Change detail

The zero-fill detector sampled a subset of frames, which aliases against periodic signals — a stream can read as silent when it is not, and vice versa. This removes the subsampling and consolidates the two divergent detector implementations into one shared path so both callers agree.

Cross-platform analysis

• Expected macOS behavior: identical. The classifier has no process.platform branch — it operates on PCM buffers after capture, downstream of the platform-specific capture layer.
• Expected Windows behavior: identical.
• Existing implementations reviewed: both the macOS system-audio capture path and the Windows WASAPI loopback path feed this classifier. The change is in shared post-capture analysis, not in either capture pipeline.
• Impact radius: system-audio health classification and whatever downstream logic reacts to a silent-stream verdict.

Validation

• Covered by automated macOS branch tests — the classifier tests have no platform branch, so one run covers both.
• Covered by automated Windows branch tests — same suite, executed on Windows.
• Reviewed but not executed on macOS
• Requires physical macOS verification — end-to-end system-audio capture behavior with the new detector.
• Requires physical Windows verification — this development machine has faulty audio hardware, so live capture cannot be trusted here as a signal either way.

Commands executed

git cherry-pick -x   # 1 commit, zero conflicts

Remaining risks

Removing subsampling means the detector now examines every frame, which is more CPU per buffer. The tests assert correctness, not cost — worth a glance at capture-loop timing under load on both platforms.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UWCt3EHyDEzRYhZXJwudSf

[greptile-apps]

Greptile Summary

Consolidates zero-fill amplitude analysis into a shared full-sample peak-to-peak detector.

• Removes subsampling that could alias periodic audio signals.
• Requires three above-threshold chunks before permanently latching meaningful audio on both system and microphone paths.
• Adds behavioral and structural regression coverage for DC bias, aliasing frequencies, transient chunks, and detector delegation.

Confidence Score: 5/5

The PR appears safe to merge, with the shared detector and both latch paths behaving consistently under the covered signal and silence scenarios.

The full-sample peak-to-peak scan removes the documented aliasing mechanism, both capture paths use equivalent cumulative latch behavior, lifecycle state resets with each wired capture, and no concrete changed-code failure remains.

Important Files Changed

Filename	Overview
electron/audio/systemAudioHealthClassifier.mjs	Replaces strided sampling with a complete PCM scan and introduces a configurable cumulative three-chunk meaningful-signal latch.
electron/main.ts	Reuses the shared detector in the microphone path and aligns its latch threshold with system-audio classification.
electron/audio/systemAudioHealthClassifier.d.mts	Declares the new meaningfulChunkCount classifier option.
electron/audio/tests/SystemAudioHealthClassifier.test.mjs	Adds behavioral coverage for DC offsets, aliasing tones, transient rejection, sustained audio, and latch configuration.
electron/services/tests/ZerofillDetectorPeakToPeak.test.mjs	Updates structural regression checks for the shared detector and lazy source extraction.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[PCM audio chunk] --> B[Scan every int16 sample]
  B --> C[Compute max minus min]
  C --> D{Peak-to-peak above 100?}
  D -- Yes --> E[Increment loud chunk count]
  E --> F{At least 3 loud chunks?}
  F -- Yes --> G[Latch meaningful signal]
  F -- No --> H[Continue observing]
  D -- No --> I{Observation window elapsed without latch?}
  I -- Yes --> J[Report sustained silence]
  I -- No --> H

Loading

_{Reviews (1): Last reviewed commit: "fix(audio): remove subsampling from the ..." | Re-trigger Greptile}