[bug] #490 added three paths that discard captured audio with no report — the idle recycle eats the clip and the live words, and a backgrounded tab discards every clip

[serge-ivo]

#490 added three paths that throw audio away with no report — the invariant #377/#420/#425 established

The owner: "sometimes it wasn't capturing or transcribing my voice even." These are the paths that can do that and leave no trace, so neither he nor the error log can tell it happened.

1. The idle recycle discards the clip AND the live words, silently

packages/sdk/src/voice/use-voice.ts:468-475

} else if (decision === "idle") {
	vadStateRef.current = initVad();
	idleRecycleRef.current = true;
	clearVoiceText();
	sttRef.current?.stopDiscard();
}

No gate consultation, no reportClientError, no bubble left behind. Compare the "end" branch thirteen lines above (use-voice.ts:455-458), which asks endOfTurnAction(gateSnapshot()) before discarding. The two paths disagree about whether a proven-alive Web Speech gate may vouch for the turn — and the idle path is the one that throws away both the audio and the words the user is watching appear on screen.

vadStep returns "idle" after 15 s when onset never fires (vad.ts:196-198; the caller passes no idleMs, so the 15 000 default applies — use-voice.ts:446).

User-visible signature: the pill pulses "Listening", live words appear, and 15 s later everything vanishes and the mic silently reopens. Saying it again usually works, because the VAD state is fresh. That is verbatim the #377 failure shape ("watched their own words appear and then vanish"), live again on this branch.

It also cannot produce a watchdog row: clearVoiceText() runs before stopDiscard(), so no turn is left in transcribing and the 50 s watchdog (use-voice.ts:70, :1803) never arms. There are zero (watchdog) rows in the last 83 error rows — consistent with this path, and by construction it could never appear there.

2. _peakLevel === 0 now discards — and a backgrounded tab guarantees it

packages/sdk/src/voice/stt.ts:364-372

if (this._peakLevel === 0) {
	// No energy data at all. Discard rather than uploading an unknown clip.
	this.onEnd();
	return;
}

The comment #490 deleted stated the old contract: "Skipped when nothing ever called noteLevel(), so a caller with no analyser keeps working rather than going silently deaf." #490 inverted exactly that.

Two problems:

• The comment above the branch claims the discard happens "AND there is no Web Speech gate available to vouch for the turn". There is no gate check in the branch. The comment describes a design the code does not implement.
• noteLevel has exactly one caller — the requestAnimationFrame tick at use-voice.ts:427. rAF callbacks do not run for a document that is not visible. With the console tab backgrounded or the window minimised — entirely normal for someone talking to a coding agent while watching a terminal — _peakLevel stays 0, the 60 s max-dictation cap fires (use-voice.ts:779-788), and every clip is discarded. Before [bug] STT transcribes silence in hands-free mode — Whisper hallucinations posted as phantom user turns ("Pottery Barn", "Thank you for watching") — gate fails on iOS Safari and when SpeechRecognition stalls #490 that clip transcribed and sent. The visibilitychange handler (use-voice.ts:1835-1843) only restarts on becoming visible; nothing stops the loop while hidden.

Inferred, not observed: no live row proves a hidden-tab discard occurred — and by construction there cannot be one. That is the point.

Why this belongs with #291

#377, #420 and #425 are all closed, and together they established the invariant that every drop path reports itself. #490 broke it in three places. #291 ("Stop swallowing runtime errors across voice, coder, and worker paths") is open and is exactly this subject — but both paths above postdate its last pass (50bea16, "the last 29 swallowed failures"), so they are uncounted by it.

Acceptance criteria

1. The idle-recycle path consults the gate the same way the "end" path does, or states in code why it must not.
2. Every path that discards captured audio reports itself to the durable log — including the idle recycle and the zero-peak discard.
3. The zero-peak branch either implements the gate check its comment describes, or the comment is corrected to match the code.
4. A backgrounded/minimised tab does not silently discard every clip. Either noteLevel gets a source that runs while hidden, or the zero-peak case is treated as "no analyser data" (the pre-[bug] STT transcribes silence in hands-free mode — Whisper hallucinations posted as phantom user turns ("Pottery Barn", "Thank you for watching") — gate fails on iOS Safari and when SpeechRecognition stalls #490 contract) rather than "no speech".
5. Voice error rows carry enough context to diagnose from the log alone: at minimum peakLevel, noiseFloor, onsetFloor and whether the gate heard speech. Today context is only {code} / {transcript,path} / {sttWhisper}, which is why this took a code read rather than a query.
6. A test for the idle path asserting it reports, and one for a zero-peak clip with a live gate.

[serge-ivo]

Fixed in 97bf5a4 (on main; CI green on that SHA; live in the console bundle — verified below).

What shipped, against the acceptance criteria

1. The idle-recycle path consults the gate. Both of vadStep's terminal decisions now go
   through one pure verdict, planTurnClose (turn.ts), on planNoiseRejection's rule: a
   proven-alive gate that heard words outranks the energy heuristic, and a gate that never ran
   vouches for nothing in either direction. Onset failing while Web Speech is returning text is a
   false negative of the energy gate, not silence, so that turn is transcribed instead of binned.
2. Discard paths report. With one deliberate exception, below.
3. The comment's gate check now exists. SttOptions.gate is wired from gateSnapshot, and it
   is the only thing that can condemn an unmeasured clip — which is what the comment always claimed.
4. A backgrounded tab no longer discards every clip. Took the second option you offered:
   planClipGate separates no measurement (frames === 0 → upload unless a live gate condemns
   it, the pre-[bug] STT transcribes silence in hands-free mode — Whisper hallucinations posted as phantom user turns ("Pottery Barn", "Thank you for watching") — gate fails on iOS Safari and when SpeechRecognition stalls #490 contract) from a measurement of zero (frames > 0 → real silence, still
   discarded, [bug] STT transcribes silence in hands-free mode — Whisper hallucinations posted as phantom user turns ("Pottery Barn", "Thank you for watching") — gate fails on iOS Safari and when SpeechRecognition stalls #490 intact). I did not give noteLevel a source that runs while hidden: the same
   rAF tick drives end-of-turn detection, so that change would also move when a hands-free turn
   ends, which is a bigger change than this ticket and wants its own issue.
5. Rows are diagnosable. Every drop path carries captureDiagnostics: peakLevel,
   noiseFloor, the onsetFloor they were judged against, frames, gateAlive,
   gateHeardSpeech. onsetFloorFor is extracted so the onset gate and the onstop gate cannot
   compute that threshold differently.
6. Tests: planTurnClose (idle-with-live-gate transcribes; end and idle agree on identical
   evidence; a dead gate neither saves nor condemns) and planClipGate (zero frames uploads;
   zero frames + live silent gate discards; measured zero still discards; gate outranks the floor).

Deviation from criterion 2, stated plainly

An idle-out with the peak under VOICE_FLOOR reports nothing. Hands-free recycles a quiet mic
every 15s while you think, so a row per recycle would bury the rows this issue exists to make
findable. Energy above the floor with no onset — the suspicious case — always reports. If you would
rather have every recycle counted, say so and I will add it at a lower level.

Verified in production

Deploy Host Worker succeeded and the code is live at https://proagentstore.online/console/. From
the deployed (minified) bundle:

function ss(e){let t=!!e.gate?.isAlive&&e.gate.heardSpeech,n=!!e.gate?.isAlive&&!e.gate.heardSpeech;
return e.frames===0?n?{action:`discard`,reason:`no-speech`}:{action:`transcribe`,reason:`no-data`}
:t?{action:`transcribe`,reason:`gate-vouched`}:...}
levelSnapshot(){return{peakLevel:...,noiseFloor:...,onsetFloor:as(this._noiseFloor),frames:this._levelFrames}}

Not verified — and what would close it

I have no microphone, so no drop was reproduced end to end. What is verified is the decision
logic (unit tests), that it is deployed, and that the rows now carry the numbers.

The measurement that closes this: use hands-free for a session, and when it misses you, pull
list_errors for source:voice. Each row now carries peakLevel / onsetFloor / frames /
gateAlive. Three signatures, three different answers:

• frames: 0 → the hidden-tab path. Confirms the rAF-while-hidden work is worth doing.
• peakLevel present and below onsetFloor → the gate is still too strict; that is [bug] The #490 noise gate is wrong in both directions — no_speech_prob is dead code on the default model, and the adaptive floor discards peaks the same file calls speech #511 and
  the numbers re-derive the constant directly.
• path: "idle" with energy → onset never fired on speech that was there.

Before this commit none of those rows existed, which is why the issue needed a code read.