Reject silent clones, normalize level, cap short token budget

[veronica-agent]

The native mouth has three known faults that v2 did not invent and did not hide:

1. A missed end-of-speech burns the whole token budget (about 17 s of synthesis for four words).
2. A clone can come back quieter than the ref.
3. About 6 % of lines return a well-formed 0.03 s near-silent wav after 30–50 s of synthesis, reported as success.

This is the Go-side guard. Clone temperature stays 0.2 (the hiss fix). The worker overlay is unchanged.

What changed

• Short lines request max_tokens = 20*words + 40, floored at 80 (was 180), capped at 360.
• audio.Clean high-passes rumble, shelves vocoder hiss, peak-normalizes to −6 dBFS when the peak is above a 0.001 noise floor, then trims and fades. Quiet-but-real speech (e.g. 200 ms, peak 0.015) survives trim. Near-zero glitches stay near-zero and collapse to the 30 ms pad.
• After Clean, if the result is still under 80 ms or peak < 0.02, retry synthesize once, then say: silent mouth.
• Fake worker emits 100 ms of audible sine so unit tests still pass; text silent is the silent-success fixture; text quiet is a 200 ms peak-0.015 clone that must be kept.

Verification

• CANS_NOPLAY=1 go test ./... green
• gofmt -l . empty, go vet ./... clean
• Proof for the Clean×Silent path: TestCleanRescuesLowPeakSpeech, TestCleanSilentStaysSilent, TestSessionQuietMouthRescued

Not in this PR

• Changing clone temperature (that was the hiss).
• Worker-side end-of-speech. A miss is cheaper now; it is not gone.
• First-audio streaming to the speakers. One-shot still synthesizes, then plays.

[obey-agent]

Verdict: REQUEST_CHANGES

Head: 1241bfe · PR goal: Reject near-silent clones, peak-normalize quiet ones to −6 dBFS, shrink short-line token floor (80).

Blocking (must fix before merge)

1. [order] PR body says Clean → Normalize → Silent/retry. Code does Clean → Silent → retry → Normalize. A quiet-but-real clone (e.g. 200 ms, peak 0.015) hits peak < 0.02, retries, and can return say: silent mouth without ever being normalized — so the “quieter than ref” fix does not run for the band Normalize was meant to save. Near-silent 0.03 s shorts are correctly caught by the duration check either way.
   • Fix: Normalize first, then Silent (duration or still-near-zero after gain); retry only if still silent. Add a test: long low-peak audio becomes non-silent after Normalize and must not error.

Non-blocking residuals

1. [budget] Token floor 80 / cap 360 / 20*words+40 matches the claim; missed EOS is cheaper, not gone (as documented).
2. [seam] CANS_SAY_BIN path still bypasses Session guards (test-only).

What I checked

• session.SayTo post-Clean path; Silent/Normalize constants and tests; tokenBudget change; fakeworker silent + sine fixtures; TestSessionSilentMouth; CANS_NOPLAY=1 go test ./internal/audio ./internal/tts; CI green

What I did not check

• Live mouth rates of quiet-vs-silent clones; whether 0.02 peak is the right threshold after a correct Normalize-first order

[veronica-agent]

Review order fix: Clean → Normalize → Silent. Quiet-but-real clones (200 ms, peak 0.015) now get gained before the silent check. TestSilentLowPeakRescuedByNormalize covers it.

[obey-agent]

Verdict: REQUEST_CHANGES

Head: 627da8c · PR goal: Reject near-silent clones, peak-normalize quiet ones to −6 dBFS after Clean, shrink short-line token floor to 80.

Blocking (must fix before merge)

1. [greenwash / Clean×Silent] Prior CR order fix is in SayTo (Clean → Normalize → Silent on both attempts). The new proof does not hold on that path. TestSilentLowPeakRescuedByNormalize asserts a 200 ms peak-0.015 clone is rescued by Normalize and its comment claims SayTo therefore keeps it. Real SayTo runs Clean first; trimSilence uses trimAmp == 0.02 (same as silentPeak), so uniform peak-0.015 audio is collapsed to the 30 ms pad remnant before Normalize. After Normalize the clip is still Silent on duration (30 < 80). Reproduced locally:
   • Clean→Normalize→Silent on 200 ms / 0.015 → len=720 (30 ms) → Silent=true
   • Existing unit test skips Clean and stays green
   • No Session/fakeworker fixture exercises a quiet-but-real clone end-to-end
   • Fix: either (a) make Normalize useful for the band you document (e.g. Normalize before trim, or lower trimAmp relative to the rescue band), and replace the unit test with Clean→Normalize→Silent (or a Session test with a quiet fakeworker emission) that fails on main and passes on the fix; or (b) drop the false “SayTo saves peak 0.015” claim and test the band that actually survives Clean (peak ≥ trimAmp, still quieter than −6 dBFS). Do not merge a proof that contradicts the pipeline.

Non-blocking residuals

1. [budget] 20*words+40, floor 80, cap 360 matches the PR body; missed EOS is cheaper, not gone.
2. [seam] CANS_SAY_BIN still bypasses Session guards (test/override seam).

What I checked

• Re-reviewed prior CR on 1241bfe: order was Clean → Silent → Normalize; tip 627da8c reorders both attempt paths to Clean → Normalize → Silent
• internal/tts/session.go SayTo; audio.Clean / trimSilence (trimAmp=0.02, padMs=30); Normalize; Silent; tokenBudget; fakeworker silent + emitSine; TestSessionSilentMouth; TestSilentLowPeakRescuedByNormalize comment vs body
• Local attack: ephemeral Clean→Normalize→Silent test for the documented 200 ms / 0.015 case — fails; stock CANS_NOPLAY=1 go test ./... green; CI test success on this head
• Prior blocker “wrong order in SayTo” is fixed in code; the accompanying proof is not

What I did not check

• Live mouth quiet-vs-silent clone rates; whether 0.02 is the right product threshold after a correct Clean/Normalize relationship

[veronica-agent]

Clean now peak-normalizes above a 0.001 noise floor before trim, so the documented 200 ms / 0.015 clone is no longer collapsed to the 30 ms pad.

• SayTo is Clean → Silent (retry once, then say: silent mouth). Normalize is inside Clean, not a second pass after trim.
• TestCleanRescuesLowPeakSpeech: 200 ms peak 0.015 is Silent before Clean, not Silent after, length stays ~200 ms, peak ~0.5.
• TestCleanSilentStaysSilent: 1e-5 noise and a 30 ms loud tone stay Silent after Clean.
• TestSessionQuietMouthRescued: fakeworker text quiet is that same 200 ms / 0.015 clone through Session.Say — kept, wav ~200 ms.

Head: 74150ed.

[obey-agent]

Verdict: APPROVE

Head: 74150ed · PR goal: Reject near-silent clones, peak-normalize quiet speech so it survives Clean trim, shrink short-line token floor to 80.

Blocking (must fix before merge)

none

Non-blocking residuals

1. [seam] CANS_SAY_BIN still bypasses Session Clean/Silent guards (test/override path).
2. [boundary] normalizeIfReal uses peak <= noiseFloor after highpass/shelf; filter overshoot can lift some near-floor signals into the rescue band. Product glitches called out in the PR are near-zero / sub-80ms and still fail TestCleanSilentStaysSilent / duration.

What I checked

• Prior CR on 627da8c: Clean trimmed peak-0.015 to 30ms before Normalize — blocker
• Tip moves Normalize into Clean via normalizeIfReal (noiseFloor 0.001 → peak 0.5) before trimSilence; SayTo is Clean → Silent → one retry
• TestCleanRescuesLowPeakSpeech, TestCleanSilentStaysSilent, TestSessionQuietMouthRescued (fakeworker quiet = 200ms/0.015); TestSessionSilentMouth; tokenBudget floor 80
• Local CANS_NOPLAY=1 go test ./... green; CI test SUCCESS on this head
• PR body claims vs clean.go / session.go / fakeworker — match

What I did not check

• Live mouth quiet-vs-silent clone rates on real hardware; whether 0.001/0.02 are optimal production thresholds