feat: hotato autopsy — zero-config forensics on any call recording

[quantumCF]

PR-A of the activation pivot: the front door becomes forensics. Bring a recording you already have; incidents in seconds; testing becomes the graduation path.

• hotato autopsy <recording> — zero config. Stereo runs the existing deterministic scan path unchanged and emits [CRITICAL]/[WARNING] incident blocks (BARGE-IN, TALK-OVER, DEAD AIR, LATENCY SPIKE, ECHO SUSPECTED) with timestamps, measured magnitudes, and the plain-English sentence. Critical bars reuse the same ceilings investigate label pins.
• Mono is best-effort and says so: energy-VAD silence timing with a measured confidence per finding (dB margin below the recording's own VAD threshold, derivation printed beside the number) and a one-line functional scope; mono can never emit BARGE-IN/TALK-OVER by construction; pure silence yields zero incidents. The mono refusal of run/scan/trust/contracts is untouched.
• One self-contained HTML report — waveform + incident markers + per-incident cards, zero scripts, zero external refs — at a content-addressed path (hotato-output/autopsy-apx-<12hex>.html), byte-identical on re-run with no SOURCE_DATE_EPOCH needed.
• Content-derived ids: apx-<sha256[:12]> of the input bytes, incidents addressed <id>#<rank> — the shape a future hotato pin resolves without re-running.
• est. cost renders only with --cost-config (operator figures, validated fail-closed); no config = zero dollar figures anywhere.
• Three rendered example calls (house seeded renderer, regenerate-on-absence) demonstrating barge-in-say-do, latency-dead-air, and talk-over; no-args prints usage + the quick start on the bundled example. sdist ships scenarios, never the WAVs.
• 17 new tests; docs/AUTOPSY.md indexed everywhere; CHANGELOG additive; llms-full regenerated.

Adversarially verified in an isolated venv (pass on all 9 checks): the 10-second test measured at 0.23s from a clean shell; two-run byte-identity on CLI text and HTML; mono-honesty incl. the no-fabrication silence probe; cost-guardrail greps on both surfaces; examples grep-verified as described rendered examples; investigate/trust behavior unchanged; full suite green.

Spec: hotato-growth/research/AUTOPSY-FIRST-SPEC-2026-07-23.md

https://claude.ai/code/session_015njkxtoAKioqTPT2k14xFr

[github-actions]

hotato turn-taking eval

8 of 8 scenarios pass. 0 fail. No regression.

scenario	expect	yielded	time to yield	talk over	result
01-hard-interruption	yield	yes	0.50s	0.50s	pass
02-backchannel-mhm	hold	no	-	1.57s	pass
03-filler-start	yield	yes	0.65s	0.56s	pass
04-correction	yield	yes	0.50s	0.50s	pass
05-telephony-8khz	yield	yes	0.50s	0.50s	pass
06-double-talk	yield	yes	1.05s	1.05s	pass
07-echo-bleed	hold	no	-	3.00s	pass
08-rapid-turn-taking	yield	yes	0.50s	0.50s	pass

Regressions

None.

_{Reproducible timing measured locally from call audio. Swap the bundled self-test step for your own captured recordings to gate on your agent. github.com/attenlabs/hotato}