feat(anti): add bounded PCM WAV listening profile

[ch1ps3t]

Anti previously discarded audio input rather than forwarding it. This adds explicit local PCM16 WAV capture and a bounded advisory listen profile: one/two clips, selected Gemini model, upload opt-in, one helper generation and one backend attempt, at most2,048 generation tokens and90seconds. Original bytes/MIME/order are preserved; malformed/oversized media and unsupported routes refuse. Listen disables retry, fallback, automatic route selection, account rotation and source pre-reading. Tracks#98; Keyspilli’s consumer is draft#216.

Supported native Gemini text.format requests now reach upstream as JSON MIME/schema fields, without weakening constraints through the function-tool sanitizer. Unsupported constraints and strict guarantees refuse before account acquisition; other native families/image routes do not advertise effective structured output. Explicit helper --response-schema includes canonical schema bytes in prompt policy identity and records their hash. Reported Google thought counts survive streamed/non-streamed gateway/helper usage; missing counts remain unknown and listen’s total generation ceiling stays2,048. Complete output and meaning still require caller validation.

Exact advertised model IDs take priority over equivalent aliases. Previously set iteration could rewrite gemini-3.1-pro to gemini-3.1-pro-high, or add high reasoning to a base Flash request. Alias fallback remains available when the exact ID is absent. Regression tests force the alias-first ordering and verify both exact IDs and unchanged Flash alias behavior.

Validation for 56ade46237c6e57efbd7aed37e57e1e3c0b2f6d6: guarded local routing/helper/WAV/checkpoint/run-control/local-only checks268passed/2subtests; focused correctness lint and compilation pass. All11checks pass in CI37227468017, including Linux/macOS/Windows, Python3.10 compatibility and installed wheel/rebuilt-sdist checks. The Linux Python3.10 suite reports3,647passed/245subtests/1skip. Earlier native schema/request/protocol178checks and helper/WAV/usage155checks, owned fake upstream assertions, asset manifest and capability checks remain in the implementation evidence; the preceding b190db0 guarded full local run passed3,646tests/243subtests/1skip on Python3.14.

Live Pro recognition remains unqualified. A separately authorized consumer-compatible screen accepts bare JSON or one exact JSON fence, preserving strict schema, content and control checks. Of three serial calls, the missing-audio control passed and the two authored speech clips were transcribed exactly in order. The same two pinned WAVs supplied in reversed order then produced HTTP 200 / STOP with a raw “missing” response and zero clips. The mandatory control failed and permanently stopped the screen; the three piano controls were not sent. Prepared requests differ only in audio-part order, with prompt/schema/system instruction/route/settings/timeouts identical, and raw provider text equals helper output. This establishes a successful speech request but does not establish reliable audio delivery or recognition. Account/project lease identity was not captured or pinned across separate calls; one trial per condition cannot isolate those effects or establish that reversal caused the failure. Exactly three provider POSTs, two audio requests / four media parts, no retry/fallback/judge; cumulative follow-up calls are 21. Source/runner/artifact/prior-ledger pins and all 63 frozen corpus pins remain unchanged; the owned gateway is closed. Local evidence is in .superpowers/sdd/2026-10-04-anti-recognition-pro-consumer-format/ (frozen protocol 9ddd0f0fc346eb687f9d7dec4a2c7cb211f604af6b6dd615abc543816e8452a1).

A subsequent separately authorized single-clip speech screen passed all three frozen controls: truthful missing audio, then exact authored transcription of each clip supplied individually as A. All three requests used the same account/project identity within that session, enforced before dispatch through an in-memory equality gate and recorded only as opaque lease-1; no email, project ID, token or header was serialized by the instrumentation. Exactly three provider POSTs, two audio requests / two parts, HTTP200/STOP and response-reported gemini-3.1-pro-low, no retry/fallback/judge. All raw provider text equals helper output; each complete JSON fence was decoded locally. Generation including reported thought counts stayed within2,048 and each call within90seconds. Cumulative follow-up calls now24/24; owned gateway37701/port51779 is closed. The harness passed29offline admission/grammar/grading/leased-transport checks; application source is unchanged. Evidence root .superpowers/sdd/2026-10-04-anti-recognition-pro-single-clip/, protocol 280205e39e5e458a5423b37cec6188ccf7392196fc58686ca73e5656af6e495f.

This is one successful request per isolated speech clip, not reliability, paired recognition, piano or musical qualification. Its prompt/clip count changed and the previous study did not pin lease identity, so it does not establish the cause of the earlier failure. All previous failed grades/stops are preserved. The owner then explicitly authorized the distinct five-call single-piano screen. It passed missing audio, descending/eight attacks and ascending/eight attacks, but failed m04: the repeated tone requires two attacks at steady pitch; Pro reported three attacks and unknown direction. The output was complete, HTTP200/STOP and schema-valid, so this is a semantic recognition failure. The study stopped permanently after four POSTs / three single audio parts; held-tone m05 was not sent. No retry/fallback/judge, within-session opaque account/project equality held, all raw provider text equals helper output, and generation/time stayed within2,048/90seconds. Cumulative follow-up28 against ceiling29; unused allocation cannot resume the stopped study. Gateway11881/port53045 is closed. Thirty-three offline harness checks passed; source/runner/artifact/prior-ledger and63corpus pins match. Evidence root .superpowers/sdd/2026-10-04-anti-recognition-pro-single-piano/, frozen protocol ada43306f329aafa65ad91815a252a409d3f8a2b994cdd6d6a6537c08b2ae8bb. Speech and scale successes remain bounded observations; precise attack counting, reliable rhythm, held-tone perception and musical judgment remain unqualified. The next investigation uses local PCM/onset/transcription estimates with source checks, without repeating live screens or changing defaults.

The preceding Pro no-audio study remains a separate failed bare-JSON gate: it truthfully reported missing audio within one complete JSON fence, so its five audio controls were not sent. Its raw output, failed grade and stop are preserved without reinterpretation. Native JSON MIME/schema fields were transmitted in both studies, but native formatting enforcement remains unverified; the existing Keyspilli decoder validates output locally.

Earlier actual-Player and low/high Flash screens remain stopped after unsupported missing-audio claims. The existing owner listening observations and frozen corpus are preserved. This PR fixes transport, bounded refusal, identity and budget evidence; musical judgment, arrangement diagnosis, reliable pitch/count/order perception and automatic publication remain unqualified. Keep this PR draft pending its separately scoped recognition/acceptance gates. No merge, production deploy, global gateway/default change or corpus regeneration is included.

The subsequent offline audit reproduced the existing Keyspilli PCM inspector on both frozen repeated/held WAVs. Energy-onset estimates were0.080/0.600seconds versus authored starts0.000/0.500, and0.120 versus0.000 for held; all are within/at the frozen120ms tolerance. No analyzer defect was reproduced. The known local Basic Pitch executable is absent, so acoustic pitch remains unknown. A joined hybrid-observation.json now keeps pinned authored events, measured PCM estimates/configuration and the unchanged raw Anti claim separate. Its local reproducer verifies11PCM fields per clip, media/source hashes and provider provenance with zero provider calls. This is offline disagreement evidence, without profile/default changes, tuning, automatic repairs or musical acceptance. See .superpowers/sdd/2026-10-04-anti-recognition-pro-single-piano/hybrid-followup.md for the next implementation.

Local consumer evidence report — 5 October 2026: Keyspilli draft PR216 now includes an optional offline report at commit b75067169a1959c31199f277c4fec3ba4f96f47d. It keeps pinned authored events, measured PCM and saved single-clip Anti claims distinct, rejects contradictory incomplete/fallback markers, and makes zero provider requests. The actual repeated/held report shows local repeated2starts versus unchanged savedAnti3, and held advisory unavailable. Acoustic pitch and musical acceptance remain unestablished; all prior failures/stops and63corpusfiles are preserved, follow-up count stays28.

Keyspilli full local2546tests/278files, focused67audio tests, workspace typechecks and production build pass. Exact-head CI37241815916 passes including full E2E and isolated backup/restore; deployment/publication skipped. Skill0.1.0+codex.20261005003602 has26-file canonical/install parity and both validators pass. The previous cache was restored after installer eviction and all26files match pre-install hashes. Anti source is unchanged; its11current-head checks were freshly confirmed successful. No new live screen, retry, fallback, qualification, merge or deployment.