docs: the S3 wet-0.30 audible rows examined — bounded lines at the plate's own modes, loop-dependent

[tap]

Docs only. docs/reverb-afc.md said three times that the reverb-only rows at S3, wet 0.30, are audible below the dry loop's limit in some seeds, "not examined". This PR replaces that with what an investigation on the same host, seeds and material found.

What the criterion flags there is a sustained narrowband line at one of the plate's own strongest low modes — 93.8 Hz or 832 Hz in 7 of the 10 rows, both within 3 dB of the plate's impulse-response peak below 2 kHz (164.7 Hz) — 26–31 dB over the envelope for 0.5–0.8 s, first crossing at a forward gain 0.1–20.4 dB below the dry limit.

Controls:

• Feedback path open (gain held at −300 dB for 120 s), plate on the held note and on speech, decay 0.5 and 0.7, all 5 seeds: zero flagged tracks in 20 of 20 runs. The loop is necessary.
• Canceller alone at S3, same seeds: +10.75 / +12.37 (sane). Same plate at S1: +10.14 on both the chain output and c.
• Calibration coverage and ramp alignment were complete in every run; the ramp's onset gains back-calculate from the schedule. Not a tooling offset.
• It does not run away: the bisected cost at these settings is ≤ −0.10 dB and the ramp's runaway is the highest in the table; the lines stay bounded 20–30 dB over the envelope. In 3 of 10 rows the chain output and c disagree by more than 5 dB on different lines — the cross-check doing its job.

What the doc now says: real, feedback-dependent, bounded, at the plate's own modes; the bisected cost and the audible limit answer different questions in this cell; the mechanism that lets a loop 10–20 dB under its limit sustain a 30 dB line at a plate mode is not identified (the canceller is in the loop; the effect is absent without the plate and absent at S1); read the per-seed range, not the median; whether it is audible on real singing is a phase 1 question.

No source, test or driver change. No number in the new text comes from anywhere but those runs.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DDhgJqxUsdmWPKh6pKrV5A