TTS: Hear It does not play actual rendered sound

[genwave-radio]

I think it renders and plays the first one you hit Hear it for, but every time after that, it plays the same one. I tried a sample one, hit Get IPA and then Hear it, hit Play, played as expected. Didn't like it so I made a small change, hit Get IPA, Hear it and Play, same exact one played. Thinking I had messed something up, made a more major change, same effect. So I changed the word completely (ie. Pattern: MacCleod and Respelling:This-Is-A-Test) and the sound played was still the original, so something is wrong in this workflow somewhere.

[genwave-radio]

🎯 Root cause: a legacy speech correction is shadowing the rule — nothing is stale

The audio is not a replayed first render. Every audition rendered fresh — and every render was byte-equivalent by construction, because a legacy speech correction on the demo box (From: MacLeod → To: Maa-cloud, the old pre-IPA respelling mechanism) rewrites the text before pronunciation rules ever match:

1. Audition phrase built from the pattern: Now playing: MacLeod.
2. NormalizingTtsSynthesizer (the F68 chokepoint) applies the correction → Now playing: Maa-cloud.
3. The candidate rule {pattern: MacLeod, ipa: …} now matches nothing — the word it targets no longer exists in the text
4. Kokoro renders the same rule-less text every time → identical audio, whatever the IPA

The first audition "played as expected" because the correction was doing the respelling — the old rule made it say Maa-cloud, not your new one.

🧾 Evidence

• api log, every one of the 7 auditions (10:29–10:31): TTS correction fired: from=MacLeod voice=am_michael
• kokoro log, all 7 renders: identical 23-char input Now playing: Maa-cloud. — no [word](/ipa/) annotation — identical 61200-sample output
• derive was healthy the whole time: 7 requests, distinct respellings (lengths 9,9,8,8,7,8,11) → distinct IPA
• dev repro: same request renders 96,680 B without the correction vs 84,150 B (= the no-rule render, ±2 B kokoro noise) with it. Correction present → candidate IPA has zero effect
• browser flow verified end-to-end on dev (Playwright, real clicks): each Hear It POSTs the current IPA, fresh blob URL, different audio each round. UI, derive, preview endpoint, blob handling all correct

⚠️ Scope: not audition-only

On-air renders normalize through the same chokepoint, so any word with both a correction and a pronunciation rule has a dead rule on air too. The audition is faithfully reproducing air — that parity is working as designed; it's the collision that's silent. (T280's wire proof happened to use "MacCleod", double-C, which the correction doesn't match — why the loop tested green.)

🩹 Immediate unblock

Settings → TTS → speech corrections: delete the MacLeod → Maa-cloud row. Your IPA rule then matches and the audition tracks every change.

🔧 Fix directions (needs a ruling)

• (a) Authoring-time collision guard — rules editor/API and corrections write path warn when a correction From overlaps a rule Pattern (case-insensitive). Cheap, fail-loud, no render-path change.
• (b) Precedence ruling — pronunciation rules (the newer, better system) suppress a same-word correction at render time; NormalizingTtsSynthesizer already holds context.Rules, so it can skip a correction whose From matches a resolved rule pattern. Changes on-air behavior → wants a deliberate ruling.
• (c) Both — guard now, precedence at the next /design.