fix(asr): harden v3 multilingual chunk seams

[vdt4534]

Summary

This updates the existing stacked draft after the #596 overhaul. I retested the current #596 head (bfa14a17) with a larger fixture set. The no-mel direction still looks right and keeps the fast path, but a few longer/more varied fixtures still drift into English at chunk seams.

This patch keeps the same basic direction and speed profile, then adds a v3-only acoustic boundary/warmup path for no-mel batch chunks.

It does not use language hints, vocabulary filtering, language-specific token rules, or a second decode.

What changed

• For Parakeet TDT v3 with melChunkContext == false, choose chunk starts from local acoustic boundaries instead of a fixed grid only.
• Use local energy contrast around the boundary candidate, not hard-coded French/English rules or token vocabulary.
• Decode a hidden 7-frame real-audio prefix only on short troughs, suppress prefix-region output tokens, but still let those tokens update decoder state.
• Skip warmup across stable quiet/true pause regions.
• Avoid pulling a chunk boundary earlier when that would force the next boundary into high-energy speech.
• Harden overlap merging so equal-length/leading contested gaps use token confidence instead of always choosing the older chunk.
• Keep this gated to v3/no-mel; v2, tdt-ja, and zh-CN CTC are not routed through this path.

Why

On the current #596 branch, the original notes_1408_clean.wav repro is fixed, and speed is much better than the old pinned baseline. The expanded fixtures showed remaining drift:

• climate voice memo: gouvernements and entreprises...
• long FR clip: départemental, the camions...
• Perrault FR 2min: the degree of pénétration...
• Candide FR 10min: and Montepulciano... Chypre and Samos

The failures do not look French-specific in mechanism. They look like seam/decoder-context issues where v3's multilingual prior falls into English after a bad boundary. The fix only looks at audio energy, frame indices, and model token confidence.

Validation

20 fixtures locally:

• 8 Amnis/repro fixtures: FR, EN, mixed FR->EN, long FR
• 12 LibriVox public-domain fixtures: EN/FR/ES/PT at roughly 2, 5, and 10 minutes

Timing:

Variant	Total / 20 fixtures	Avg
pinned bb96003	30.748s	1.537s
current #596 bfa14a17 no-mel	17.153s	0.858s
this PR	17.032s	0.852s

Known target snippets with this PR:

• notes_1408_clean: faire un effort sur le progrès
• wwii_belgique_fr: En Belgique... a été jour férié
• user2_2026-05-12: blouses médicales blanches et portant
• climate voice memo: gouvernements et entreprises ont revu
• mixed FR/EN: oblasts occupés de Donetsk... A decade on
• long FR clip: départemental les camions rouges
• Perrault: degré de pénétration
• Candide: et les invite... et à boire

Checks:

• swift test --filter ChunkProcessorTests
• swift test --filter TdtRefactoredComponentsTests
• swift test --filter TdtDecoderV2Tests
• swift build -c release --product fluidaudiocli
• git diff --check

Fixtures

The already-uploaded issue fixtures are enough to reproduce part of this. I also tested additional longer clips. The LibriVox clips are public-domain and can be attached or regenerated as clipped WAVs if useful. The private Amnis/voice fixtures are not committed here; I can attach specific ones only if we decide they are okay to make public.

[Alex-Wengg]

i actually overhauled this change . #596

[vdt4534]

Updated this draft after your #596 overhaul. The previous version of this PR is superseded; the current branch is now based on bfa14a17 and keeps the no-mel direction, with the extra acoustic-boundary/warmup handling described in the updated body.

I also added a fresh note on #594 with the newer fixture results and timing numbers.