[ivan-digital]

Summary

The packed-LSTM Kokoro models (#518) speak at upstream Kokoro's rate, so one pass of the 5 s graph holds about 56 tokens. Two things still assumed the faster models:

• synthesizeChunks first tried any text up to 128 tokens in one pass and discarded the result on overrun (Render long Kokoro text chunk by chunk #516). Most two-sentence replies now overrun, so their first audio waited for a pass that was thrown away.
• E2ENemotronHarshAudioTests.testStreamingMatchesBatchOnCleanLongUtterance has failed on main since the Core ML re-export (recall 0.923 and 0.949 against a 0.95 bar). CI skips the E2E suites, so it didn't show.

What changed

• Chunk up front. Text goes straight to the chunker, as in speech-core. Text within the 56-token budget is still one chunk and one call. A short last chunk now merges only within 64 tokens (chunkTokenHardCap), the hard cap of speech-core's 5.0 s profile (Pin Kokoro's packed-LSTM graphs and size chunks for their speech rate speech-core#160), instead of within the 128-token input. The plan lives in KokoroTTSModel.chunkPlan, and three tests pin its budget and cap.
• Parity metric. The Nemotron test now scores both transcripts against the paragraph Kokoro spoke. The stream must stay within 0.05 of the batch pass, about two words, and the batch pass must reach 0.85, so a broken fixture still fails. The old metric scored the stream against the batch transcript, so every word the batch pass misheard counted against the stream. On main the two passes only misheard "costs declined. Our" differently ("Ledecland" against "Ledeckland", "Our" against "are").
• testLongTextIsNotClippedToOnePass accepts a peak of exactly 1.0, which the ±1 clamp from Size Kokoro chunks and trimming for the packed-LSTM models #518 can produce.
• docs/models/kokoro-tts.md gives the current budget (56) and cap (64). It also notes that a sentence longer than one chunk is split between words.

Measurements

M5 Pro, re-exported aufklarer/Kokoro-82M-CoreML, generateStream, median of 7:

First audio	main	This branch
Two-sentence reply, 119 tokens	273 ms	131 ms
Its first sentence alone, 47 tokens	135 ms	131 ms

The Nemotron parity paragraph is now 22.8 s of audio. On this branch, batch and stream both recover 0.949 of the spoken text.

Tests

• KokoroTTSTests: 102 tests, 0 failures, 1 skipped (env-gated German parity). This includes the three new plan tests and the E2E tests against the re-exported models.
• Nemotron E2E suites (E2ENemotronHarshAudioTests, E2ENemotronLangTagLeakTests, E2ENemotronStreamingASRTests): 18 tests, 0 failures, 2 skipped (English-only bundle not installed), in two full runs.
• CI

testNoLangTagOnOverlappedSpeech returned an empty transcript in one of five runs. Its 4.05 s two-voice mix is the same on main and here, and its transcript varies from run to run ("Please verify your credentials and confirm the booking time", "Sidensify your credentials"). That flake predates this change and isn't addressed here.

Related: #516, #518, soniqo/speech-core#160.