fix(whisper): decode each window without the previous window's text

[cemkocabasa]

Problem

A 5:50 dictation on local Whisper (whisper-turbo-q5, auto-detect, no hints) was transcribed cleanly up to 3:12. From there it returned one short sentence 81 times and nothing else until the final two words. About 2.5 minutes of speech were lost. The result was identical on every run: the live dictation, zee -transcribe, and four direct Ctx.Transcribe runs.

Cause

no_context = true only clears context between whisper_full calls (whisper.cpp:6921). Inside one call, every 30 s window is still conditioned on the text decoded so far (:7111, up to 224 tokens). So the design-notes line that ruled out the cross-chunk prompt was based on a misreading of that flag.

A -DWHISPER_DEBUG trace shows what happens. Every window from 207 s to 349 s decoded the same sentence with entropy = 2.46849, against an entropy_thold of 2.40, so the temperature fallback never ran. The sentence's 8 text tokens alone score 2.07. The timestamp tokens around each repeat change value every time, and they lift the window just over the threshold. Each next window then starts from a prompt full of the repeated sentence and continues it with confidence (avg_logprobs between −0.25 and −0.13).

Change

internal/whisper/whisper.go: n_max_text_ctx = 0, or the hint's token count + 1 when hints are set. The pinned hint still fills every window's prompt and no decoded text gets in. Clips under 30 s are a single window, so they decode exactly as before.

Measurements

M4, 16 GB, macOS 27. Same clip, only n_max_text_ctx changed:

n_max_text_ctx	Result	Time
default (224 effective)	loop from 3:12, tail lost	29.2 s
64	complete	20.8 s
32	complete	16.0 s
16	complete	16.5 s
0 (this PR)	complete	16.5 s

Trade-offs, measured:

• Punctuation drifts in later windows. On the same dictation there are 5.6 sentence marks per 100 words at 0, against 10.5 at 16 tokens.
• Turkish TTS (78 s, say -v Yelda): WER against the source text went from 6.2% to 7.5%. That is one word: "konşimento" → "con shimento".
• English TTS (116 s, say -v Samantha): WER 3.3% both ways, with no word changes.
• Hints still reach every window. With a four-term hints file, a company name spoken in the last window (5:40) comes out right; without hints it is misspelt.

Rejected alternatives (a small rolling context, a higher entropy_thold, splitting at pauses) and the reasoning are written up in docs/design-notes.md, under "Why whisper decodes each window without the previous window's text". The old timestamps entry is marked partly superseded, following the repo convention.

If you'd rather keep more punctuation, a 16–32 token rolling context also cleared this clip. I didn't choose it because a window that does loop would still pass its last repeats on to the next window.

Tests

• go test -race ./... passes, with the whisper model present, so TestTranscribeKnownClips ran.
• No regression fixture: the failing recording is a private dictation, and synthetic long clips don't loop.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G7rq8KjcC7Y2vQ79CCoTo4