fix(stt): stop fabricating Qwen3-ASR stream timestamps

[ainergiz]

Summary

Closes #927.

Qwen3-ASR derives partial streaming timestamps from token-budget progress, so changing max_tokens changes the reported timing for identical text. Generated text has no real audio alignment, so this change makes text deltas untimed and keeps only coarse audio chunk extents.

Changes

• Return start_time=None and end_time=None for text-delta events.
• Keep numeric start/end values on empty-text chunk-boundary events.
• Clamp chunk boundaries to the real input duration before model-only padding.
• Buffer untimed text into the matching chunk in the CLI.
• Preserve an untimed final text segment in JSON.
• Raise a clear error instead of silently writing empty SRT/VTT output when no timed boundary exists.
• Document the streaming event contract and point users to the forced aligner for text-level timestamps.

Compatibility

Streaming consumers must handle nullable start_time and end_time on text deltas. The server serializes these as null. Numeric values are coarse audio chunk extents, not transcription timestamps.

Validation

• Focused streaming tests: 12 passed.
• Full STT suite: 283 passed, 9 skipped, 18 subtests passed.
• Black, isort, and git diff --check passed.
• Confirmed identical partial timestamp metadata for max_tokens=64 and 8192.

[github-actions]

⚠️ The pre-commit checks failed in CI.

Run the following from the repository root, then commit and push any changes:

python -m pip install pre-commit
pre-commit run --all-files

The project's pre-commit configuration includes Black and the other required formatters.