[thatsrue]

What changed

Three fixes from the review in #77:

• Every input goes through convert_to_wav, so a 16 kHz mono WAV is peak-normalised like any other container. The copy path and needs_conversion are gone.
• Opening a progress file cuts an unfinished last line, so the next run's chunks start on their own line instead of being glued onto the fragment.
• The inference record names only the transcription JSON. A deleted VTT now fails only the vtt stage and is rewritten from the JSON, with no inference.

After updating: the cache version is not bumped, so finished sessions are not re-transcribed. Records written before this keep their old contents: a 16 kHz mono WAV that was already converted keeps its unnormalised copy until force=1 or a source change, and an older inference record still lists the VTT, so deleting that VTT re-infers once more.

Why

The same track gave VAD and whisper different audio depending only on its container. A second kill mid-run lost every chunk the previous run had checkpointed. Deleting a VTT redid hours of inference that the code comment promised to skip.

Closes #77

How you checked it

• make lint, make typecheck, make test-fast: exit 0, 297 passed.

• The three new tests fail on main and pass here: test_quiet_16k_mono_wav_is_peak_normalised, test_missing_vtt_is_rewritten_from_json_without_inference, and the resume step added to test_chunk_checkpoint_roundtrip_and_truncation.

• test_quiet_16k_mono_wav_is_peak_normalised replaces test_cache_miss_replaces_existing_normalized_audio, which tested the removed copy path. test_cache_miss_removes_existing_audio_before_conversion still covers stale derived audio.

• Slow tests: 6 passed, 1 skipped. test_segmented_transcription and test_basic_transcription could not run because there is no whisper model in this checkout.

• A bug fix comes with a test that fails without it

• Docs match how it works now

• No secrets or private information