On-disk Whisper transcription cache

[roelvangils]

Summary

• Repeat runs of `dpub convert --transcribe` against the same audio + model + language combination now skip Whisper entirely.
• One JSON file per (audio, model, language) tuple in `~/.cache/dpub/transcripts/` keyed by SHA-256 of the inputs.
• Modifying any input invalidates the entry naturally; failures are non-fatal (corrupt cache → re-transcribe).
• `DPUB_NO_TRANSCRIPT_CACHE=1` bypasses both reads and writes for debugging.

Measured speedup (cavia book, 4h 22m audio, 109 sections)

Run	Wall time
Cold (Whisper on every file)	722 s
Warm (109/109 cache hits)	21 s
Speedup	34×

EPUBCheck stays clean (0/0/0).

Implementation

• `Segment` and `Word` in `dpub-whisper` gain `serde::Deserialize` alongside the existing `Serialize` (round-trip prerequisite).
• New `transcript_cache` module in `dpub-convert` (~280 lines, 8 unit tests). `CachedTranscriber` wraps `Transcriber`, hashes the model once at construction, hashes audio per call, stores a JSON envelope with diagnostic metadata + segment payload.
• `inject_transcripts` swaps in `CachedTranscriber`; the existing in-memory `HashMap` cache stays to avoid re-hashing audio across sections that share a file.

Test plan

• `cargo test --workspace` — all green, 8 new unit tests
• Cold run on cavia book — 722 s, populates 109 cache files (2.6 MB)
• Warm run — 21 s, 109/109 cache hits, EPUBCheck clean
• (Manual) Modify a single byte of one audio file, re-run: expect cache miss for that file only

🤖 Generated with Claude Code