Merge Whisper segments into prose-shaped paragraphs

[roelvangils]

Summary

• Post-process Whisper output before injection: merge ~10–30 s segments into prose-shaped paragraphs of ~3–6 sentences each, instead of one <p> per segment.
• Pure single-pass greedy state machine. Sentence-terminator detection guards against decimal numbers (3.14) and Dutch+English abbreviations (bv., enz., o.a., Dr., blz., …). Capitalises the first letter at finalize. Force-flushes at max-sentences/max-chars to bound the worst case (Whisper hallucinations with no punctuation).
• Each cleaned paragraph carries a stable id="tx-<section>-<para>" so a future per-paragraph Media Overlay sync milestone (M6.5 in the design notes) can reference it without a re-render.
• Default-on. Pass --no-text-cleanup to keep the raw per-segment output for debugging.
• Zero new dependencies — ASCII-only checks, std::char::to_uppercase for the capitalisation fix.

Why

The current --transcribe output is technically correct but unreadable: paragraph breaks every twenty seconds, often mid-sentence, often starting with a lowercase letter. Apple Books (and any reader that doesn't play Media Overlays) shows the resulting EPUB as text-only — and that text is currently a wall of fragmented paragraphs. This change makes it read like prose.

The merged paragraph's [start, end] interval is the union of its constituents (first.start_seconds, last.end_seconds). Whisper segments are non-overlapping and chronological, so this is the correct span for any future per-paragraph audio sync. Today's SMIL Media Overlays anchor on <h1> and pagebreak IDs, not on the injected <p>s, so this change is a no-op for current overlay behaviour. The paragraph IDs are infrastructure for the M6.5 follow-up.

Algorithm in one paragraph

Append segments to a running paragraph builder. After each append, check whether the buffer ends at a real sentence terminator (. ! ? …, with the digit-decimal and abbreviation guards). If yes AND the paragraph has ≥ 3 sentences AND ≥ 300 chars: emit and start a new paragraph. If the paragraph reaches ≥ 6 sentences OR ≥ 600 chars: emit anyway (safety valve). At end of input, flush whatever's still open.

Test plan

• 12 unit tests cover: basic 3-sentence merge, min-chars hold, max-sentences split, max-chars safety valve, tail flush, decimal-number false positive, Dutch abbreviation false positive (enz.), Dr. abbreviation, first-letter capitalisation, empty-segment skip, timing preservation, empty input.
• Existing synthetic-fixture tests stay green (3 tests, EPUBCheck-clean assertion intact).
• Existing real-book tests stay green when run with DPUB_TEST_BOOK (3 tests, including EPUBCheck-clean on the reference book).
• cargo clippy --all-targets -- -D warnings clean on both touched crates.
• cargo test --workspace all green.
• Manual end-to-end: re-run dpub convert --transcribe nl --whisper-model ggml-medium.bin --audio opus --bitrate 32 against the reference book once the in-flight comparison run finishes; eyeball one section's XHTML to confirm the cleanup reads as prose. (Will land as a comment on this PR.)

Out of scope (filed separately)

• dpub-whisper: transcribe() reloads the GGML model on every call (~3–5 min wasted per book) #10: per-file Whisper model reload (perf bug; independent fix).
• Cover lookup: --cover <path> + --auto-cover via Open Library #11: cover-image lookup (--cover and --auto-cover).
• M6.5: per-paragraph Media Overlay anchor wiring — the tx-NNN-MMM IDs land in this PR but the SMIL doesn't reference them yet.