feat: chunk long SRT cues within each beat's fixed voice window

[htlin222]

Closes #4.

Problem

build_srt() emitted exactly one SRT cue per voiced beat, spanning the beat's entire voice.start -> voice.end window with the whole beat.text as a single block. Fine for short beats, but intro/outro beats can carry a full paragraph — worst case in production: one cue, 40.0s, 216 chars shown as a single static block. Bad subtitle UX.

Fix

build_srt() now calls a new chunk_cues(text, start, end, max_units=34, min_dur=1.2) per beat instead of emitting the beat verbatim, and renumbers cues across the whole file (not per-beat). The audio-reconciled [voice.start, voice.end] window itself is never touched — only the caption layer inside it is re-sliced, so this has zero interaction with detect_anchors.py/splice.py/mux()'s amix/adelay timing.

Algorithm (matches the issue's sketch, with the two open pieces filled in):

• cjk_width(text): rendered subtitle width via unicodedata.east_asian_width — Wide/Fullwidth chars (CJK ideographs, kana, hangul, fullwidth punctuation) count as 2 units, everything else as 1.
• split_on_punctuation(text, max_units): split on sentence punctuation (。！？) first; any sentence still over max_units is split further on clause punctuation (，、,); a clause still over budget with no punctuation at all (e.g. a long English run) falls back to a hard width-based wrap. Concatenating the returned chunks always reproduces text exactly.
• chunk_cues: redistributes [start, end] across the chunks proportionally by cjk_width; the last sub-cue's end is pinned to end to absorb rounding drift, so total duration is preserved exactly.

Merge-back rule (the trailing-fragment edge case called out in the issue): punctuation splitting is width-aware but duration-blind, so a short window can produce a chunk whose proportional share of the duration falls below min_dur (an unreadably brief flash). _merge_short_chunks repeatedly folds the first offending chunk into its neighbour (forward, or into its predecessor if it's the last chunk) until every remaining chunk's proportional share clears min_dur, or only one chunk is left. This can leave a merged chunk over max_units — documented trade-off: a floor on reading time trumps the char-count budget, since a legible-but-long cue beats an unreadably brief one.

A beat short enough not to need splitting collapses back to exactly the old one-cue-per-beat behaviour (verified explicitly in tests) — all pre-existing test_overlay.py tests pass unchanged.

Tests added (tests/test_overlay.py)

• test_chunk_cues_short_beat_collapses_to_one_cue_spanning_full_window — short text reduces to one cue, same window.
• test_chunk_cues_long_beat_splits_into_multiple_readable_sub_cues — ~193-char/40s realistic zh-TW narration splits into multiple cues; verifies contiguous windows, exact duration sum, exact text reconstruction via concatenation, and per-cue reading-speed/width bounds.
• test_chunk_cues_merges_short_trailing_fragments_to_respect_min_dur — 8 comma-separated clauses crammed into a 3s window would otherwise yield 8 sub-1.2s cues; verifies the merge-back rule collapses them while still respecting min_dur, exact duration sum, and exact text reconstruction.
• test_chunk_cues_single_leftover_chunk_cannot_grow_past_the_fixed_window — degenerate case where the window is shorter than min_dur itself.
• test_build_srt_renumbers_cues_across_the_whole_file_not_per_beat — numbering is one contiguous sequence across multiple beats' sub-cues.

Test plan

• Baseline before changes: 98 passed, 0 failed.
• After adding tests + implementation: 103 passed locally.
• After git rebase origin/main (picked up unrelated upstream commits fix: escape quotes in VHS tape Type lines #10-docs: document ledger.json downstream reuse patterns #13): 111 passed.
• All pre-existing overlay tests (test_srt_cues_come_straight_from_ledger_voice_windows, test_launch_subtitle_is_clean_no_slot_prefix, test_srt_orders_by_voice_start) still pass unchanged.

https://claude.ai/code/session_012EwptawkV4dcGhoGefzxgV