v0.2.0 — Robustness & speed fixes

Hardening pass from a real 13-clip edit that surfaced several silent failures, plus a transcription speedup. 32 tests pass.

Correctness

• Ordering: natural-sort directory inputs so 1,2,…,10,13 merge in numeric order without zero-padding (was lexical 1,10,11,2,…).
• Concat: always re-encode via the concat filter with per-input normalization (scale/pad, fps, SAR, yuv420p, 48k, mono). The old stream-copy demuxer could exit 0 while producing frozen video / drifting audio / clicks on mixed time bases, resolutions, and mono-vs-stereo inputs.
• A/V sync: lock each clip's audio to its frame-aligned video length (apad + -shortest) and force true 30fps CFR, so per-clip gaps no longer accumulate into audible drift.

Subtitles

• Split over-long Whisper cues at punctuation with proportional timing (--max-chars, default 20) so burned-in subtitles no longer wall the frame.

Speed

• Default model large-v3-turbo (~2× faster on CPU, negligible quality loss); --model large-v3 for maximum accuracy.

Docs

• READMEs (EN + 中文): natural ordering, new model, --max-chars, and a "speeding up a multi-clip edit" note (use the mtime cache; never pairwise-merge).