v0.3.0

Added

• Improved sentence-level export with a dependency-free sentence boundary engine that avoids splitting abbreviations (Dr., Mr., e.g., etc.), decimals and version numbers (3.14, v2.4.1), domains and technical names (example.com, Node.js, Next.js), and initials (J. R. R. Tolkien, U.S.). Closing quotes and brackets after terminal punctuation stay inside the sentence.
• Sentence JSONL records now include cue_coverage, timing_precision, padded playback_start/playback_end playback ranges, and cue provenance fields (start_cue_index, end_cue_index, cue_count, start_char_in_first_cue, end_char_in_last_cue, boundary_engine).
• Cue-internal sentence boundaries are now estimated with weighted token interpolation (word length and punctuation pause weights), with the previous character-ratio approach kept as a fallback.
• Strong cue gaps (over 0.6 s) can split unpunctuated text when the next cue starts with an uppercase letter; well punctuated continuing sentences are never split because of a gap.

Changed

• The legacy timing_strategy sentence field is now derived from the new cue_coverage and timing_precision fields with unchanged values (cue_exact, cue_merge, heuristic, unknown).