Implements pack/unpack_embedding, cosine_similarity, find_best_speaker_match
(max-per-speaker grouping, known/ambiguous/unknown thresholds), and embed_audio
with lazy resemblyzer import so tests run without the docker-only dependency.
10 tests passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
SQLite database layer with thread-local connections, WAL journal mode,
foreign key enforcement, and FTS5 full-text search on utterances via
content-table triggers. TDD: 5 tests written first, all passing.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>