v0.4.7

What's Changed

• fix(chunker): table-aware boundaries and hard max_chunk_size cap (#57) by @vkehfdl1 in #63
• feat(verify): detect duplicate-vector corruption from pre-#59 Metal embeddings (#61) by @vkehfdl1 in #62
• feat(native): warn on max_length truncation and report chunks_truncated (#60) by @vkehfdl1 in #64
• perf(sync): embed in cross-file windows with content-hash reuse and length-sorted batching (1.3-1.9x) by @HaD0Yun in #66
• chore: release v0.4.7 by @vkehfdl1 in #75

Full Changelog: v0.4.6...v0.4.7

Upgrading notes

• Markdown table chunking changed (#57/#63): the recursive chunker now keeps table rows intact and splits tables at row boundaries. Because chunk IDs are keyed on chunk_schema_id + content hash and the schema id is unchanged, existing indexes will not re-chunk unchanged files. Run minsync sync --full once if you indexed markdown tables with ≤0.4.6 and want the new row-atomic chunking applied to your whole corpus.
• Built with pre-#59 (≤0.4.5) on Metal? minsync verify now detects duplicate-vector corruption from the old Metal embeddings (#62) — run minsync verify to check, and minsync sync --full to repair flagged indexes.