Repository data root. Three categories live here:

• sources.json (committed): canonical source manifest. One entry per artifact (ECMA-376 PDFs, ECMA Part 4 XSD zip, future MS-OI29500, etc.) with url, edition, sha256, and a license note. bun run sources:sync upserts these rows into the reference_sources table. Edit by hand; the sync script reads it.

• xsd-cache/ (gitignored): local XSD download cache. Populated by bun run xsd:fetch. Contents are not load-bearing for the schema graph itself - the graph lives in Postgres - they're just the source artifacts the ingest reads. Safe to delete; regenerated on the next fetch.

• behavior-notes/ (committed when populated): curated YAML files documenting how Microsoft Office actually behaves vs. the spec. A future ingest will load these into the behavior_notes table so structural tool responses can carry "what Word actually does" alongside the schema-level answer. Empty until that workflow lands.

What does NOT live here:

• Generated build output: dist/, dev/data/extracted/, dev/data/chunks/, dev/data/embedded/ (all under dev/, gitignored).
• Database state: lives in Postgres; reproducible from the manifest + ingest scripts.