Sync protocol#

Repos are keyed by their normalized git remote URL (fallback: .scry.toml [project] name, then directory basename), and every path is stored relative to the repo root. Any checkout of the same remote, on any device, addresses the same index rows.

A sync is a manifest diff:

1. Client walks the repo (.gitignore, .scryignore, hidden files, and default binary patterns excluded; empty and oversized files skipped) and hashes each file with xxh64.
2. POST /v1/manifest returns the server's {relpath, xxh64} list; only files whose hash differs are uploaded, in batches, and files missing locally are deleted server-side.
3. The server chunks and embeds each upload. Chunks carry a content hash; a chunk whose hash already has a vector anywhere in the store reuses it, so edits re-embed only what actually changed.
4. Committing a file flips any memory anchored to it with a different hash to stale.

scry watch runs the same diff on debounced filesystem events; a second checkout of an already-indexed repo syncs with zero uploads.