[ajianaz]

What

• New embedding.rs: Uteke::embed (lazy init + retry), embed_once (no retry) and embedder_max_seq_len. The embedder lock is held for one call, never across a loop.
• The ten call sites (recall, explain, document search, document chunks x2, room recall, update, import, embed_text, writes, reembed) use it; the five expect("embedder ensured above") are gone and nothing outside embedding.rs locks the embedder.
• Behaviour change: query, document and room embedding now retry a transient failure like writes (3 attempts, 200/400/800 ms). reembed takes the lock per row instead of for the whole loop and still makes one attempt per row. Writes keep their policy: init failure is an error, generation failure is FTS5-only with a warning.
• The missing-embedder message is now the plain "No embedder configured".

Why

#1538, part of #1543. Four patterns for the same call meant retry on writes only, panics that were safe only by call order, and a lock held across the reembed loop.

Testing

• New tests with the shared fake (test(core): let tests supply an embedder through the public open path (#1542) #1545): query, document search and embed_text retry once; a persistent failure errors after 3 attempts; embed_once does not retry; a write that cannot embed is FTS5-only with a warning; reembed makes one attempt per row.
• cargo test --workspace (core lib 877 passed, 0 ignored), clippy -D warnings, fmt.
• Cora MCP review of the new module: no findings.

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT