feat(core): import stores first and embeds later (deferred embedding, controlled repair --reembed) #1608

Problem

uteke import embeds every row while it imports (import_export.rs). With a local ONNX model that is CPU-heavy: a large file keeps the machine busy for the whole import, and a backend that is down or rate-limited stops the import at that row (the rows already imported stay; importing again skips them through the duplicate check). remember already has the opposite policy: it stores the row keyword-only and uteke repair --reembed adds the vector later.

Uteke Cloud does the same split: a write returns before the vector exists and an outbox worker embeds it afterwards.

Proposal (owner direction, 2026-10-30: store first, embed later; next update, not 0.22.0)

• import stores the rows without a vector (keyword-searchable at once) and reports how many still need one.
• uteke repair --reembed does the embedding in a controlled way: --limit N per run, a pause between batches, resumable (it only picks rows with no vector), progress output, and a nice-able CPU use. The same for document chunks (done for chunks in feat: import restores the vectors of imported documents; import --full (#1587) #1592).
• Optional: an embed worker inside uteke-serve that drains the backlog in the background at a low priority, like the cloud outbox worker.

Open questions

• Default or flag? A flag (import --defer-embed) keeps today's behaviour for everyone, including the LongMemEval harness, which imports with an embedder and must not change (benchmark invariant). Making it the default is a behaviour change for CORE POST /import.
• Duplicate check without vectors. The cosine duplicate gate (>= 0.95) needs a vector. Without one, dedup falls back to an exact content hash (source_hash exists), and near-duplicates are only caught when the vector is added later. Decide whether that is acceptable.
• Contradiction and edge wiring at import also read the vector; they would run when the row is embedded.
• Backlog visibility: uteke stats / verify should show "N memories without a vector".
• Server worker vs CLI only for the first step.
• CORE impact: the response of POST /import and the meaning of "imported" (uteke-cloud contract issue first).

Related

#1606 (a store with no embedding backend imports keyword-only), #1592 (chunk vectors are re-embedded by repair --reembed), #1607 (CHANGELOG wording about embedding failures during import).