feat(core): cross-document entity resolution for batch extraction (#431)

[CheeryProgrammer]

Closes #431.

Problem

Per-chunk extraction resolves entities by exact canonical name (resolveOrProposeNode), so the same real entity written under variant names across a batch's documents — "OrderService" in one doc, "Order Service" in another — lands as separate proposals. A batch gives us the whole document set at once, so we can do better.

What

After a batch's items are all applied, PollExtractionBatches runs resolveBatchDuplicates over the identity scopes it touched:

• union-find proposed nodes that share a normalized name or alias — the same normExpr the review-time merge proposer uses (lowercase, strip non-alphanumeric, so "Order Service" ≡ "order-service" ≡ "OrderService");
• merge each cluster into the oldest survivor: fold aliases, repoint proposed edges (RepointProposedEdges), retire the duplicate to historical.

Conservative by construction

• Exact normalized-string identity only — never fuzzy/embedding similarity (that stays a review-time proposal).
• Only proposed nodes are touched — nothing live changes, and a human still reviews the merged proposal.
• Scope-keyed — same-named entities in different projects never merge (Scoped node identity: name + declared discriminator set (default project; empty = global) #117/Consolidate scoped to discriminator set (no cross-project merges) #118).
• Runs only on the batch path; the synchronous path already dedups current nodes via Consolidate.

Tests (DB-gated, run in CI)

• Two docs in one batch mention the same entity under variant names → collapse to one proposal (asserted order-independently, since a batch applies items in map order), alias folded, edge repointed.
• Distinct entities (OrderService / PaymentService / Kafka) are never merged.

Follow-up

• Embedding/semantic cross-document clustering (near-synonyms, typos the normalizer misses) needs proposal-summary embeddings that fresh proposals don't yet carry + threshold tuning — filed as Embedding-based cross-document entity clustering (follow-up to #431) #437.

🤖 Generated with Claude Code