fix(engine): don't hold pooled DB connections across embedder/LLM calls (consolidation + mental model)

[nicoloboschi]

Problem

Several memory-engine paths held a pooled PostgreSQL connection checked out for the entire duration of a slow external call — an embedder or LLM round-trip. The pools are already bounded (asyncpg, default max_size=20) and per-process, so this is not a leak — it's saturation: enough concurrent operations park the pool on multi-second external calls and the poller / remaining work then block until acquire_timeout.

This is a reimplementation of the idea behind #2434 (which was opened before #2917 rewrote these paths around the pluggable memories store) onto current main.

What this PR fixes

update_mental_model — the embedding is now computed before acquiring a pooled connection. Its text depends only on the incoming name/content, never on DB state.

Consolidation — _process_memory_batch and its executors / dedup helpers no longer receive a long-lived connection. Recall, the batch LLM call, every per-action embed, and dedup adjudication run with no connection held; each helper self-acquires a short-lived connection only around its own SQL.

Moving the slow calls off the connection widens the decision→write window, so the held-transaction serialization is replaced with explicit guards:

• Each source-liveness check (FOR SHARE) is paired with its write in one short transaction, so a concurrent delete can't orphan an observation between check and write.
• Dedup CREATE/UPDATE folds are RETURNING-gated and re-filter live sources inside the fold transaction (sources-before-observation lock order, matching the normal write paths). A twin deleted during the now connection-free window → the caller still CREATEs; a source deleted → it's dropped from the fold, never written back.
• A cheap non-locking preflight restores the pre-refactor "skip before embedding when every source is already gone" short-circuit.

The separate-store (non-SQL) branches and the Oracle-safe search_vector clause (#3021) are preserved throughout.

Scope / deliberately deferred

update_memory_unit (and update_mental_model's sibling edit path) still embed inside the acquired connection, but their fix requires a two-phase read/embed-off-connection + short write transaction with re-lock/abort/retry that must be reconciled with the pluggable store's begin_txn/decide_txn cross-store transaction coordinator — and validated against a real database. That's a follow-up PR (linked once opened) rather than bundling an unvalidated concurrency rewrite of the interactive-edit path here.

Verification

• ruff + ty clean.
• 45 deterministic no-DB tests pass — they pin the dedup fold guards (RETURNING-gating, live-source filtering, twin-vanished → still-CREATE, created/skipped propagation) and the pre-embed short-circuits/rejections directly.
• Live-DB consolidation / invalidation / document-transfer tests updated to the new short-acquire signatures, plus a new in-flight-delete race test (source deleted during the off-connection embed → the in-txn FOR SHARE guard skips the update). These run against live PostgreSQL in CI.

Refs #2434