refactor(memory): optimize exact retrieval and batch writes

[Grivn]

Problem and change

Bounded Memory reads still scan/sort corpus-sized candidate sets, local BFS loads the whole store, and exact recall retains token/vector intermediates and repeats cosine work. This change uses indexes that match active ordered reads, indexed local BFS, bounded semantic/diff selection, scalar recall signals, a reusable vector decode buffer, and transaction-local prepared inserts.

Synced with master at 5c98ff9c (including #148). Preserve its covering edge indexes, caller-loaded snapshots, intent-ranked transition cache and rank-before-visit-budget rule. Cached transitions consume the same precomputed cosine scores as vector anchors and reranking. Scores, thresholds, per-anchor budgets, supersession demotion, causal ordering and commit boundaries retain their current-master contracts; no approximate search is introduced. Independent traversal parity is checked against the upstream implementation.

Performance against current master

Same benchmark sources, N=1/10/100/1,000/10,000, five alternating runs per version: 1,080 final samples. Apple M4, Go 1.25.4, one P, real temporary SQLite WAL, local 128-dimensional vectors. Values below are median process CPU time; raw wall time and allocations are also committed.

Case	Current master	Final PR
Hybrid recall, N=100	3.038 ms	2.555 ms
Hybrid recall, N=10,000	93.062 ms	66.531 ms
Latest by source, N=10,000	0.602 ms	0.0146 ms
Bounded two-hop neighborhood, N=10,000	40.306 ms	0.0595 ms
Atomic 10,000-row insert	524.331 ms	431.604 ms
Dense recall, N=100	18.324 ms	16.866 ms

Hybrid recall allocation at N=10,000 drops from 78.85 MB to 49.10 MB per call (37.7%). An isolated comparison with the merge-only version attributes about 10.24 MB of that reduction to decode-buffer reuse.

The earlier 6.51× dense-recall gain belongs to the historical baseline; current master already includes most of that graph-cache improvement. Small cases retain tradeoffs: N=10 BFS is 50.09 → 59.34 µs; single-record semantic paths add roughly 2 µs. Known-entity scanning at N=10,000 is about 4% slower; retention is near neutral. Exact vector scanning remains O(Nd), dense neighborhood ranking and negative substring/entity scans remain explicit costs. Shared-host timing is not a production latency promise.

Validation

• go build -o mnemon .
• make test (including vet and architecture gates)
• go test -race ./internal/memory/... ./cmd/memory -count=1
• MNEMON_EMBED_ENDPOINT=http://127.0.0.1:1 bash scripts/e2e_test.sh: 258 passed
• npm test --prefix npm/cli: 13 passed
• Local OpenCode runtime and ZCode shell suites passed; the Windows PowerShell-specific probe was skipped on macOS.
• Upstream traversal parity across four intents, three budgets, ties, repeated anchors, deleted nodes, zero/negative vectors and dimension mismatch; caller snapshot ordering; decode buffer resizing and special float bits; existing index migration, BFS parity, rollback/panic and compaction tests.

CI is green on 30d61deb: Linux deterministic/npm checks and Windows Memory checks passed. The first Linux attempt failed in the unchanged Agency test TestStoreConcurrentPutHasOneFinalEffectAcrossInstances with promotion did not settle. That case passed 100 local repetitions on both the pure master baseline and this branch; the failed Linux job then passed on rerun. The failure was not reproduced locally, and its root cause remains unconfirmed.

Paid live providers and the Agency/Docker integration umbrella were not run. Benchmarks exclude CLI startup, provider latency and cold filesystem-cache timing.

Reports and complete samples: English analysis, 中文报告. The original measurements remain explicitly labeled as historical evidence.