dream(performance): HNSWIndex efSearch query-time default (evaluated, REJECTED)

[ruvnet]

1. Hypothesis

Given AgentDB's HNSWIndex-backed memory search, when the query-time default candidate-pool size is decoupled from efConstruction via a dedicated efSearch default (50), then per-query search latency should decrease relative to baseline, subject to: (1) recall@10 must not drop below 0.90 at any measured corpus size, (2) build/insert-time behavior is unchanged, (3) existing tests remain green.

2. Candidate

v3/@claude-flow/memory/src/hnsw-index.ts + types.ts — 2 code lines changed, 1 interface field added:

• HNSWConfig.efSearch: number (new field, default 50)
• search()'s implicit-default line: Math.max(k, efConstruction) → Math.max(k, efSearch)

Previously, every caller of HNSWIndex.search() that omitted ef (which is every production caller in agentdb-adapter.ts — no in-repo caller currently passes ef) implicitly used the build-time graph-quality knob (efConstruction, 200) as the query-time latency/recall knob. This decouples them, matching standard HNSW practice.

3. Evaluation Receipt

New seeded benchmark: v3/@claude-flow/memory/benchmarks/results/scripts/efsearch-default-benchmark.mjs. Clustered synthetic corpus (256-dim, N=3000/8000, 60 queries/N), live brute-force recall@10 ground truth (never stored), paired per-query timing.

N	Latency Δ	Recall@10 Δ	Floor (0.90)
3000	-55.9% (t=-30.5, n=60)	-3.17pp (0.995→0.9633)	held
8000	-57.5% (t=-32.7, n=60)	-7.5pp (0.9517→0.8767)	BREACHED

4. Baseline Comparison

Baseline = git stash the candidate diff, rebuild, rerun the identical script; candidate = restore diff, rebuild, rerun. Same script, same seeds, two code states — see receipt-baseline.json / receipt-candidate.json / comparison-efsearch-final.json in this PR.

5. Darwin Lineage

Skipped — candidate failed its own pre-declared quality invariant (STEP 12 precondition: "candidate passed basic evaluation" not met). Running bounded generations over a rejected baseline would tune toward passing a floor that already caught a real problem.

6. Flywheel Evidence

No signed @metaharness/flywheel bundle — bespoke deterministic benchmark, not an LLM-task corpus the replay/verify tooling targets. Evidence retained as committed receipts + the issue/gist report.

7. Reward Hack Check

Manual checklist clear: gold computed live (not stored), no threshold tuned to pass (0.90 floor set before the final run and the candidate fails it), both N sizes reported including the failing one, seeds identical across baseline/candidate labels, zero API cost.

8. Security Review

Not security-sensitive: in-process vector-index query-default change only. No credential/network/filesystem/MCP-authority surface touched.

9. Regression Analysis

Existing test suite 456/456 minus 1 pre-existing environmental failure (root bypasses a read-only-file-permission test, confirmed identical with/without the candidate via stash/rebuild) → 455/456 both before and after. Zero regressions attributable to the candidate.

10. ADR

None — simple parameter-default change, and REJECTED; doesn't meet this pipeline's bar for an architectural decision record.

11. Research Gist

docs/dream-cycle/dream-gist-2026-08-15.md (this session had no gist-creation tool available, consistent with every prior dream-cycle night).

12. Issue

#3033

13. Witness

Field	Value
Session commit	45e65b5dae5d2c312e70cd5ba90df0701ea05c28
Report SHA-256 (pre-witness content)	2f2de7910ca69da3752d18c67964927871cf8a8b8ce1c470d1d798170f5e81fd
Witness stamp	d756e6d92234275d44976fe47668e8713fa359481abd3d9d85e983ec94a4dd2d

14. Merge Policy

Human review required. Do not self-merge. Do not autonomously promote Flywheel state. This PR is REJECTED evidence, retained for reviewability and follow-up, not intended to be merged as-is — see Recommendation §15 in issue #3033 for the proposed follow-up direction (higher/scale-adaptive efSearch default, fix HNSWIndex.deserialize()'s persistence gap, re-test at production-realistic 1536-dim).

---

Generated by Claude Code

[ruvnet]

The "🤝 Agent Coordination Tests (performance-benchmarker, 1)" failure on f2f5a20cd is a transient infra flake, not a code issue: Unable to download artifact(s): Failed to GetSignedArtifactURL: Unable to make request: ECONNRESET while pulling a setup artifact between CI jobs — unrelated to this PR's diff (v3/@claude-flow/memory/src/{hnsw-index,types}.ts + the new benchmark scripts under v3/@claude-flow/memory/benchmarks/results/).

That commit is already superseded by ae2b1b59b (ledger update), which triggered a fresh set of CI runs currently in progress. No action taken here; will follow up if the new run fails for a real reason.

---

Generated by Claude Code