research: S-T mincut namespace routing for agent memory vector search

[ruvnet]

Summary

• Adds crates/ruvector-namespace-merge — a zero-dependency Rust proof-of-concept for principled multi-namespace routing in agent memory vector search
• Implements S-T mincut (Edmonds-Karp max-flow) over a namespace similarity graph to identify the coherent semantic cluster of namespaces a query belongs to
• Key algorithmic fix: relative per-query normalisation of cosine similarities makes routing invariant to absolute cosine magnitude across dimensionalities and noise levels
• Adds ADR-298 and nightly research README + gist

Changes

New crate: ruvector-namespace-merge

Three routing strategies sharing a NamespaceRouter trait:

Strategy	Mechanism	Recall@10	Dist ops vs AllSearch
AllSearch	Scan everything	1.000	100%
CentroidFilter	Skip if cosine(q, centroid) < threshold	0.945	38%
MinCutRoute	S-T flow partition on namespace similarity graph	0.985	41%

Flow network (per query, N namespaces):

• S → nsᵢ capacity = round(q_sim_norm[i] × 10000) — query affinity
• nsᵢ → T capacity = round((1 − q_sim_norm[i]) × 10000) — inclusion cost
• nsᵢ ↔ nsⱼ capacity = round(inter_sim[i,j] × 10000) — cohesion penalty

After Edmonds-Karp max-flow, the source-side reachable set in the residual graph = namespaces to search.

Benchmark results (64-dim, noise=0.30, 300 queries)

Variant          Mean(µs)  p95(µs)   QPS      Recall   NS searched  Dist ops
AllSearch           133.7     157     7,481    1.0000      5.00        2500
CentroidFilter       49.7      63    20,125    0.9453      1.91         957
MinCutRoute          54.2      68    18,449    0.9853      2.05        1025

MinCutRoute: 98.5% recall at 41% of AllSearch's distance computations, 2.47× faster.

Tests

6/6 pass:

• all_search_recall_one — AllSearch recall vs itself = 1.0
• centroid_filter_high_recall — recall ≥ 0.75 at threshold=0.20
• mincut_route_searches_fewer_ns_than_all — recall ≥ 0.70, ns < total
• flow_unit_two_cluster_query — direct flow unit test
• flow::tests::test_simple_max_flow — max-flow correctness
• flow::tests::test_source_side — source-side BFS correctness

Docs

• docs/adr/ADR-298-namespace-merge-mincut.md — decision record with context, alternatives, failure modes, security, migration path
• docs/research/nightly/2026-08-08-namespace-merge-mincut/README.md — full research document
• docs/research/nightly/2026-08-08-namespace-merge-mincut/gist.md — public summary

Test plan

• cargo test -p ruvector-namespace-merge — all 6 tests pass
• cargo run --release --bin benchmark — all acceptance criteria pass
• cargo fmt --check -p ruvector-namespace-merge — no formatting issues
• Review ADR-298 for architectural coherence with existing ADRs (especially ADR-026 tiered routing, ADR-297 ACRP)
• Consider wiring MinCutRoute into ruvector-agent-memory query path (Phase 2 per ADR-298)

---

Generated by Claude Code