fix(ruvector-mincut): deterministic partition() + witness complement fix

[ruvnet]

Hypothesis

The 2026-09-05 nightly run (ADR-345, mincut-gated forgetting) measured RuVectorGraphAnalyzer::partition() returning different, sometimes empty, results across repeated calls on byte-identical graphs, and filed two open "Next Research" items: find the root cause, and check whether the crate's lower-level API avoids it. This PR is that follow-up: given the same graph, when the two root causes are fixed, repeated partition() calls should return byte-identical, correct results — subject to zero regressions in the existing test suite and the original benchmark's verdict changing only if the bugs (not the algorithm's fitness for that task) explain its prior REJECT.

Root cause (two independent bugs)

1. BoundedInstance::brute_force_min_cut/search_for_cuts (crates/ruvector-mincut/src/instance/bounded.rs) built their vertex/seed enumeration order from HashSet<VertexId> iteration. Rust's default hasher reseeds on every HashSet/instance construction, so first-match tie-breaking among equal-cost minimum cuts was a function of construction-time entropy, not the graph. Fixed with sort_unstable() on both vectors.
2. WitnessHandle::materialize_partition() (crates/ruvector-mincut/src/instance/witness.rs) computed the cut's second side as 0..=max(membership) minus membership — silently wrong (often empty) whenever the winning side didn't contain the graph's actual highest-ID vertex. Fixed at the one production call site, RuVectorGraphAnalyzer::partition() (crates/ruvector-mincut/src/integration/mod.rs), by computing the complement against the analyzer's real graph vertex set.

crates/ruvector-agent-memory/src/graph_forget.rs's MincutGatedForgetting no longer needs its mincut_trials majority-vote workaround now that the underlying call is deterministic; its default drops 3 → 1 (unit tests: 10 → 1).

Architecture

See the mermaid diagram in the nightly report: docs/research/nightly/2026-09-21-mincut-determinism-fix/README.md.

Files changed

• crates/ruvector-mincut/src/instance/bounded.rs — determinism fix (bug 1)
• crates/ruvector-mincut/src/integration/mod.rs — witness complement fix (bug 2)
• crates/ruvector-agent-memory/src/graph_forget.rs — drop the now-unnecessary retry workaround, doc updates
• docs/adr/ADR-346-mincut-witness-partition-determinism.md (new)
• docs/adr/INDEX.md — index entry for ADR-346
• docs/research/nightly/2026-09-21-mincut-determinism-fix/README.md, gist.md (new)

Benchmark command

# Determinism probe (200 trials on a fixed 19-vertex graph)
cargo run --release -p ruvector-agent-memory --example mincut_determinism_probe --features mincut-forget

# Original 2026-09-05 acceptance benchmark, run unmodified
cargo run --release -p ruvector-agent-memory --example mincut_gated_forgetting_bench --features mincut-forget

Real benchmark results

Three stages, each re-running both commands unmodified:

Stage	Determinism probe (200 trials)	84-vertex bench verdict
Baseline (unmodified main)	88/200 (44%) empty/wrong	REJECT
Bug 1 fixed only	200/200 (100%) empty — now consistently wrong	REJECT
Both bugs fixed	0/200 (0%) empty, 200/200 (100%) correct	REJECT (unchanged, expected)

The middle row is the key intermediate result: fixing only the tie-breaking bug made the outcome fully deterministic and fully wrong (it reliably converges on the tie-broken cut that triggers the complement bug), which is strong independent confirmation both bugs are real and both needed fixing.

Latency is flat (~1.0–1.06s/call) across all three stages — both are correctness bugs, not performance bugs; the crate's separately-documented latency cost is untouched and remains open (see Next Research in the report).

cargo test --release -p ruvector-mincut --lib: 512 passed, 0 failed, 5 ignored (unchanged count from baseline). cargo test --release -p ruvector-agent-memory --features mincut-forget --lib: 31 passed, 0 failed. cargo clippy --release -- -D warnings clean on both crates. 30/30 repeated runs of the mincut_trials = 1 bridge-detection tests, no flakes.

Acceptance result

ACCEPT for the determinism/correctness hypothesis (0% empty, 100% correct, reproducible across repeated runs). The original mincut-gated-forgetting production hypothesis (ADR-345) remains REJECT, unchanged and expected — its rejection was independently attributed to the global min-cut not isolating human-perceived "bridge" memories on this benchmark's synthetic data, orthogonal to both bugs fixed here.

Darwin result

Not run. This is a targeted, root-caused bug fix with one clear correct implementation (sort before first-match; compute the complement against the real vertex set) — no meaningful parameter space to evolve. See the nightly report's "Evolution Results" section for the full reasoning.

Flywheel result

Recorded in docs/research/nightly/2026-09-21-mincut-determinism-fix/README.md: root cause, fix, before/after evidence, and three explicit future-research items (a WitnessHandle constructor-level fix, a latency investigation via DynamicMinCut, and a broader audit of fragment//expander//jtree/ for the same bug class) are filed so a future session doesn't have to rediscover them.

Security review

No new attack surface, no new dependencies, no new unsafe code, no parsing of untrusted input. Determinism is a strict improvement for any downstream witness-verification or replay consumer.

Main limitations

• Only the call path reachable from RuVectorGraphAnalyzer::partition()'s one current production caller was fixed and verified; WitnessHandle::materialize_partition() itself is still wrong for any future direct caller until a deeper constructor-level fix lands (filed as Open Question 1 in ADR-346).
• Other HashMap/HashSet-heavy modules in the crate (fragment/, expander/, jtree/) were not audited for the same bug class.
• The crate's separately-identified per-call latency is unaffected by this fix.

Production recommendation

Land as-is. Both fixes are minimal, change no public API signature, are covered by the existing test suite plus this PR's three-stage before/after benchmark, and remove a correctness confound from any current or future consumer of RuVectorGraphAnalyzer/BoundedInstance/WitnessHandle.

Research document / ADR / gist

• Research report: docs/research/nightly/2026-09-21-mincut-determinism-fix/README.md
• ADR: docs/adr/ADR-346-mincut-witness-partition-determinism.md
• Gist: docs/research/nightly/2026-09-21-mincut-determinism-fix/gist.md

---

🤖 Generated with claude-flow

https://claude.ai/code/session_01MzSGbyatr2eZwuM4JVeP5g

---

Generated by Claude Code