ruvector-mincut: LocalKCut::check_cut does an O(m) edge scan per call

Found during the 2026-09-05 PR-review mission while reviewing #941 (agent pr-review-ruview-ruvector; the PR itself is being closed as its own verdict was REJECT and it collides with main's ADR-340).

Observed: crates/ruvector-mincut/src/localkcut/mod.rs:368 — LocalKCut::check_cut scans every edge (O(m)) on each call, so the local k-cut check is not local in cost; repeated checks during repair scale with total graph size rather than the cut's neighbourhood.

Ask: bound the scan to the incident edges of the candidate cut (adjacency-indexed), or document the cost and cap the call count in the repair loop. A benchmark that grows m while holding the cut size fixed would show the current behaviour.

Not established: whether any production path calls check_cut in a loop today; the finding is from reading the function.

🤖 Generated with RuFlo

https://claude.ai/code/session_019xHM4rAH4aaShb4DTr1n6s

Start in crates/ruvector-mincut/src/localkcut/mod.rs at LocalKCut::check_cut around line 368, and trace its callers in the repair loop to establish whether repeated checks occur. Add a benchmark that grows m while keeping the candidate cut size fixed. Done means the scan cost is bounded to incident edges, or the existing cost and repair-call limit are explicitly documented and enforced.