performance(mincut): replace O(m) edge scans in LocalKCut::check_cut with indexed lookup

Confirmed defect

LocalKCut::check_cut currently handles each boundary adjacency by calling self.graph.edges() and linearly searching the returned full-edge Vec by edge_id:

self.graph.edges().iter().find(|e| e.id == edge_id)

This is reachable on the default branch in crates/ruvector-mincut/src/localkcut/mod.rs. DynamicGraph already maintains edge_index and exposes get_edge(u, v), an endpoint-indexed lookup.

Evidence

PR #941 isolated the effect while testing min-cut-guided HNSW deletion repair:

• individual find_cut calls: 24.8–158.0 seconds at HNSW construction density m0=32
• completed three-deletion run: 318,024.78 ms for LocalCutGuided versus 0.78 ms for EagerRepair
• the same order of magnitude appeared in three independent attempts

The broader HNSW repair hypothesis remains rejected because small-world topology also violates the local algorithm's bounded-degree assumptions. This issue concerns only the independent full-edge scan defect.

Minimal candidate

Use the existing endpoint lookup for each boundary adjacency, preserving edge weight and cut-edge semantics:

if let Some(edge) = self.graph.get_edge(v, neighbor) {
    cut_edges.push((v, neighbor));
    cut_value += edge.weight;
}

Acceptance criteria

• Existing ruvector-mincut correctness tests remain green.
• Add a regression that verifies identical cut membership and weight before/after the lookup change on weighted undirected graphs.
• Add a fixed seeded dense-graph benchmark comparing the current scan implementation with the indexed implementation.
• Indexed lookup is at least 100× faster on the recorded dense fixture without changing the returned LocalCutResult.
• cargo test -p ruvector-mincut, cargo clippy -p ruvector-mincut --all-targets -- -D warnings, and the relevant workspace CI shard pass.
• Do not revive or claim acceptance of the rejected min-cut-guided HNSW repair design without a separate preregistered experiment.

Start in crates/ruvector-mincut/src/localkcut/mod.rs, focusing on LocalKCut::check_cut and DynamicGraph's existing get_edge lookup. Run cargo test -p ruvector-mincut and inspect the correctness tests before adding the weighted undirected regression and fixed-seed dense-graph benchmark. Done means unchanged LocalCutResult values, at least a 100× benchmark improvement, and passing the specified clippy and workspace CI checks.