Make Leiden faster than Louvain

[forhappy]

Summary

• follow up merged PR Optimize Leiden community detection performance #315 with the remaining Leiden hot-path optimizations
• reuse dense neighbor scratch space, cache per-level graph statistics, and roll back rejected passes with a move journal
• borrow the first graph level, skip redundant full-graph subgraphs, and defer member materialization until convergence
• reduce typed-topology and evidence allocation while preserving canonical ordering and exact floating-point accumulation
• preserve byte-identical partition and quality output with reference-equivalence regression coverage

Performance

Fresh release-process measurements used 5 warmups and 51 alternating observations per algorithm on FastAPI pinned at 0c2b6aafd7a2e3a5bf1055ea0ed0a41da15ba5f4.

Graph profile	Louvain median	Leiden median	Leiden / Louvain
max inference, 4,152 nodes / 10,038 links	205.449 ms	202.579 ms	0.986x
low inference, 2,335 nodes / 3,370 links	78.761 ms	74.355 ms	0.944x

Leiden is 1.4% faster on the max-inference graph and 5.6% faster on the low-inference graph. In the release in-process compact qualification, fixed Leiden is 0.611x Louvain latency (39% faster). Median peak RSS on the max graph is 136.13 MiB versus 134.19 MiB for Louvain, a 1.44% overhead.

The fixed-Leiden max-graph output SHA-256 remains exactly 4f26ed16c66b5f939e2436255dfff884defa431392b26a3e322c331f957a4e75; community count and quality metrics are unchanged.

Verification

• cargo fmt --all -- --check
• cargo test -p compass-graph --locked
• cargo clippy -p compass-graph --all-targets --all-features --locked -- -D warnings
• cargo clippy --workspace --lib --bins --locked -- -D warnings
• cargo test --workspace --lib --bins --locked
• ./scripts/qualify_code_graph_v1.sh --fixtures-only
• ./scripts/qualify_code_graph_v1.sh --community-quality --report /Volumes/Workspace/crabbuild-target/compass-c112/community-quality-optimized.json
• sh scripts/check_product_boundary.sh
• git diff --check

The generated community-quality report is byte-identical to the checked-in qualification artifact.