Optimize algorithm hot paths with typed-array CSR kernels

[davidkpiano]

Summary

Performance overhaul of the algorithm hot paths. Every rewrite preserves observable behavior — same results, same ordering, same laziness and error contracts — and all 1,852 existing tests pass unchanged.

• getStronglyConnectedComponents — iterative typed-array Tarjan over the CSR out-arcs (single pass, stack-safe on deep graphs). ~15x faster.
• getTopologicalSort — Kahn's algorithm over CSR arcs with a typed ring queue and per-CSR cached in-degrees, replacing the Map-based version whose queue.shift() was O(n²). ~20x faster.
• isBipartite / getMaximumBipartiteMatching — 2-coloring runs directly over the cached CSR arcs instead of rebuilding an undirected adjacency on every call; the self-loop sweep now only runs on the success path. ~60x faster on repeated queries.
• Bellman-Ford — typed relaxation over cached compact arc arrays (same relaxation order, so tie-predecessor enumeration is unchanged), plus a scalar-predecessor single-pair fast path for getShortestPath that returns the same path the full enumeration would yield first. ~15x faster.
• Floyd-Warshall — flat Float64Array distance matrix with copy-on-write tie-predecessor pair lists, replacing nested object arrays with per-cell clones and some() dedup scans. ~4x faster.
• genBFS / genDFS / genPostorder — hand-rolled Generator-protocol iterators (no generator resume machinery), with BFS/DFS batching work in 1024-node chunks served from a buffer. Yield order is unchanged (property-checked against the previous implementation across random/scale-free graphs and all directions), validation still throws on first next(), and mid-iteration mutation snapshots behave as before — now backed by the CSR's node snapshot instead of a per-iterator array copy. ~1.5–2x faster full traversals.
• Dijkstra / A* / bidirectional search — default edge weights come from a per-arc Float64Array cached on the CSR, removing edge-object loads (and ?? 1 fallbacks) from the inner loops.
• getDegree — served from a per-version degree map (one hashed lookup per call); a one-entry memo in front of the index WeakMap speeds all point-query bursts against the same graph.
• All-targets shortest-path reconstruction — each path materializes once via a shared backtracking buffer, replacing per-recursion-level array spreads (O(L) instead of O(L²) per path). Also roughly halves the getAllPairsShortestPaths runtime on path-like graphs.

Notes for review

• New caches (arc weights, edge-order arcs, in-degrees, degrees, nodes snapshot) all key off the existing GraphIndex/CSR version, so they inherit the documented staleness contract — mutation-API changes invalidate them, in-place field edits still require invalidateIndex().
• Dangling-edge handling is normalized in a couple of spots (e.g. Floyd-Warshall previously crashed on edges with missing endpoints; they are now skipped like everywhere else).
• The bipartite conflict edge named in getMaximumBipartiteMatching's error may differ from before when several edges prove non-bipartiteness (any witness is valid; tests only assert the message shape).

Testing

• pnpm verify (typecheck, repo typecheck, generated schemas, conventions, full test suite, build, publint + package smoke test) passes.
• Added-behavior equivalence was additionally property-checked for traversal ordering against the previous implementations.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QXF7Ba1P2JvvtCfiYbX9xe

---

Generated by Claude Code

---