feat(experimental): compute GPU weakly connected components

[ibgreen-openai]

Goals

• Publish deterministic GPU-resident weakly connected component identifiers for direct graph coloring and community exploration.
• Preserve the optional Arrow-free luGraph API and stack narrowly on deterministic shortest paths feat(experimental): add deterministic GPU graph shortest paths #2942.
• Keep convergence and overflow state on GPU without implicit submission, CPU synchronization, or graph-owned scratch.

Changes

• Add LuGraphConnectedComponents and LuGraphConnectedComponentsProps exclusively to @luma.gl/experimental/lugraph.
• Compute weak connectivity for directed and undirected graphs using only forward CSR adjacency, symmetric integer-atomic root hooking, and explicitly synchronized pointer-jumping passes.
• Publish each converged component's lowest stable vertex ID; preserve isolated vertices, duplicate edges, self-loops, and valid edges while excluding invalid endpoints.
• Support caller-configured bounded iterations and an optional caller-owned GPU convergence scalar; status one means the final iteration proved a fixed point, while zero conservatively reports incomplete/overflowed work.
• Fail closed on forward adjacency overflow using 0xffffffff labels; reverse adjacency and reverse overflow are deliberately irrelevant to weak connectivity.
• Publish convergence one for valid empty graphs; support omitted convergence storage without allocating graph-owned transients.
• Apply bounded three-dimensional dispatch and portable five-or-fewer storage bindings throughout initialization, hooking, pointer jumping, and globally synchronized convergence resets.
• Validate packed caller-owned uint32 outputs and reject physical aliases against graph/topology allocations.
• Add 45 focused Node contract tests and 19 real-WebGPU CPU-oracle scenarios covering directed weak connectivity without reverse CSR, canonical minimum IDs, conservative convergence, partial iteration, weighted/undirected graphs, overflow, repeated encoding, byte offsets, and actual [2,2,2] dispatch.

Verification

• nvm use: passed (Node v22.22.1).
• yarn install: attempted normally and with escalation; configured registry rejects unavailable dependencies with HTTP 403. Verification uses an isolated APFS-cloned compatible dependency tree and ignored Yarn state; no package/lockfile changes.
• yarn lint fix: passed (1,544 files).
• Final yarn build: passed across every workspace after final formatting and test changes.
• Focused Node connected-components tests: 45 passed.
• Focused real Chromium/WebGPU tests: 19 passed.
• Complete CI=1 yarn test: 752 Node tests and 1,626 real Chromium/WebGPU tests passed; existing skips only.
• yarn website:build: passed.
• (cd website && yarn build): passed.
• yarn examples:typecheck: all 46 example workspaces passed.
• yarn bundle-size: all seven bundle-size budgets passed.
• Built ESM and CommonJS package-subpath smoke checks: passed; LuGraphConnectedComponents remains absent from the experimental root.
• Normal commit hooks: lint and all 752 Node tests passed.

Risks and follow-up

• A convergence status of zero can conservatively include labels that became canonical during the final bounded iteration; status one guarantees a stable fixed point.
• This implements weak—not strong—connected components. Directed edge orientation is intentionally ignored for component membership.
• Atomic operations are integer-only; no floating-point atomics, device-wide shader barriers, CPU readbacks, or automatic fallback are introduced.
• Stacked on feat(experimental): add deterministic GPU graph shortest paths #2942; target codex/lugraph-breadth-first-search until prerequisites land.