feat(experimental): add deterministic GPU graph shortest paths

[ibgreen-openai]

Goals

• Publish deterministic, GPU-resident unweighted shortest-path distances and predecessors for interactive graph neighborhood exploration.
• Preserve the optional Arrow-free luGraph package boundary and stack narrowly on GPU vertex-degree/adjacency foundations feat(experimental): compute GPU-resident vertex degrees #2940 and feat(experimental): construct GPU-resident graph adjacency #2937.
• Support dynamic selection and bounded traversal without CPU synchronization, graph-owned frontier scratch, or hidden command submission.

Changes

• Add LuGraphBreadthFirstSearch, LuGraphBreadthFirstSearchProps, and directional traversal types exclusively to @luma.gl/experimental/lugraph.
• Compute bounded BFS distances and deterministic lowest-stable-vertex-ID predecessors directly from caller-owned CSR adjacency using integer atomicMin and explicit per-depth GPU passes.
• Preserve arbitrarily chunked multi-seed vectors, invalid/duplicate seed handling, directed outgoing/incoming/bidirectional traversal, undirected deduplication, and a 1,024-hop compile-time ceiling.
• Support optional GPU-resident active seed counts and active depth controls that change between encodings without recompilation.
• Publish an optional source-aligned zero/one reachability mask; roots and unreachable predecessors use the explicit 0xffffffff sentinel.
• Fail closed when any required adjacency overflow flag is set: all distances/predecessors remain unreachable and masks remain clear.
• Avoid frontier/transient allocations by expanding only vertices whose current atomic distance equals the active hop.
• Use bounded three-dimensional dispatch for every seed and vertex pass while remaining within the portable eight-storage-binding limit.
• Reject physically aliased outputs and preserve all caller-owned source chunks/buffers without concatenation, readback, submission, or destruction.
• Add 63 focused Node contracts and 25 real-WebGPU CPU-oracle scenarios covering deterministic diamond ties, dynamic chunked controls, reverse/both directions, overflow matrices, byte offsets, repeated encoding, and actual [2,2,2] dispatch.

Verification

• nvm use: passed (Node v22.22.1).
• yarn install: attempted normally and with escalation; configured registry denies unavailable dependencies with HTTP 403. Verification uses an isolated APFS-cloned compatible dependency tree and ignored Yarn state with no package or lockfile changes.
• yarn lint fix: passed (1,540 files).
• Final yarn build: passed across every workspace after final code and formatting changes.
• Focused Node BFS tests: 63 passed.
• Focused real Chromium/WebGPU BFS tests: 25 passed.
• Complete CI=1 yarn test: 707 Node tests and 1,607 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; LuGraphBreadthFirstSearch remains absent from the experimental root.
• Normal commit hooks: lint and all 707 Node tests passed.

Risks and follow-up

• Each compiled hop scans all vertices; the existing 1,024-hop ceiling prevents unbounded graph recording, while no CPU-side convergence polling occurs.
• Truncated adjacency deliberately publishes no partial traversal; the topology overflow flag remains GPU-resident and caller-visible.
• Deterministic predecessors choose the numerically smallest stable parent vertex ID among equal-length paths.
• Stacked on feat(experimental): compute GPU-resident vertex degrees #2940; target codex/lugraph-degree-metrics until prerequisites land.