Graph analytics catalog: 20 new kernels, and a workspace CI that can pass

[alexy]

Tier A of docs/goals/graph-analytics-catalog.md: the procedure registry grows from 10 projection kernels to 30, and Nutmeg serves every one of them with no code of its own.

Kernels

Family	Procedures
Community	louvain, leiden, labelPropagation
Centrality	betweenness, closeness, harmonic, eigenvector, katz, hits
Structure	kCore, triangleCount, localClusteringCoefficient, bridges, articulationPoints, biconnectedComponents
Similarity	nodeSimilarity
Paths and flow	spanningTree, maxFlow, minCut
Embeddings	fastRP

Each has an independent oracle in its test (brute force, a dense-matrix definition, or exhaustive search) plus hand-computed cases, and cancellation, budget and scratch-release checks. Parallel kernels return bit-identical results and charge identical work at 1, 2, 3 and 8 threads; louvain, leiden and labelPropagation are sequential on purpose, because reproducibility is the contract.

Plumbing

• run_on_projection and public projection_options, so an embedder runs any kernel by name on its own projection.
• NodeTable: one result shape for per-node values, for rows a kernel keys itself (node pairs, edges), and for fixed-length Float32 vectors (Arrow FixedSizeList). A table with no rows yields one empty batch carrying its schema.
• A second positional node argument, target.
• parallel.rs and meter.rs: fixed blocks combined in order. The other agent's WorkMeter on work/rayon-kernels supersedes Meter; see codex-to-codex.md for the proposed merge order.

CI

The workspace workflow has failed on every run for weeks, for stacked reasons each hidden by the one before:

1. grust-ladybug linked zstd twice under --all-features on Linux. It is now tested on its own.
2. The benchmark runner's tests read benchmarks/lsqb/upstream/lsqb, which is gitignored. fetch-upstream.sh now downloads the pinned archive and verifies its size and SHA-256.
3. One timing test in the runner failed once and passed once on the same code. It is not fixed; its assertion now prints the error it saw.

Superseded pull-request runs now cancel themselves.

Not in this PR

articleRank (an option inside pagerank.rs), and bellmanFord/astar, which need projections that admit negative weights; the design question is recorded in the plan. Nutmeg's run-on-projection branch must merge together with this.

🤖 Generated with Claude Code