perf: analyze(minimum_spanning_tree) grows superlinearly (469 s at 100k nodes, 0.5.2)

[DecisionNerd]

Found while recording XYG's GraphForge composition scale evidence (CurateLabs/xyg#930). In @curatelabs/graphforge 0.5.2, analyze("minimum_spanning_tree", ...) grows much faster than linear on a sparse graph, while the other algorithms on the same graph stay near-linear.

nodes (edges)	pagerank	louvain	minimum_spanning_tree	dijkstra
1,000 (2,000)	5.6 ms	7.8 ms	97 ms	3.7 ms
10,000 (20,000)	46 ms	94 ms	4,140 ms	27 ms
100,000 (200,000)	783 ms	2,195 ms	468,618 ms	744 ms

Each 10× step costs about 43× then 113×, which looks closer to O(V·E) or O(V²) than Kruskal/Prim's O(E log V). #582 (closed) polished MST scale earlier, so this may be a regression or a path #582 didn't cover (weighted, directed=false).

Setup

• Graph: two weighted edges per node, a seeded small-world pattern (i → i+1 and i → i+k), loaded through bulk construction contract v1 (publishBulkNodes/publishBulkEdges, UUIDv7 ids).
• Inputs: CurateLabs/xyg benchmarks/gen_graphforge_scale_inputs.py --out DIR --sizes 1000,10000,100000.
• Call: g.analyze("minimum_spanning_tree", "Person", null, false, "w").
• Machine: linux-x64, Node 22.22.1, AMD Ryzen 7 3800X (16 threads), timed in-process.

Repro

const { GraphForge } = require("@curatelabs/graphforge");
const g = new GraphForge();
g.publishBulkNodes(uuidv7(), fs.readFileSync("nodes-100000.arrow"));
g.publishBulkEdges(uuidv7(), fs.readFileSync("edges-100000.arrow"));
console.time("mst");
g.analyze("minimum_spanning_tree", "Person", null, false, "w");
console.timeEnd("mst");

Raw numbers: spec/benchmarks/graphforge-compose-local.json in CurateLabs/xyg#930 (graphforge.timings_ms).