docs(adr-001): BENCHMARK.md — phase-2 SubLinear numbers + class catalog

[ruvnet]

Summary

Closes the last "fully SOTA" gap from the ADR-001 status preamble. BENCHMARK.md was stuck at the Linear-only baseline (NeumannSolver / OptimizedConjugateGradientSolver) and didn't reflect the change-driven SubLinear paths shipped in PRs #26–#35.

Changes

1. Change-driven solve paths — Linear vs SubLinear. Numbers from PR bench(adr-001 #2/#6): empirical SubLinear delta-solve comparison #31's `delta_solve` criterion group: cold_full / warm_full / sparse_closure across n = 64, 256, 1024. Demonstrates the architectural property that `sparse_closure` stays roughly constant while the Linear paths grow with n.

2. Event-driven anomaly example numbers (PRs docs(adr-001): event-driven anomaly example using SubLinear primitives #33, docs(adr-001): wire coherence-gated skip into event-driven example #35): baseline 88 µs Linear, gate `<1 µs` `O(|δ|)`, per-event ~2-3 ms SubLinear.

3. Complexity-class catalog enumerating every phase-2 primitive with its declared class, plus the MCP enforcement / surfacing surface.

ADR-001 SOTA status

Item	Status
All 6 roadmap items SubLinear-or-stronger end-to-end	✅
Phase-2 MCP budget enforcement on solve handlers	✅
Wire-queryable catalog via `estimateComplexityClass`	✅
Empirical SubLinear bench receipt	✅
README complexity-class language	✅
BENCHMARK.md complexity-class language	✅ (this PR)
Pi hwmon J/solve backend	hardware-dependent — needs Pi runner

Test plan

• Markdown renders correctly in GitHub preview
• Full CI (docs-only, no code touched)

🤖 Generated with claude-flow