P2 — Measurement v2: run harness, benches, doc==tool, grade bands #112

Phase P2 of docs/V1-ULTRAPLAN.adoc. The scoreboard's credibility. Closes G6.

WS-2.1 — Run harness (the missing corpus)

• Versioned runner: 14-probe suite + scenario sets × N repeats, seeds where supported, timestamps → vexometer-trace-v1 bundles
• Aggregation through ISA_Calculator + Compare_Models (Welch + bootstrap CI already in-tree post-fix: scoring correctness for probe regexes, RCI, comparison, and the README table #109)
• ≥3 local models (Ollama) + ≥2 API models (explicit consent) each producing a re-derivable public run bundle
• validate-style recomputation accepts every published bundle

WS-2.2 — Benchmarks ("proven tests and benches")

• Scoring throughput bench, harness wall-time bench, JSON report-generation bench (join existing Ada + Criterion suites)
• Regression gate: warn on >10% regressions vs bench history
• Numbers land in the state file with producing commands

WS-2.3 — Doc == tool for every printed number

• Generalise the fix: scoring correctness for probe regexes, RCI, comparison, and the README table #109 table test: one module recomputes every figure the site renders (ISA, grades, deltas) from raw aggregates
• Editing a rendered number without recomputation fails just test

WS-2.4 — Grade bands + rank rules (D-2 ruled: hybrid)

• Fixed bands canonical: A ≤0.15, B ≤0.30, C ≤0.45, D ≤0.60, F >0.60 (0–1 lower-is-better) — confirm cut points against WS-2.1 corpus before freezing
• Percentile letter line (rank-relative per epoch, labelled relative)
• docs/GRADES.adoc (or METRICS.adoc section): both systems pure functions, exhaustive boundary tests

Acceptance: every leaderboard figure re-derives in tests; grade maths deterministic at boundaries.