v0.3 closed: conformance task — DFG fitness × precision → F

[protosphinx]

Stacked on top of #10 (CSV ingest). Merge order: #2 → #3 → #4 → #5 → #6 → #7 → #8 → #9 → #10 → this.

Summary

• Adds the fifth and final v0 task: conformance checking via DFG fitness × precision. v0.3 closes — every task in the README's table now ships with a CPython baseline and a leaderboard entry.
• DFG-based, no pm4py dep. The structure is right-shaped for a future alignment-based variant (PR v0.3 closed: conformance task — DFG fitness × precision → F #11+) behind a [discovery] extra.

What's new

• score_conformance — F = 2fp / (f+p) where f and p are computed from set overlap of submitted DFG vs test-partition DFG. Catches both the "model too small" (loses fitness) and "model too big" (loses precision) failure modes.
• pm_bench/conformance.py — DFG extraction, model JSON r/w. Submission format is a JSON file with a transitions list of [a, b] pairs.
• New CLI verb pm-bench discover — discover <name> --baseline dfg --out model.json discovers a DFG from training cases. Score path takes --dataset + --split (not --prefixes) since the model is global, not per-prefix.
• leaderboard/conformance/synthetic-toy.json with the dfg-ref entry: F=0.857 (fitness 1.0, precision 0.75 — the model covers all test transitions but carries 2 extras the test never uses).
• leaderboard.py, CLI standings printer, and STANDINGS.md markdown all learn the conformance column set. pm-bench leaderboard --all --verify now walks 4 boards.
• 11 new tests (test_conformance.py) — extract_dfg, every score corner case (perfect/tiny/big/disjoint), JSON round-trip, board verification, e2e CLI. 108 total.

Smoke

$ pm-bench discover synthetic-toy --split split.json --out model.json --baseline dfg
wrote model with 8 transitions to model.json (baseline=dfg)

$ pm-bench score model.json --task conformance --dataset synthetic-toy --split split.json
{
  "task": "conformance",
  "fitness": 1.0,
  "precision": 0.75,
  "fscore": 0.857...,
  "n_test_transitions": 6,
  "n_model_transitions": 8
}

$ pm-bench leaderboard --all --verify
bottleneck/synthetic-toy: OK — 1 entry(ies)
conformance/synthetic-toy: OK — 1 entry(ies)
next-event/synthetic-toy: OK — 1 entry(ies)
remaining-time/synthetic-toy: OK — 1 entry(ies)

Test plan

• pytest -q — 108 passed (was 97 on PR CSV ingest — pm-bench works on any event-log CSV without registry plumbing #10)
• ruff check pm_bench tests — clean
• Score corner cases (perfect, too-small, too-big, disjoint, empty)
• e2e CLI — discover → score conformance round-trip on synthetic-toy
• STANDINGS.md staleness canary caught itself when conformance column added; regenerated and re-passes

Roadmap impact

• v0.3 ✅ closed in README and GOALS.
• The stack now spans v0.0 through v0.4 plus the CSV ingest user-facing win. Open milestones beyond this PR series are: pinning the first real BPI dataset (one-time TOS step) and gnn baseline integration — neither blocked on more code from us.

[protosphinx]

Merged into main as part of the audit-cleanup stack (commit 9c00b47). The full content of this PR is now on main.