End-to-end loop on synthetic-toy: prefixes + predict + score

[protosphinx]

Summary

• Wires up prefixes, predict --baseline markov, and score so the README's command sequence actually runs end-to-end on synthetic-toy.
• Adds the first-order Markov reference baseline — the floor any submission must clear. On synthetic-toy: top-1 0.976, top-3 1.000.
• Locks the file formats the leaderboard CI will depend on (prefixes.csv, predictions.csv) via tests/test_e2e.py.

What's new

• pm_bench/prefixes.py — extract (case_id, prefix_idx, prefix, true_next) targets from a split; skips length-1 cases. CSV round-trip.
• pm_bench/predictions.py — predictions CSV (case_id, prefix_idx, predictions).
• pm_bench/baselines/markov.py — fit on train cases only; unigram fallback for unseen last-activity. No torch / sklearn — just CPython.
• CLI: pm-bench prefixes <name> --split split.json --out prefixes.csv, pm-bench predict <name> --split split.json --prefixes prefixes.csv --out predictions.csv --baseline markov, pm-bench score predictions.csv --prefixes prefixes.csv --task next-event.
• README install/use section now reflects the loop that actually works; STATUS.md added; GOALS.md ticked.

Why this matters

• v0.1's dataset fetch is gated on a 4TU TOS step that can't be automated. Locking the file format on synthetic-toy first means external contributors can build models against pm-bench now, on synthetic data, and slot the same code into BPI/Sepsis/Helpdesk the moment v0.1 lands.
• Sets up the second reference baseline cleanly: adding gnn is a --baseline gnn choice once the dataset machinery is in place.

Test plan

• pytest -q — 24 passed (was 17)
• ruff check pm_bench tests — clean
• Manual end-to-end smoke: split → prefixes → predict → score returns top1 0.976 / top3 1.000
• tests/test_e2e.py runs the same sequence via click runner — guards the file format

Roadmap impact

• Bumps v0.0.1 checkbox in README + GOALS; v0.1 (dataset fetch) is the next gate.

[protosphinx]

auto-deferred for human review: LOC delta exceeds gate (652 added + 15 removed = 667 lines, threshold is 250)

---

Generated by Claude Code