v0.3: outcome task — AUC scoring + prior baseline

[protosphinx]

Stacked on top of #6 (remaining-time). Merge order: #2 → #3 → #4 → #5 → #6 → this.

Summary

• Adds outcome as the third task — binary classification, pure-Python AUC. CLI is fully task-aware: 3-of-5 v0 tasks now ship.
• Last-activity-conditioned prior baseline: the dumbest model that uses any prefix information. Tying it means the model isn't reading the trace.
• No leaderboard entry on synthetic-toy yet — seed=42's test partition has n_pos=0, AUC degenerates. The full pipeline still runs and is tested; a real entry lands when a BPI dataset gets pinned.

What's new

• score_outcome — rank-sum AUC with average-rank tie breaking. Degenerate single-class case returns 0.5 by convention rather than NaN, so leaderboard rows stay readable.
• OutcomeTarget + extract_outcome_targets — repeats the case's final 0/1 outcome at every prefix length. Models see the same target with progressively more context — a clean way to measure "how soon can you tell?"
• prior_outcome baseline — last-activity → empirical positive rate, with global-rate fallback. ~30 lines of CPython.
• _synth.is_positive_outcome — synthetic-toy rule: case ends with delivery_confirmed.
• CLI dispatch — --task outcome, --baseline prior. UsageError on mismatched (task, baseline) pairs. Outcome rule resolved by dataset name.
• 8 new tests (test_outcome.py) — extraction, baseline determinism, per-last-activity rates, CSV round-trips, e2e click-runner pipeline. 73 total.

Smoke (synthetic-toy)

$ pm-bench prefixes synthetic-toy --split split.json --out o.csv --task outcome
$ pm-bench predict synthetic-toy --split split.json --prefixes o.csv \
    --out opreds.csv --baseline prior --task outcome
$ pm-bench score opreds.csv --prefixes o.csv --task outcome
{ "task": "outcome", "auc": 0.5, "n": 41, "n_pos": 0 }

The n_pos: 0 is honest about what the data supports; the rest of the pipeline is verified by tests/test_outcome.py against a hand-built event set with controlled class balance.

Test plan

• pytest -q — 73 passed (was 59 on PR v0.2: remaining-time task + mean reference baseline #6)
• ruff check pm_bench tests — clean
• AUC math: perfect separation, perfect inversion, all-tied, single-class, hand-checked 8/9 case
• e2e CLI on synthetic-toy completes cleanly even with degenerate AUC

Roadmap impact

• v0.3 (5-task scoring): now 3 of 5 (next-event, remaining-time, outcome) ✅; conformance + bottleneck remain
• The outcome leaderboard slot is real — it just waits on a dataset whose test split has both classes

[protosphinx]

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