v0.3: bottleneck task — NDCG@10 + mean-wait baseline + leaderboard

[protosphinx]

Stacked on top of #7 (outcome). Merge order: #2 → #3 → #4 → #5 → #6 → #7 → this.

Summary

• Adds bottleneck detection as the fourth task — per-transition wait-time ranking, NDCG@10. CLI is fully task-aware; the leaderboard scaffold now hosts 3 entries on synthetic-toy (next-event, remaining-time, bottleneck) plus the outcome pipeline.
• Mean-wait baseline scores NDCG@10 0.9786 on synthetic-toy. Strong floor — ranking by training-mean already nails the ordering.

What's new

• score_bottleneck — pure-CPython NDCG@k. Predicted scores rank transitions; truth is the held-out per-(a,b) mean wait time. Missing predictions sink to the bottom (refusing to predict doesn't earn credit).
• pm_bench/bottleneck.py — BottleneckTarget, BottleneckPrediction, extract_bottleneck_targets, CSV r/w. Per-transition shape (activity_a, activity_b, mean_wait_seconds, n_observations) — different from the per-prefix tasks, intentionally so.
• pm_bench/baselines/mean_wait.py — train-mean-per-transition with observation-weighted global fallback. ~30 lines.
• CLI dispatch — --task bottleneck, --baseline mean-wait. UsageError on mismatched (task, baseline) pairs, consistent with the other tasks.
• leaderboard.py — _rescore_bottleneck and a new dispatch branch in rescore; standings now sorts by ndcg_at_k for bottleneck and auc for outcome (correctly higher-better for both).
• CLI pm-bench leaderboard prints task-appropriate columns: mae_days for time, auc for outcome, ndcg@10 for bottleneck, top1/top3 for next-event.
• 7 new tests (test_bottleneck.py); 86 total, ruff clean.

Smoke

$ pm-bench prefixes synthetic-toy --split split.json --out bt.csv --task bottleneck
wrote 6 prefixes to bt.csv (task=bottleneck partition=test)

$ pm-bench predict synthetic-toy --split split.json --prefixes bt.csv \
    --out bpreds.csv --baseline mean-wait --task bottleneck
wrote 6 predictions to bpreds.csv (task=bottleneck baseline=mean-wait)

$ pm-bench score bpreds.csv --prefixes bt.csv --task bottleneck
{ "task": "bottleneck", "ndcg_at_k": 0.9786..., "k": 10, "n_transitions": 6 }

$ pm-bench leaderboard --all --verify
bottleneck/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 — 86 passed (was 73 on PR v0.3: outcome task — AUC scoring + prior baseline #7)
• ruff check pm_bench tests — clean
• NDCG math: perfect ranking, inverted ranking, missing-predictions, hand-checked 3-transition example
• Drift canary on the new leaderboard entry runs through the existing --all --verify workflow

Roadmap impact

• v0.3 (5-task scoring): now 4 of 5 ✅ — conformance is the only remaining task. That one will need pm4py (process discovery + token replay), so it's the natural moment to introduce a [bpi] extra.

[protosphinx]

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