Fix stale T2 aggregation documentation in paper_metrics_core.py

[NarayanSchuetz]

Problem

The compute_skill_scores docstring in src/imputation_evaluation/evaluation/paper_metrics_core.py (currently around lines 419–437) describes an obsolete aggregation and scope layout:

• overall as a flat geomean over all 19 per-channel tasks across scenarios
• collapsed scopes named cat_collapsed:sleep / cat_collapsed:workouts
• an overall_binary_collapsed scope

That no longer matches the executable implementation. Since the C2/C3 consolidation, T2 overall is the universal three-level B.2 aggregation:

1. Within each (scenario, bucket), mean log(clip(R_task)) over constituent tasks.
2. Within each scenario, mean over buckets present.
3. Across scenarios, mean the scenario-level log ratios.
4. Return 1 - exp(mean_log_ratio).

Bucket sourcing is:

• activity: continuous ch_0..ch_4
• physiology: continuous ch_5..ch_6
• sleep: cat_collapsed:sleep
• workouts: cat_collapsed:workouts

Per-channel binary rows ch_7..ch_18 do not contribute directly to per-scenario, semantic, or overall scopes. They enter those scopes only through the collapsed binary rows. The emitted headline scope is overall; overall_binary_collapsed no longer exists.

Reimplementing from the stale docstring produces a materially different, channel-count-weighted result (especially overweighting workouts).

The rank aggregation docstring near aggregate_task_ranks_to_scopes is stale in the same way, and several nearby comments still mention overall_binary_collapsed despite no executable code emitting that scope.

Current source of truth

• Three-level skill aggregation: paper_metrics_core.py around lines 512–530 and 590–597
• Bucket mapping: b2_bucket_for_channel and _bucket_log_R_per_scenario
• Canonical documentation: src/imputation_evaluation/METRICS.md, section 5.3.1
• Historical rename/removal: commit e863802 (overall_binary_collapsed → overall; legacy flat overall deleted)

Acceptance criteria

• Rewrite the compute_skill_scores scope/aggregation docstring to describe the universal three-level B.2 form.
• Document that ch_7..ch_18 are excluded from headline aggregation and enter via cat_collapsed:*.
• Remove references implying that overall_binary_collapsed is an emitted scope.
• Update the aggregate_task_ranks_to_scopes docstring and nearby stale comments consistently.
• Ensure scope names match current output: cat:activity, cat:physiology, cat:sleep, cat:workouts, semantic, and overall.
• Add or update a regression test that asserts no overall_binary_collapsed scope is emitted and that overall follows bucket × scenario balancing.