feat(d2): normalized gain + paired-difference significance

[DCCA]

Summary

Two statistical upgrades to the effectiveness (D2) dimension, the two highest value-per-effort items from the deep-research competitor survey (SkillsBench, and Anthropic's "Adding Error Bars to Evals", arXiv 2411.00640):

1. Hake normalized gain — g = (skill − baseline) / (1 − baseline), reported alongside absolute lift in the scorecard D2 line and JSON. Corrects for baseline headroom: a +5pp lift on a 0.9 baseline (g=0.5) reads as the strong result it is, not a rounding error next to a +5pp lift on a 0.4 baseline (g≈0.08).
2. Paired-difference significance — significance now runs on the per-eval paired difference when ≥2 evals appear in both configs. Shared per-eval difficulty cancels, shrinking the SE so a genuine lift clears the noise with fewer trials. Falls back to the unpaired se_of_difference for a single eval (no between-eval variance to exploit).

baseline_lift and its unpaired SE are unchanged, so existing consumers keep working; the new normalized_gain / paired_lift / paired_se fields are additive.

Design notes

• New stats.normalized_gain and stats.paired_diff are pure, dependency-free, and unit-tested; aggregate/dimensions/scorecard thread the fields through.
• Significance is still decided at full precision (before rounding) so display rounding can't flip the verdict.

Test evidence

169 passed in 0.53s          # +6 new tests (stats, aggregate, scorecard)
50 files already formatted    # ruff format --check
All checks passed!            # ruff check
self-validation: 100 / A / Ship

Dogfooded end-to-end on a two-eval benchmark — D2 line renders:
pass 68% ± 25%, lift +20% (significant), gain +0.38, and the paired path flips a lift that the unpaired SE would call n.s. to significant.

🤖 Generated with Claude Code