refactor(statistics): consolidate divergent pearsonR copies into one shared helper

[drewstone]

Closes the pearsonR-consolidation item from #77 (the named post-#263 follow-up).

Problem

Seven source files each carried a private Pearson-correlation implementation, with divergent edge-case handling:

file	n < 2	both constant	one constant	length guard
rl/reward-hacking.ts	0	0	0	yes
pipelines/judge-agreement.ts	NaN	1	0	yes
judge-calibration.ts	NaN	1	0	yes
builder-eval/correlation.ts	(none → NaN)	1	0	no
contract/analyze-runs.ts	0 (allows n=1)	0	0	n===0 only
meta-eval/rubric-predictive-validity.ts	NaN	1	0	yes
meta-eval/correlation-study.ts	NaN	1	0	yes

The rank-with-ties and Spearman helpers were duplicated alongside them (4 ranks/rankWithTies copies).

Change

One pearsonR / spearmanR / ranks added to src/statistics.ts — already the home of the project's stat primitives (pairedBootstrap, cohensD, mcnemar, …) — and every call site routed through it. Newly exported from the root entry.

Edge-case contract (explicit, documented in-file):

• length mismatch or n < 2 → NaN (correlation is undefined; distinct from a measured 0)
• both series constant → 1 (degenerate perfect agreement)
• exactly one series constant → 0 (no covariation to detect)

This matches the dominant (5/7) convention and is the statistically defensible one.

analyze-runs inter-rater kappa now averages only the finite pair correlations, so a degenerate pair (a single jointly-rated run, or a constant rater) yields NaN and is dropped rather than dragging the mean toward an artificial 0 — the same pattern judge-calibration's avgPairwise and correlation-study's bootstrap already use.

All 7 local copies (pearson/pearsonR/spearman/spearmanR/ranks/rank/rankWithTies) deleted. Net -52 LOC.

Verification

• pnpm run typecheck — clean
• pnpm run lint — exit 0 (pre-existing warnings only, none in touched files)
• pnpm run build — clean (tsup + openapi)
• pnpm test — 246 files, 2515 pass, 2 skipped
• New describe blocks for pearsonR/ranks/spearmanR cover the full edge-case contract (n<2 → NaN, both/one constant, length mismatch, ±1 linear, non-linear monotone Spearman, average ranks for ties).

[tangletools]

✅ Auto-approved PR — 4b72e2ae

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-22T13:40:02Z}

[tangletools]

✅ Auto-approved PR — e7ca9d1c

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-22T13:42:38Z}