feat(tier3): report pass@k uncertainty and paired outcomes

[AdemVessell]

Summary

• add two-sided 95% Wilson score intervals to each condition's case-level
  pass@k rate;
• add direction-preserving paired outcome counts and paired rate delta;
• add a two-sided exact McNemar diagnostic only for completely paired cases;
• render the new evidence in the Tier 3 HTML report and document its limits.

Why

An aggregate with-skill minus without-skill rate does not reveal whether gains
and regressions occurred on the same cases, and a point estimate alone obscures
small-sample uncertainty. The additional fields make the existing result more
auditable without changing its pass/fail policy.

Scientific boundary

The intervals and exact test describe one configured run. They do not establish
task independence, correct leakage, prove generalization, or make the test set a
random sample. Partial/unidentified pairings receive no exact-test result.

Validation

• PATH="$PWD/.venv/bin:$PATH" make test PYTHON=.venv/bin/python
  • 4,862 passed, 17 skipped, 4 deselected
• make lint PYTHON=.venv/bin/python
  • passed
• focused collector/report tests
  • 34 passed
• make build PYTHON=.venv/bin/python
  • wheel and source archive built

Compatibility

This is an additive JSON/report change. Existing pass@k values and verdict
thresholds are unchanged.

Addresses #62.

Verification checklist

• I am familiar with the Contributing Guidelines
• Added or updated focused tests
• Updated documentation for user-visible changes
• Ran make lint
• Ran make test
• Ran make build
• Did not add credentials, private datasets, or proprietary benchmark content

Release impact

• Updated CHANGELOG.md

[AdemVessell]

Pushed follow-up 444af99 after a hostile correctness review of the paired
pass@k evidence.

The follow-up:

• prevents small nonzero exact McNemar probabilities from being rounded to
  zero and preserves decimal plus exact-rational representations (with a
  null numeric approximation instead of false zero on float underflow);
• reports the minimum attainable p-value for the observed discordant-pair count
  and labels results that are resolution-limited at alpha 0.05;
• derives the arm pass-rate delta from counts, avoiding contradictions caused
  by subtracting already-rounded rates;
• renders the paired delta, suppresses the zero-pair row when pairing is
  unavailable, and keeps the exact test restricted to complete pairing; and
• adds a collector-path regression where reward IDs are absent and case IDs
  are derived from Harbor trial metadata.

Validation on the final head:

• full suite: 4,874 passed, 17 skipped, 4 deselected;
• focused collector/report suite: 81 passed;
• Ruff and diff checks: clean;
• wheel and sdist build: passed;
• Fern docs check: zero errors (redirect check skipped without Fern auth).

An independent cold re-review found no remaining actionable defect. The new
statistics remain diagnostic for one configured run; they do not establish
task independence, generalization, or causal skill benefit. This follow-up also
corrects an existing pass-lift rounding defect, so the stored delta may change
by one unit in the fourth decimal while verdict thresholds remain unchanged.