fix(heldout): declare candidate-group DIF sample size

[seonghobae]

Stack

Draft successor of #1098 (e39a99f0). Parents #1097, #1096, #1095, #1093, #1092, #1091, #1090, #1074, and psychometric KPI owner #1067 stay open. This PR does not close #1098.

Exact head: b975c3dde9c74a862696aa49c6668d6a2e3f567a.

Problem

Held-out candidate-group DIF evidence hid DIF_SAMPLE_SIZE = 4_000, so two-cohort purification used a repository-selected population.

Change

• Require declared even sample_size on _validate_candidate_group_dif. Missing, boolean, non-positive, or odd values fail closed.
• The harness run still writes 4,000 as this run's choice and records sample_size.
• ADR 0050 is Proposed. Production route/conduct defaults stay locked.

Verification

Local smallest tests:

• 62 passed in 16.15s (tests/test_psychometric_routing.py, tests/test_psychometric_benchmark_boundaries.py, tests/test_planning_adr_identifiers.py)
• interrogate on the held-out script: 100%

This is declaration-contract evidence, not buyer-held-out accuracy, p95 latency, or protected merge.

Remaining no-heuristics work

Other harness sample sizes (reliability, judge, exploration rate) remain later slices. No production policy or release is authorized. Parent #1067 still needs independent review.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Normal merge at da4ceac preserves original DIF sample-size delta b975c3d and incorporates predecessor ae5aabf. Carries no-detection censoring repair, explicit alarm/detection denominators, null candidate when no calibration threshold is eligible, and retained calibration results. Merged-tree checks: 29 boundary and ADR identifier tests passed in 3.44 seconds, Ruff passed, diff whitespace check passed. CodeGraph indexing completed. DIF generation still uses the explicitly declared even sample count; no synthetic results are promoted to customer-data accuracy evidence. No force push, discarded delta, production change, or release claim. Hosted checks must validate this new head.