Guard patient partitions and IRLS convergence; add reproducible comparator benchmark

[aRealGem]

Summary

This draft combines the reviewed patient-partition guard, IRLS convergence guard, and reproducible alternatives benchmark.

• Validate non-missing patient IDs and integer cross-validation partitions in all six released IMPROVE random-forest tables. Reject a patient appearing in multiple partitions before either fixture is written. The documentation distinguishes metadata validation from an audit of upstream training or NNAlign split assignments.
• Return unavailable (None) when IRLS exhausts its iteration budget without satisfying the existing convergence test. Keep the estimator, tolerance, and iteration limit unchanged. Add quasi-separation, finite-MLE budget-exhaustion, recalibration, and report regressions.
• Add a bounded comparator benchmark using a held-out WDBC fixture and a synthetic repeated-patient fixture, with matched estimands and an independent sklearn/statsmodels control. Execute Python/PyCalEva and R probably/rsample comparisons, and clearly label the isitfair adapter as source-reviewed but unexecuted.
• Preserve historical benchmark evidence and add separately labeled integrated verification. Generated row-level fixtures/predictions are excluded.

Verification on the reviewed, identical file tree

• Combined Python suite: 282 passed, 20 optional fixture-dependent skips, no failures or expected failures.
• Comparator contracts: 6 passed, including the formerly expected-failure quasi-separation case.
• Ruff lint, Ruff formatting, and git diff --check: passed.
• Two integrated Python comparator reruns matched all 13 stable input/result hashes.
• Seven retained R evidence hashes verified. R computation was not rerun during integration; the adapter and exported inputs were unchanged.
• GitHub CI completed successfully on published head 4829570: lint and formatting passed; 276 tests passed, 20 skipped. The configured CI suite runs tests; the six comparator contracts were additionally covered by the combined local suite above.

See integration verification and the benchmark report.

Evidence limits

• The 20 skipped checks require optional PCam, GUSTO, or IMPROVE fixture/golden inputs; full golden-output revalidation is not claimed.
• Both comparator gain intervals cross zero. They condition on fixed held-out predictions and omit correction-training uncertainty; they do not establish repeated-sampling coverage.
• The convergence guard can make a valid finite-MLE fit unavailable when more iterations are needed. It is not a universal separation detector.
• Competitor corrections with different estimands/configurations are reported separately. No clinical validation, universal superiority, measured developer-time savings, or between-tool speed claim is made.
• The separate statistical-stress study at local commit df23b1c, including validation/statistical_stress/** and tests/test_statistical_stress.py, is excluded.
• This PR remains a draft and unmerged. Existing PR IMPROVE third-party predictor panel, plus a fix for the coverage-delta column it broke #2 is unchanged.

Publication provenance

GitHub connector publication assigns new commit metadata, so published commit IDs differ from the reviewed local IDs. Every one of the six published file-tree hashes was verified identical to its reviewed local counterpart, preserving the original commit sequence and file contents. The source hashes and local-only statements retained in the benchmark records describe the historical verification stage before this publication.

Reviewed local head: 150998f76df1141c73b6c3621233e693c4eb0b49
Published head: 4829570fd3809c1fda779e8ea24fc9017fcb43a4
Identical complete tree: cacfe06f6bad5942962ce2b1699d4a95aa0dce1f

Local → published commit mapping:

• 99ca3d72f1d4f8340df9a59de3616277e9ebdb7a → b5fe8a39466c469576b5955ae4b6851cf607a5f5
• dae6e1ad1c9c20bcd179e5db7cf9f71e3d888beb → af4c88c913dd45889332a24947061ba51f64fd56
• 756e5a1956578a340eddb90806813ddee62e30fc → eb22fde21e5d18e6f57f9f354f717ad2a31ee0e5
• 868276ef960f0bc7f2bc75cc79926ff7b3c2e1f1 → 9533c0b75ddea5c6a39abc59746d6ae020a7139f
• 617ddac480849ac48de41c1c1b1d050ddc427492 → 7792ba8402556a972d4aa1533233769636e39a35
• 150998f76df1141c73b6c3621233e693c4eb0b49 → 4829570fd3809c1fda779e8ea24fc9017fcb43a4

Author and committer were verified as aRealGem <203440826+aRealGem@users.noreply.github.com> on each published commit.

Co-developed-with: OpenAI Codex