[ellispatrick]

Part of a stack of 8 pull requests from the October 2026 audit of ClassifyR (AUDIT.md on branch audit-2026-10). Each part is based on the one before, so they need to be merged in order. Each one can be accepted, changed or rejected on its own merits. Draft PR #125 holds all eight together. Every part installs and passes its own tests.

Stack, merge in order: #126 tests and core fixes → #127 interface fixes → #128 metrics and plots → #129 pathways and criss-cross → #130 maintenance → #131 faster cross-validation (identical results) → #132 shared folds (changes results) → #133 release notes. The regression harness and timings are in SydneyBioX/ProjectClassifyR (private).

Summary

Changes results (Ellis Patrick's decision). All assays, classifiers, selection methods and combinations in one crossValidate call are evaluated on the same training and test splits.

Why

Each assay, classifier and combination used to get its own random splits, drawn from the global random stream that earlier cross-validations had advanced. Two assays' fold assignments agreed for only 19% of samples, which is chance for 5 folds. So comparisons between assays in performancePlot and samplesMetricMap, as in the protocol paper, were not paired and included split-to-split noise. With shared splits the comparisons are paired.

What changes

• The first cross-validation of a call keeps its seed and splits, and every later one uses them.
• With shared splits, the final models no longer need to be fitted one after another in the main process to keep the random streams reproducible, so they are fitted in the worker pool, from the random state that follows making the splits.
• ?crossValidate documents the shared splits.

Do results change?

Yes:

• The first cross-validation of a call is the same as before.
• Later ones differ, because their folds differ.
• The results are again identical for any number of cores.

Evidence

• Tested: every assay, classifier and merge combination has the same fold assignments.
• The regression harness (part 6 → part 7): scenarios with one cross-validation per call are identical (DLDA, random forest, SVM, leave-one-out, criss-cross), and scenarios with several differ in the later ones.

To reject

Close this pull request. Part 6 alone keeps separate splits, with identical results to before.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4