[ellispatrick]

For testing only, not for merging. Holds #135, #136, #137, #138 and #139 together on top of #134, as #125 does for #126–#134. The later stacked PRs (defaults and dependencies, classifier registry, new classifiers) are based on this branch.

What it contains

PR	Branch	Contents
#136	review/10-metrics-defaults	metrics and plots (Q6–Q12, Q19)
#135	review/11-classifier-fixes	classifier fixes (Q13, Q22–Q24, Q26)
#137	review/12-speed-defaults	speed and defaults (Q1, Q2, Q5, Q15, Q25, Q31, Q32)
#139	review/13-glm-cv-lambda	penalised GLM chooses lambda by cross-validation (Q14)
#138	review/14-crossvalidate-null-outcome	prepareData stops on a NULL or empty outcome

Decisions: ProjectClassifyR docs/DECISIONS.md, 10 October 2026.

Conflicts resolved

• Penalised GLM (12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137 and 13/ Penalised GLM: lambda by 5-fold cross-validation; lambda filter considers every class (changes results) #139): two classes use the binomial model (12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137), and lambda is chosen by cross-validation of that model (13/ Penalised GLM: lambda by 5-fold cross-validation; lambda filter considers every class (changes results) #139). The non-empty-lambda filter accepts the binomial coefficient matrix as well as the multinomial list.
• Ensemble selection (11/ Classifier fixes: naive Bayes crossovers, Fisher score, kNN exact matches, previousSelection, ensemble selection, tuning warning #135 and 12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137): 12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137 split .doSelection into a cache wrapper and .selectFeatures; 11/ Classifier fixes: naive Bayes crossovers, Fisher score, kNN exact matches, previousSelection, ensemble selection, tuning warning #135's rewritten ensemble selection is kept. Its nested-CV tuning now uses the inner scheme of the training set (.innerCVperformance), like other selection, instead of rerunning the outer scheme. A test covers it.
• Tests: the tests of both GLM PRs are kept.

Evidence

• 631 test expectations pass.
• Regression harness (18 scenarios) against each PR's own build. Every difference comes from the PR that changes it:
  • 12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137: random forests (asthma_RF, proc1_ghist_RF, proc2_RSF);
  • 13/ Penalised GLM: lambda by 5-fold cross-validation; lambda filter considers every class (changes results) #139: GLM lambda (asthma_GLM_LASSO, the clinical model of proc3_pathways);
  • 10/ Metrics and plots: no internal rounding, tied risks count 1/2, fold grouping for all metrics, chance line per metric (changes results slightly) #136: no rounding (proc4_crisscross).
  • The other 11 scenarios are identical to 12/ Speed and default changes: random forest ranking and mtry, binomial GLM, likelihood ratio, XGB rounds, merge selection cache, light nested CV #137's build.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TfsjwkbDbHkX4kJZNe4ER4