Performance blocker: make every canonical estimator operation beat sklearn/Python

[godofecht]

Goal

The canonical flow-scikit benchmark should have zero supported estimator operations materially slower than sklearn/Python on the same machine and numerical backend.

Current canonical-v2 evidence has moved beyond the older SVC.predict ~0.67x claim. #481 records current KernelSVC prediction as faster than sklearn on both Iris and Digits and explicitly identifies the older architecture-map inference claim as stale. Do not treat SVC prediction as the first blocker unless a fresh local operation-level run reproduces a loss.

The current concrete losses are tracked by focused children, including LinearSVC fit (#479), KernelSVC fit (#480), DecisionTree Digits (#482), RandomForest Digits fit (#483), GaussianNB Digits predict (#484), KMeans Digits fit (#485), Lasso fit (#486), and KernelRidge fit/predict (#487/#488). Artifact/provenance drift is tracked in #493.

Required work

1. Re-run the canonical suite with current main, pinned BLAS thread counts and the five-repeat measurement protocol.
2. Classify every loss by substrate: Python/control overhead, allocation/copy, scalar loop, BLAS/LAPACK dispatch, algorithmic mismatch, cache/layout, or measurement noise.
3. Fix losses one operation at a time, preserving sklearn-equivalent semantics and existing parity tolerances.
4. Add scaled workloads for every canonical row so wins are not artifacts of fixed interpreter/import overhead.
5. Track fit/predict/transform/predict_proba separately; no estimator-level average may hide a slow operation.
6. Add peak RSS/allocation evidence for representative fits and inference paths.

Current priority

Use operation-level evidence, not stale estimator-level summaries. Before optimizing a row, confirm the current local result and provenance. When a previously losing operation becomes a win, update generated architecture/performance artifacts through #493 instead of carrying the obsolete classification forward.

For KernelSVC specifically, keep fit and predict separate: #480 owns training efficiency; #481 owns prediction-only scaled verification and stale-artifact cleanup.

Exit criteria

• zero canonical supported rows below 0.95x Python after noise filtering;
• target state: every row >= 1.05x Flow win;
• ties must have a filed optimization hypothesis rather than being silently counted as wins;
• machine-readable report lists wins/ties/losses by operation and workload size;
• benchmark documentation explains why Flow wins each class rather than only publishing ratios.

Parent programme: flooooooooooow/flow#727.

Validation and benchmark qualification are local-only. Do not add, enable, trigger, inspect, wait for, or rely on GitHub Actions or other hosted CI/workflows.

[godofecht]

Full audit decomposition created from the current canonical v2 artifact (19/19 eligible, 9 Flow wins, 10 sklearn wins).

Current-loss implementation issues: #479 LinearSVC fit; #480 KernelSVC fit; #481 KernelSVC predict/evidence cleanup; #482 DecisionTree Digits; #483 RandomForest Digits; #484 GaussianNB Digits predict; #485 KMeans Digits fit; #486 Lasso fit; #487 KernelRidge fit; #488 KernelRidge predict.

Measurement/evidence holes: #489 operation-level rows; #490 scaling surfaces; #491 RSS/allocation/copy evidence; #492 learned-model-state diagnostics; #493 generated-artifact provenance/drift; #494 dynamic execution-substrate coverage; #495 Tier-1 performance qualification; #496 Python interop/zero-copy overhead; #497 every implemented estimator operation benchmarked; #498 shared hot primitive audit.

Important audit correction: architecture_performance_map.json is stale relative to the current frozen canonical result in at least some rows (e.g. Digits LogisticRegression), so #493 is a prerequisite for trusting derived roadmap classifications. Also, several end-to-end losses are fit-only while prediction already beats sklearn; #489 prevents totals from hiding that phase information.

#478 should close only when these child surfaces are green or explicitly removed from supported scope.

[godofecht]

Two additional audit layers after the first decomposition: #499 adds hardware-counter/hot-symbol profiling for every canonical loss; #500 forces an explicit pure-Flow vs native-library backend strategy for liblinear/libsvm-class workloads rather than treating external-native gaps as generic compiler problems.

[godofecht]

Fairness-contract additions: #501 materializes and checks estimator hyperparameters/stopping rules/effective work budgets; #502 makes dtype/layout/preprocessing/conversion work explicit. These prevent apparent speedups from being accepted when the two sides are doing different numerical work or timing different conversion boundaries.