PythiaStage.evaluate scores a trained algorithm against a fabricated all-false truth column (mirrors now-fixed MATLAB #58) #345

Description: andremun/InstanceSpace issue #58, filed while reviewing this repo's PR #322, documented that MATLAB's PYTHIAevalMode scores every trained classifier against its test-set truth column unconditionally — including a trained algorithm the test set has no data for, whose truth column is a reconciliation artifact (NaN >= x evaluates to false in MATLAB), not real ground truth. That issue's own text states "the Python side was found to correctly mirror it" at the time — i.e. this port deliberately replicated the bug for parity, per this repo's "MATLAB is the behavioral authority" policy.

MATLAB's issue #58 was closed today (2026-09-26), fixed by andremun/InstanceSpace@c62408f. This port has not been updated to match.

Where the old (now-stale) behavior lives:

• instancespace/stages/pythia.py::PythiaStage.evaluate loops over every non-None fitted classifier slot and scores it against y_true, with no way to know whether that column reflects real test data or _build_test_algo_matrix's reconciled all-NaN/all-False placeholder.
• instancespace/instance_space.py::InstanceSpace._build_test_algo_matrix already computes a has_ground_truth mask (used for compute_binary_performance's widened comparison) but _explore_evaluate (instance_space.py:1215) discards it: y_raw_test, _ = self._build_test_algo_matrix(...).
• The docstrings at instance_space.py:1183-1187 and pythia.py:472-479 both currently document the old MATLAB behavior as intentional ("exactly as MATLAB does").

Evidence this is real, not theoretical: a report-only CI job (matlab-oracle-verify.yml, added on PR #343) ran the real MATLAB exporter against andremun/InstanceSpace's current master and, after separating platform/run noise from real code-driven drift (control run pinned at the old gold commit vs. current master), found exactly 8 files whose content is attributable to actual MATLAB code changes since the pinned commit — all 8 are PYTHIA's own hyperparameter/selection/eval outputs for the legacy_svm and trace3_pythia_skip variants (build_data/pythia/legacy_svm/outputs/{hyperparameters,raw_metrics,selection0,selection1,summary,ysub}.csv, explore_data/pythia/legacy_svm/outputs/eval_summary.csv, explore_data/pythia/trace3_pythia_skip/outputs/eval_summary.csv). Full breakdown on #344.

Suggested fix, mirroring MATLAB's own resolution:

• Thread has_ground_truth through to PythiaEvaluateInput (new field).
• In PythiaStage.evaluate, skip the confusion-matrix computation for a trained-algorithm index where has_ground_truth[index] is False, leaving its accuracy/precision/recall as NaN and its cvcmat row as zero — the same treatment already given to an empty (None) classifier slot.
• Update instance_space.py::_explore_evaluate to pass the real mask instead of discarding it.

Safety check already done: no current tests/fixtures/matlab/current-based parity test reads the specific 8 affected files for a PythiaStage.evaluate comparison — test_current_matlab_trace_parity.py's similarly-named raw_metrics.csv/eval_summary.csv reads are under build_data/trace//explore_data/trace/, a different stage directory from the affected pythia/ files. test_current_matlab_stages.py's one PythiaEvaluateInput call site is the trace3_pythia_skip variant, whose classifier slots are already all None (training was skipped), so the fix doesn't change that test's outcome. Fixing this code should not require touching tests/fixtures/matlab/current or its pinned manifest.

Compatibility tag: [Behavior-changing] — InstanceSpace.explore()'s reported accuracy/precision/recall/cvcmat for a trained algorithm absent from a test set's metadata changes from a fabricated score to NaN/zero-row, matching corrected MATLAB behavior.