S4: numpy AP and MTD kernels alongside scikit-learn, with the full parity suite — no dispatch change yet (#64 gate, part 1)

[Polichinel]

Epic: #66 · Wave 1 · Gates: S5 (#64) · Blocked by: none · Ships no behaviour change

Background

#64's gate: scikit-learn stays the running implementation until a comprehensive parity suite exists, passes against it, and has been seen failing on a deliberately broken numpy kernel. This story is that gate. It adds the numpy kernels as private functions next to _crps_ensemble_numpy (native_metric_calculators.py:46-77, the precedent) and the parity suite, and changes nothing that dispatches.

Measured on scikit-learn 1.7.2 (2026-09-17), the semantics to reproduce:

• AP — _binary_clf_curve: order = argsort(y_score, kind="mergesort")[::-1]; y_true == 1; thresholds at np.where(np.diff(sorted_scores))[0] plus the last index (ties collapse to one operating point); tps = cumsum(float64)[idx], fps = 1 + idx - tps. precision_recall_curve: precision tps/(tps+fps) with zeros where the denominator is 0; recall tps/tps[-1], or all-ones with warnings.warn("No positive class found in y_true, recall is set to one for all thresholds.") (plain UserWarning) when tps[-1] == 0; reversed, with (1, 0) appended. AP = float(max(0.0, -np.sum(np.diff(recall) * precision[:-1]))). A literal transcription matched sklearn to 0.0 on 300 tie-heavy random cases. Accepted y_true: at most two distinct values and, if two, 1 among them ({1,2} computes with 1 as positive; {0,1,2} and continuous raise).
• MTD — p<0: 2*(where(y>0,y,0)^(2-p)/((1-p)(2-p)) - y*mu^(1-p)/(1-p) + mu^(2-p)/(2-p)); p==0: (y-mu)^2; p==1: 2*(xlogy(y, y/mu) - y + mu) with xlogy = where(x==0, 0, x*log(y)) under errstate(divide="ignore", invalid="ignore"); p==2: 2*(log(mu/y) + y/mu - 1); else the generic form; plain mean; Python float. Domain: p<0 → ValueError("Mean Tweedie deviance error with power={power} can only be used on strictly positive y_pred."); 1<=p<2 → "…non-negative y and strictly positive y_pred."; p>=2 → "…strictly positive y and y_pred."; p==0 no check; 0<p<1 → ValueError. Matched to 0.0 over 11 powers × 200 cases on float64.

Maintainer decisions (2026-09-17): float64 always (sklearn computes float32 in when all inputs are float32, ~3e-7 off — documented deviation); clean ValueError for non-binary AP truth, naming the observed labels, while accepting exactly sklearn's input set.

The work

1. _average_precision_numpy(y_true_flat, y_score_flat) -> float and _tweedie_deviance_numpy(y_true_flat, y_pred_flat, power) -> float in native_metric_calculators.py, on the flattened form the kernels already use (np.repeat(y_true, S), y_pred.flatten()). Not called by anything yet. sklearn import untouched.
2. TestAPParityWithSklearn and TestMTDParityWithSklearn in tests/test_metric_calculators.py, modelled on TestCRPSParityWithProperscoring (function-local from sklearn.metrics import … oracle, pytest.approx(abs=1e-10), seeded RandomState). Content is the S5: Switch AP and MTD to the numpy kernels; scikit-learn to dev deps; import-purity guard; close C-05 (#64 gate, part 2) #64 gate list: AP — perfect, inverted, all-tied, all-positive, all-negative (value and warning message), single positive, N=1, N=2, duplicated scores with mixed labels, S in {1, 10, 100}, 200 seeded draws over N in {5, 50, 5000} × S × prevalence 1%–99%, {-1,1} and {1,2} label sets, and the raise cases ({0,1,2}, continuous) asserting ValueError on both sides. MTD — p in {-2, -1, -0.5, 0, 1, 1.0001, 1.5, 1.9999, 2, 2.0001, 2.5, 3, 5}; exact zeros in y where permitted; y == mu (deviance exactly 0); y ≪ mu and y ≫ mu; S in {1, 10, 100}; 200 seeded draws per branch; every domain rule as a raise test with the message text; float32 inputs at abs=1e-6 with the deviation named in the docstring.
3. Warning parity: each random case run under warnings.catch_warnings(record=True) on both paths, asserting the same warning categories — this is what catches a stray numpy RuntimeWarning from the Poisson branch.
4. Extreme values (1e6, 1e-6) for both.

This story does not change dispatch, pyproject.toml, docs, or the register. grep -rn sklearn views_evaluation/ still finds the import when it merges.

Acceptance criteria

• Both numpy functions exist, private, uncalled by production code
• Parity suite passes at abs=1e-10 (float64) on every listed case; float32 cases at abs=1e-6
• Seen red, recorded in the PR: three injected defects each made the suite fail — drop [::-1] on the sort; skip the tie collapse (use every index as a threshold); flip a sign in the generic Tweedie branch. All reverted
• Existing test_ap_oracle_sklearn and test_mtd_known_tweedie untouched and green
• Zero behaviour change: full suite green with no test modified other than additions; ruff clean

Validation

conda run --name views_pipeline pytest tests/test_metric_calculators.py -q
conda run --name views_pipeline pytest tests/ -q
conda run --name views_pipeline ruff check views_evaluation/ tests/
git diff --stat -- views_evaluation/   # native_metric_calculators.py only, additions only

[Polichinel]

Merged to development in #77. Ritual: /review-diff, /code-review max (15 findings), /falsify guard (25 mutations + probes; 12 survivors and 1 code finding closed), /register-risk, /ship-it. 937 tests green; 26 injected defects observed red across the two passes. The gate #64 needs is in place; #64 also carries a needs-decision on the all-negative AP branch (ADR-015).