V50: metrics that agree with the reference on every edge

[dapiced]

Summary

Each edge below was measured before any code changed, and is now checked against scikit-learn instead of our own reading of it.

Edge	Before	After
Perfect batch containing 2 of 3 classes	macro F1 0.667	1.0 (scikit-learn)
Class with 1 row (100 + 1 rows)	test 1, train 0	train 1
Class with 2 rows	test 1, validation 1, train 0	test 1, train 1
Empty set	RMSE 0, MAE 0, R² 1	NaN

1. Macro averages. macroPrf and the V45 decision policy now average only over the labels present in y_true or y_pred, like scikit-learn. Previously they averaged over every class the task knows, so absent classes counted as zeros.

• The V42 generator gets a new absentClass case: four classes declared, three present, and one predicted but never true. The old code failed it.
• I regenerated the reference with scikit-learn 1.7.2. Apart from the new case, the only diff is the Python patch version.

2. Rare classes. A class with a single row now stays on the training side. Before, it went entirely to test, and a two-row class lost its other row to the validation carve.

• Classes with 3 or more rows are unaffected, and the seeded draw for every other class is byte-identical.
• If every class has one row, nothing is left to test on. Training now stops with the named refusal too-few-rows instead of reporting metrics over an empty set.

3. Empty sets. Every metric computed over zero rows now returns NaN. No caller passes an empty set today, so this is a contract for future callers.

Measurement on real demos

Three targets offered by the demos have one-row classes: titanic → parch, cafe-sales → city, and cafe-sales → product. Macro F1 on titanic → parch:

Configuration	GBDT	Forest	Ensemble
old split, old metric	0.619	0.310	0.422
old split, new metric	0.619	0.310	0.422
new split, old metric	0.614	0.340	0.470
new split, new metric	0.716	0.397	0.548

• With the old split, every class is in the test set, so the new metric changes nothing.
• The gain comes from the split. Once one-row classes stay in training, the test set no longer contains them, and the new average stops counting them as zeros.
• The model picked on validation changes from GBDT to the forest, because the validation rows changed. This is a selection effect, not a metric change.

Test plan

• metric-edges.test.ts: 6 of 7 fail before the fix, all pass after. The new golden case also fails before.
• npm run test:coverage: 906 passed, floors met (64.6 / 53.47 / 54.48 / 65.66)
• lint, typecheck, format:check, build
• Playwright, Chromium: 97/97

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELAjHYDyxa5zbbdcke1bMA

---

Generated by Claude Code