Per-locale confidence calibration — en-us v4.0.0 (#368 L2)

A separate isotonic table per locale, vs the single global table. ECE is measured on each locale's held-out split under three regimes: raw softmax, the global table, the locale table.

locale	n	accuracy	ECE raw	ECE global-table	ECE locale-table
NL	201	0.985	0.1733	0.0466	0.0096
DE	197	0.893	0.0932	0.0889	0.0411
US	5315	0.973	0.0674	0.0044	0.0037
FR	796	0.966	0.0641	0.0138	0.0117

Where the locale-table column beats the global-table column, a single global table is leaving calibration error on the table for that locale (the OOD locales especially). A multi-locale model should ship one calibration table per locale, selected by the locale gate.