Highlight supervision

Where a corpus annotates its training split, a task can train the selector against those annotations instead of leaving it to discover them:

Registry.from_key(TOY_TASK, highlight_supervision=True, highlight_coefficient=0.5)

The flag appends HIGHLIGHT_LOSS, masked cross entropy over highlight_logits, highlight_true and mask, to the model the task names, weighted by highlight_coefficient. Nothing else changes: the same model key runs either way.

The two settings are different experiments, not two points on one scale. The unsupervised one is the realistic problem; the supervised one is the ceiling it is measured against, and its val_loss carries a term the other has no equivalent for, so only the metrics compare.

Supervision needs exactly one selector. One annotation supervises one selector. A model with several selectors, such as MGR, is refused, because its selectors are there to diverge without supervision, and supervising one of them would leave the others without a role.

A corpus without training annotations is refused. Unannotated positions are padded with -1 and skipped by the criterion, so supervising a corpus annotated on test alone would train exactly as an unsupervised run does and report itself as that run’s ceiling. The task checks the training split and raises instead.

GenSPP refuses the flag outright: no gradient reaches its generator, so the loss would be built and train nothing. Guiding a genetic search means conditioning the population it draws from, which is open work.