feat(k-index): publish a level chosen per band, not one shared quantile

[junwen94]

A single shared quantile means the easy majority and the hard tail share one number: q0.6 under-predicted 76.7% of the time for structures that genuinely need rung 12 or above, while the bulk of the dataset (rung < 7) already sat at 12.1% with room to spare. Raising the one shared level to fix the tail would tax the majority with mesh it does not need.

quantile_by_band cuts structures into bands by the model's own median and picks a separate published level per band -- the cheapest one keeping that band's own under-prediction at or below 5% on validation, not the dataset average. Bands land on q0.90 below rung 7 and q0.975 at or above it. The estimator is unchanged -- same pickle, same SHA-256 -- only the decision rule changed.

Measured on the held-out test split: under-prediction falls from 22.9% to 4.2% overall and from 76.7% to 14.7% for rung 12 or above, at the cost of MAE rising from 1.135 to 2.974 and R2 going negative (-0.305) -- a real trade, not a free win.

Two things get worse in isolation, both bounded downstream rather than fixed here: small, simple cells are pushed to denser meshes than before (primitive silicon now predicts rung 38, up from rung 17), which Core's planned metallicity-gated k-distance ceiling is designed to cap; and for a very small number of simple, high-symmetry cells the published rung can exceed the number of meshes that exist on that structure's own ladder (conventional silicon predicts rung 40 against a 38-rung ladder), which Core is designing a bounds check for. Neither is something this decision rule can fix on its own.

Updates the deposit (PSDI 4050a-aas85, in place), the protocol (edited in place, per the same-forest precedent), the model card, and the notebook, re-executed against the new decision.