[NEW SUBMISSION] ProSST PIT tail rank scoring function

[emmett-bicker]

This PR adds a new ProSST (K=2048) scoring method called PIT Tail Rank. It keeps the same single wild-type forward-pass setup, but replaces the previous confidence-calibrated delta readout with a minimal equal-weight calibration score.

For each position, ProSST probabilities over the 20 canonical amino acids are converted into three centered residual signals:

• centered lower-tail residual
• protein-level PIT calibration residual
• protein-level rank residual

The submitted score is simply:

score = local_tail_residual + protein_PIT_residual + protein_rank_residual

There are no learned parameters and no tuned scalar weights.

Validation against official vanilla ProSST log-probability delta:

Model	Vanilla ProSST	PIT Tail Rank
ProSST-1024	0.484672	0.490633
ProSST-2048	~0.507	0.524495
ProSST-4096	0.498301	0.500819

The submitted model remains ProSST-2048, where this minimal method gives the strongest absolute result. The same scoring rule also improves ProSST-1024 and ProSST-4096 over their vanilla log-probability-delta baselines.

[emmett-bicker]

Hey @pascalnotin -- When you have time, would really appreciate a review of this

[emmett-bicker]

Hey @pascalnotin -- When you have time, would really appreciate a review of this