[fraware]

Post-freeze research/software correction only. Empirical main and all physical campaign semantics remain untouched.

The existing isotonic implementation initialized one PAV block per sample. For conflicting labels at exactly the same similarity score, that allowed pair-ID tie ordering to change the fitted probability at that score. That is not a valid property for a calibrator whose predictor is the score itself.

This PR:

• pools all exactly tied predictor scores into one weighted empirical-rate block before pooled-adjacent-violators regression;
• bumps the calibration algorithm identifier from sonic-twin-isotonic-calibration-1 to sonic-twin-isotonic-calibration-2 while keeping schema version 1;
• rejects one-class direct calibration fits because they cannot calibrate same-physical-specimen probability;
• adds regressions for tied conflicting labels, pair-ID/input-order invariance, ordinary PAV pooling, and one-class rejection.

No physical calibration evidence exists. These changes improve the correctness of software-only calibration mechanics and do not promote any Sonic Twin identity claim.