[Divo-Lee]

Summary

• Reuse the existing target-aware Gumbel and empirical-map functions for experimental raw-SW and predicted window-20 composition-adjusted SW fitting.
• Exclude overlapping multi-label Rfam family or clan annotations before cluster maxima; retain iterative US-align leader auditing and within-cluster backfill.
• Use a continuous endpoint-anchored power tail above the empirical knots, capped by the cluster count. Previously fitted parameter files keep their original behavior.
• Allow a wider initial E-value audit ceiling and record the highest directly supported reported E-value. Bind cached scores and US-align results to their database and query panel.

This PR changes package algorithms and fitting commands only. It includes no fitted parameters for RNAsolo, PDB-RNA, or NuFold. Each deployed database needs its own new fit and audit. LOFO is the validation protocol; final deployment parameters are fitted on the designated full panels.

Verification

• Focused empirical E-value tests: 10 passed.
• Full Windows suite: 89 passed, 2 skipped; one unrelated legacy calibrate CLI test fails because its shell-script US-align fixture cannot execute as a Win32 program (WinError 193).
• Both fitting CLI entry points start and display the new audit option.

Values beyond the observed empirical range are extrapolations and are not independently validated.