test(comparison): fix PhosphoRS tier unit bug + re-baseline reference idXMLs (#40)

filter_phosphors compared percentage site probabilities (0-100) against a
decimal threshold (prob_threshold/100 -> 0.99/0.90/0.75), so the STRICT/MODERATE/
LENIENT tiers all thresholded at <1% and selected the same set (identical
counts). Compare in percent directly. The three tiers now stratify, e.g.
PhosphoRS 986 (>99%) / 1105 (>90%) / 1122 (>75%).

Regenerate the reference outputs (data/1_{ascore,phosphors,lucxor}_result.idXML)
from the current, fixed algorithms using the exact CLI invocations the test
uses. The old references predated this PR's scoring fixes (delta-ranking,
double-counting, peak-depth optimization, full-range w), so the regression
baseline was guarding pre-fix output. New==reference now -> recall ~100%.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: timosachsenberg