docs: add v1 vs v2 validation results to stress test report

SI amplifier verified (peak 7.7 vs 6.3), coercion detection on Sonnet
improved from 0/9 to 4/9 steps, behavioral divergence now dynamic
instead of near-constant. Haiku frozen pattern documented as model limit.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>