feat(evals): isolated harness core

droid CLI wrapper with verified solver isolation (--disabled-tools +
neutral cwd + ignore-directive so the solver can't invoke other skills
or read the repo), length-matched placebo/trigger conditions, Wilson/
sign/McNemar stats, and an N-model majority judge panel.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>