docs(efficacy): round 4's report and verdicts (#1875) (#1876) (#1949) (#1954)

Both registered hypotheses are supported on a few rows at K = 3:
the thin spec (12.3) on hybrid's judged security and full's and hand's
security probes, the weaker model (12.4) on full's judged security,
hybrid's judged data protection and hand's security probes. Task
success did not grow under the weaker model. Thin hybrid's registered
scores stand beside the key-supplied re-run.

Closes #1875, closes #1876, closes #1949.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Author: braboj