The n=3 behavior regression owed since v1.4.0's snippet change is complete: 30/30 valid cells, zero quota stubs, no regression. E (full install) is again the only scope-trap PASS median — the third independent run showing the separation's direction (reps PARTIAL/PASS/PASS; '3/3 every run' remains unreplicated and we keep saying so). C produced the run's only real scope violation, again (config.json rewrite). t02 has saturated across all conditions — t04 is now the discriminating half of the regression pair. Full graded record + all 30 raw cells: eval-results-v2/REGRESSION-20260708-r4r6.md.
Also: publish-hygiene, BOOTSTRAP, and SECURITY now route repo intake through agent-zero-trust (azt scan) alongside rulebench vet.
Checked: full RELEASE-CHECKLIST AUTO set (bundle from public @ ca64529), regression at n=3 with seven boundary cells hand-verified against diffs, grader bug caught and documented (prose bullets counted as diff lines — nearly graded the headline result F/F/F). Not checked: live claude -p discovery (unchanged surface since v1.5.0's verified run).