v1.36.0

What's Changed

• feat(evals): bind every A/B delta to the actor that produced it by @CybotTM in #229
• fix(evals): grade must_not assertions in the direction they state by @CybotTM in #233
• fix(evals): isolate the A/B measurement, and report evidence-free evals by @CybotTM in #234
• feat(evals): decide every assertion by majority over repeated samples by @CybotTM in #237
• feat(validate): catch a shebang without the committed exec bit by @CybotTM in #239
• docs(composer-setup): deny the skill plugin in a repo that ships PHP code by @CybotTM in #231
• feat(evals): gate on the evidence ratio, which holds still between runs by @CybotTM in #240
• fix(validate): close four coverage gaps in the shared validation stack by @CybotTM in #241
• fix(validate-evals): validate every evals.json, not the first one found by @CybotTM in #243
• feat(tests): run PHP test entrypoints, not only shell and Python by @CybotTM in #244
• fix(ci): bound apt and every reusable job so a stall cannot hold a runner by @CybotTM in #247
• fix(ci): make the shellcheck hook and CI the same check by @CybotTM in #249
• feat(validate): enforce the Agent Skills budgets and flat discovery by @CybotTM in #250
• chore(release): v1.36.0 by @CybotTM in #251

Full Changelog: v1.35.0...v1.36.0