File the 3 run_eval.py fixes upstream at anthropics/skills #32

benchmarks/2026-07-07-toolkit-grid/trigger-kit/runner-patches.diff (PR #29) carries 3 fixes to skill-creator's run_eval.py that this environment needs; per ADR 0013 (delegate-don't-vendor) they belong upstream as issues/PRs on anthropics/skills, not vendored here:

1. Hard-fail on any first non-Skill/Read tool call → keep watching instead.
2. Child process inherits parent stdin (3s stall) → stdin=DEVNULL.
3. Trigger-detection window closes at first message_stop → reset and keep watching.

Note: filing requires GitHub access outside this repo's session scope — likely an owner task or a session with wider scope.