clawbot
50816b7415
All checks were successful
check / check (push) Successful in 3m2s
PR #89 stopped script/cibuild replaying cached check layers, but left a gap: a bare `docker build .` with no --build-arg still faked. An unset ARG is an empty string, an empty string is a stable cache key, and the check layers replay from it. That gap mattered because REPO_POLICIES.md names `docker build .` verbatim as a command that must be green, so the documented command was the one that lied. Both check stages now carry `RUN [ -n "$CHECK_EPOCH" ] || exit 1` immediately under their own ARG. Failed steps are never cached, so this fails on every invocation rather than once - a bare build now stops with a named error instead of reporting a green it did not earn. Each stage needs its own guard because ARG scope is per-stage; a gate-carrying stage without one is a silent hole if ordering ever changes. The check RUNs now reference the value (`echo "check epoch: ${CHECK_EPOCH}" && make <target>`), so the cache miss is contractual rather than resting on BuildKit's current treatment of unreferenced ARGs, and the epoch is visible in the build log. The epoch becomes "$(date +%s%N)$$" so concurrent invocations in the same second cannot collide. busybox silently drops %N and exits 0, so $$ is what makes it correct there. The bare-assignment form is retained deliberately: inlining the substitution into --build-arg would, under set -eu, yield an empty and therefore constant epoch without aborting. script/docker gets the same treatment - it is not the gate, but two entrypoints disagreeing about whether the tree is green is its own hazard, and local builds are almost always warm. Verified by negative control rather than inspection: a bare build fails twice consecutively here and succeeds twice on the parent commit, so the change is demonstrably not a no-op. The builder-stage guard was fired directly with a targeted probe build, since the lint stage otherwise fails first and would leave it unexercised.