CI/CD verification steps must produce truthful, reproducible, human-visible evidence under hard environment constraints (time, filesystem, available tools, git history). Concretely:

1) Keep evidence visible across stages

• If a verification fails and you mark a finding status: "dropped", ensure it still appears in the human-facing report stream (not just an internal fixes list). Align .fixes/.reportOnly semantics so “dropped because verification failed” is still auditable.

2) Make verification rerunnable (idempotent cleanup)

• Any scratch directories/worktrees created by the verifier must be either:
  • created with names matching the workflow cleanup glob, or
  • explicitly removed by the skill itself.

Example (safe worktree lifecycle):

# use a directory name that matches cleanup patterns (or remove explicitly)
git worktree add tmp/base-verify-tree HEAD^1
# ... run A/B ...
# always remove what you registered
git worktree remove --force tmp/base-verify-tree

3) Don’t promise evidence you can’t obtain

• If the CI checkout depth/history cannot support a requirement (e.g., per-commit attribution), detect reachability first and then:
  • verify only what is available, and
  • record the missing part as Not covered rather than fabricating tables.

4) Run only gates your container can actually run; avoid mutating the PR tree

• Before claiming lint/tests “passed”, ensure required pinned binaries are present.
• Prefer setup-only install steps, then run non-mutating checks. For example, if setup installs formatters/linters:

  node scripts/lint.js --setup
  # then only non-mutating checks (examples)
  node scripts/lint.js --actionlint
  # never call the mutating no-arg form that rewrites the PR tree
  

• If a tool cannot be installed in the verify container within budget, explicitly report “not run”.

5) Validate controls don’t accidentally execute head code

• In monorepos, symlinks or internal workspace resolution can cause a “base” run to load changed head code. Before trusting an A/B control, confirm the resolved realpaths of internal dependencies point to the intended base checkout.

6) Follow-up reports must snapshot the newest substantive evidence

• For subsequent /verify rounds, snapshot the latest substantive previous report (not transient “running/weak” marker comments) so prior unresolved findings aren’t erased.

Applying these rules prevents CI/CD from silently hiding failures, failing reruns due to stale artifacts, overstating verification certainty, or producing non-comparable A/B results—raising reliability of automated gates and maintainer trust in the evidence.