Many pushes to master left over a hundred runs queued, because every
push used its own concurrency group. Group push runs by ref and cancel
the in-flight run on master for the test, lint and E2E workflows. PR,
tag, schedule and dispatch runs keep groups that a master push cannot
cancel.
The intel cache job pushes one fixed tag, so it now keeps at most one
pending run per event and never cancels a running build.
image.yml and backend.yml are unchanged: they build from the diff of a
single push, so replacing a queued run could skip a code commit. Document
the policy in .agents/ci-caching.md.
Assisted-by: Claude:claude-sonnet-5-5 python3
Ettore Di Giacinto
f0864a47da