The first beta's release workflow failed in publish_docker at "Verify
exact tags and signed provenance": gh attestation verify copied an image
blob from Docker Hub and got a 502 Bad Gateway. The verifier calls
docker buildx imagetools inspect and gh attestation verify against
Docker Hub and GHCR with no tolerance for a transient registry error, so
one 5xx fails the whole release, and the train can rerun it only once.
Each of those reads is now retried up to four times with doubling
backoff from 10 seconds. They are idempotent, and a moved tag or an
invalid attestation fails identically on every attempt, so a real
failure is delayed by the backoff and never masked. The script's stdout
is the step output, so each attempt's stdout is captured and only the
successful attempt's is emitted: a failed read that printed half a
manifest cannot corrupt the verified digests. VERIFY_REGISTRY_ATTEMPTS
and VERIFY_REGISTRY_WAIT_SECONDS tune it, validated as small whole
numbers so the retry stays bounded.
Tests use doubles that fail like a dropped connection: transient
attestation reads are retried and only the digest proof is emitted; a
digest read that prints a partial manifest before failing leaves no
trace in the output; an exhausted attempt budget fails with an explicit
error and no output; unbounded settings are rejected. The old assertion
that an unverifiable attestation costs exactly one read now expects the
four attempts. Against the previous script the new tests fail. The
contract records the retry.
Checked: all 9 tests in the verifier's test file pass; the script and
its test are not registered in a subsystem, so the completion guard has
no proof file to require.
Second of the release-path retries after #2862 (release listing 504).
rcourtman
339539a0ca