Stop a slow CDN from eating the release it is checking

[tonytonycoder11]

The resolution check added in 2.3.0 worked, and cost that release its record.

What happened

Central published all eleven artifacts. Its CDN had not finished serving three of them — kdrant-bom,
kdrant-core and kdrant-transport-rest — inside the ten minutes per module the check allowed, so the
step failed after 38 minutes and reported exactly that. Correct, as far as it goes.

The step sat second-to-last, so failing skipped the four after it:

  Publish to Maven Central                      success
  Check the release resolves from Maven Central  failure
  Create the GitHub Release                      skipped
  Fetch the CLI binaries                         skipped
  Attach the CLI binaries to the Release         skipped
  Record the published artifacts                 skipped

2.3.0 was live and resolvable on Maven Central with no Release object, no binaries and no artifact
metadata. All three were created by hand afterwards: the Release from the workflow's own awk extraction,
the binaries from the run's cli-* artifacts, and the ten storage records from digests computed over the
jars as Central serves them.

Every artifact resolves now, including the three the check gave up on. It was a race with a CDN, not a
failed publish.

The two fixes

Ordering. A check reports on what happened; it does not decide whether the thing that happened gets
written down. It runs last now. A slow CDN still turns the run red and still names the artifacts that are
missing, and the release record exists either way.

Waiting. Thirty attempts of twenty seconds per module, sequentially, made the worst case the sum of
eleven timeouts. It polls the whole set each round now, so the budget is half an hour for the slowest rather
than half an hour each. The summary also says that the release is recorded regardless, so whoever reads a
red run knows what they are looking at.

Why this is not a new check

It is the same check. What changed is that it can no longer destroy the evidence of the release it is
checking, which is the property it should have had when it was written.

[trueup-by-nacode-studios]

This pull request adds no capability the code did not already have.

_{2 files read · 1044 ms · against cf173fc · engine 1.4.1}