Measured on v0.28.1, tagging to report took about 51 minutes. Most of the
wait was work done twice or work waiting behind something it did not need.

Release prerequisites (16 min, bounded by the Windows unit suite):
- The unit suite still runs on all three platforms, but the second run now
  happens on Linux only. The repeat catches a test that passes once and not
  twice, which is a property of the test rather than of the platform.
- The twelve cross-builds run as three jobs, one per target operating
  system, instead of one after another in the Java stability job. They now
  use -trimpath, the flag a release is built with.
- The two fuzz targets that compile most of the module graph start first,
  so the other 28 fuzz while those compile instead of the job ending on six
  minutes of compiling alone.

Release (29 min):
- The unit tests run beside the build instead of before it. publish still
  needs them to pass.
- The draft-verification jobs skip the Go cache. They only run the
  assurance tool; restoring the cache for it took five minutes on Windows.

Release assessment: the install-script jobs skip the Go cache for the same
reason.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

The unit suite no longer repeats in the release prerequisites stage. The
repeat guarded against a single lucky pass, but the same suite already runs
on every pull request, on every push to main and in the release workflow,
so an intermittent test surfaces there. The unit-portable claim now says
what is done: the suite passes on Linux, macOS and Windows.

Fuzzing was the slowest remaining job, and more targets at once would not
help: it is bounded by compiling the two targets whose packages pull in
most of the module graph. Those two now run as their own job, beside the
28 parser targets, so neither waits on the other. The fuzz check reports
two instances, parsers and engine-and-plugin.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

…onfirmed

Running the unit suite beside the GoReleaser build let the build finish
first when the suite failed, and GoReleaser opens the Homebrew, Scoop and
WinGet pull requests as it runs. The build waits on validate again.

What made the wait pointless is that preflight has already confirmed the
release prerequisites stage for this commit, which runs the same suite on
three operating systems. validate now skips its steps in that case and runs
them, ahead of the build, when preflight could not confirm it.

Also sum targets_failed across fuzz instances, now that the fuzz check
reports two; the report's top-level figure had dropped out.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

Two attempts to avoid re-running the suite in the release workflow were
each found unsafe in review, so the job is back to what it was:

- beside the build, GoReleaser could open package-manager pull requests
  for a release whose tests then failed;
- skipped when preflight confirmed the prerequisites stage, a hand-made tag
  on an untested child of a verified commit would be built without tests,
  and vet would not run at all.

The decision is recorded on the job so it is not tried a third time.

Co-Authored-By: Claude Fable 5.1 <[email protected]>