[nmbro]

Follow-up to #3790, as suggested in
#3790 (comment)

Changes

No deadline for PR builds, 57 minutes for the others

The CI deadline (CI_DEADLINE_EPOCH, set in init) stops our steps
early so that AppVeyor still has time to save the build cache before
it kills the job at 60 minutes. PR builds no longer save the cache,
so they need no such margin: init now sets the deadline only for
non-PR builds. Without it, the step scripts fall back to 3600 seconds
(this fallback was already there), so PR builds can use the whole hour.

For other builds, the deadline moves from 51 to 57 minutes. Here is
what still runs after it, from the job logs:

what	time
7z of the bundle (-mx=3, from #3790's own job)	~42 s
ccache stats, artifact upload, on_success	~10 s
cache save (master builds)	24-39 s

Master builds that hit the old deadline finished about 35 s after it
(e.g. builds 5804, 5812, 5816 ended at 51.4-51.6 min). A passed build
that ends just before the deadline needs about 1.5 minutes for the
rest, so 57 minutes leaves about 1.5 minutes of margin. With 58, it
would end at about 59.5 minutes, too close to the limit.

ccache cleanup after the test and bundle steps

With APPVEYOR_SAVE_CACHE_ON_ERROR, a failed master build saves the
cache right after the failing step. But ccache -c only ran at the
end of 04-build_script.bat and in on_success. So if the test or
bundle step failed (e.g. by the deadline) after compiling more
objects, the cache was saved without enforcing max_size, and could
get near the per-item cache size limit again.

07-test_script.bat and 08-after_test.bat now run ccache -c right
after make check and make install-win-bundle, whatever the result,
and keep their exit codes. The 08 line drops set -e for that; the
only other command there already handles its own failure. The cleanup
took about 1 s in #3790's job.

Testing

• appveyor.yml parses, all files are ASCII, and the .bat files
  keep CRLF.
• A static check finds no |, &, < or > outside double quotes
  in the changed bash -lc lines.
• The cleanup line, run with stub make (exiting with 124) and
  ccache: the cleanup ran and the exit code 124 was kept.
• Not tested: an actual AppVeyor run.
  • This PR build should print CI deadline: 3600 seconds left in
    each step.
  • Master builds should print a countdown from just under 3420
    seconds (57 minutes).

AI disclosure

Prepared with Claude Code (Claude Opus 5.5, model claude-opus-5-5):
log analysis, changes, local tests and this description. I reviewed
the changes, and I take responsibility for the submission.

🤖 Generated with Claude Code

[github-actions]

A ZIP file with standard source tarball and another tarball with pre-built docs for commit c392718 is temporarily available: NUT-tarballs-PR-3794.zip.

[nmbro]

Also fixes the CodeQL failure on master since 111ca16 (tests/nutsnprintftest.c failed under -Werror), see c3725f6.

[nmbro]

Update: 167f2fd on master already fixed the clang half (-Wformat-nonliteral), same change as here. The gcc half (-Wformat-truncation at the deliberate truncating snprintf()) is still needed - CodeQL gcc jobs on 167f2fd keep failing.

[nmbro]

Dropped the nutsnprintftest commit: master fixed both halves meanwhile (167f2fd, 4248638), so this PR is AppVeyor-only again.

The Jenkins red here is "Discover slow build matrix: Failed to build stage" after ~25 min, with all quick tests passing; #3782 and #3791 fail the same way today, so it looks like a CI farm issue rather than this change.

[jimklimov]

Also tries to fix some other bugs from recent merges, primarily exposed in Windows builds.

[AppVeyorBot]

✅ Build nut 2.8.5.5871-master completed (commit cd949c75df by @jimklimov)

• artifacts
• ccache: 397 of 410 compilations were cache hits (96.8%); cache size 748 of 750 MiB after cleanup
• A 7-zip archive with built NUT for Windows debug binaries is temporarily available: NUT-for-Windows-x86_64-SNAPSHOT-2.8.5.5871-master.7z

[AppVeyorBot]

✅ Build nut 2.8.5.5871-master completed (commit cd949c75df by @jimklimov)