[nmbro]

Problem

In the AppVeyor step scripts, cmd.exe only treats double quotes as
quoting, not the single quotes around bash -lc '...'. So an unquoted
|| (or |, &&, <, >) in there is cmd's own operator:

• bash gets only the text before it;
• cmd runs the rest as a cmd command if bash failed, or skips it.

I reproduced this with a minimal .bat. On the 02-install.bat ccache
line it means that ccache -o max_size/compression/sloppiness never
ran: the job logs show max_size 5.0 GiB (the default) in 65 of the 70
jobs I sampled. The only exceptions were 5 builds from #3784, which set
CCACHE_MAXSIZE in the environment.

That is what kept resetting the cache. Over 70 job logs (pushes to
master and PR builds, 2026-10-15 to 2026-10-20):

• ccache grew by ~155 MB per master push (each one changes
  PACKAGE_VERSION in config.h, see build: keep full git version out of config.h (ccache-friendly) #3759) and kept growing with
  PR builds, past 1 GB.
• AppVeyor then refuses to save it ("Unable to save cache": the per-item
  limit is 1,048,576,000 bytes compressed). That happened 6 times.
  The old entry is gone too, so each time the next build started with
  no ccache at all (4% hits, ~160 MB), and the cycle restarted.
• Effect on the 04-build_script step: ~15-20 min with a warm cache in
  PR builds, ~33-37 min cold. Cold jobs take ~50-54 min of the 60-min
  limit, and 8 failed jobs stopped at ~51 min, which is the CI deadline.

#3786 fixes the 02-install.bat line. This PR contains the same line
as my suggestion there
(#3786 (comment)).
If #3786 lands with that suggestion, the line drops out of this diff on
rebase (a simulated 3-way merge gave no conflicts). Otherwise, I'll
rebase and keep this version.

Changes

• 02-install.bat (= the Fix build/test errors after recent merges #3786 suggestion), 05-after_build.bat,
  07-test_script.bat, 08-after_test.bat: replace a || b with
  if ! a ; then b ; fi. The pitfall is documented in appveyor.yml.
  A static check finds no |, ||, &, &&, < or > outside
  double quotes in the bash -lc lines any more.
• Enforce the ccache size limit before AppVeyor saves the cache.
  ccache can overshoot max_size between its automatic cleanups (I saw
  ~120%), which is too close to the per-item limit.
  • 04-build_script.bat runs ccache -c whatever the build result,
    and keeps the build's exit code. With APPVEYOR_SAVE_CACHE_ON_ERROR,
    the cache of a failed build is saved right after the failing step,
    before on_failure.
  • The new on_success step 12-on_success.ps1 runs it again for
    passed builds, after the tests and the bundle have also compiled.
• 03-before_build.bat: ccache -z, so the statistics count this build
  only. They were cumulative over the cache lifetime, so the hit rates
  in the logs were misleading.
• 12-on_success.ps1 posts this build's ccache hits and cache size as
  an AppVeyor job message. The GitHub notification template already
  includes job messages for passed builds. It uses the invariant
  culture, and it is best-effort (a failure there does not fail the
  build). Format:
  ccache: <hits> of <total> compilations were cache hits (<pct>%); cache size <n> of <max> MiB after cleanup
• appveyor.yml init:
  • PR builds set APPVEYOR_CACHE_SKIP_SAVE, so they only restore the
    cache. There is one cache per project, and each save replaces it:
    43 of the 46 sampled PR builds saved it, each one evicting master's
    objects and uploading ~0.7-1 GB again.
  • APPVEYOR_SAVE_CACHE_ON_ERROR moves from environment: to init,
    for master pushes only: a failed master build still saves a
    partly warmed-up cache, and other branch builds save only when they
    pass.
  • This is the pattern from https://www.appveyor.com/docs/build-cache/,
    also used by e.g. qreal, EventGhost and simde.
• 07-test_script.bat:
  • dumps the test logs from the same bash call;
  • keeps the real exit code (e.g. 124 from timeout) instead of 1;
  • runs make check with -j 2, the AppVeyor VM size.
• 08-after_test.bat:
  • drops a line which, due to the same || problem, only ever removed
    a non-existent symlink. Had it worked, the symlink would also have
    matched NUT* for 7z and duplicated the archive contents.
  • compresses the artifact with 7z -mx=3 -mmt=on; the default level
    took ~3.8 min of the job time.

Suggestion for the maintainers

The init line is a convenience only, because a PR can change
appveyor.yml. To really keep PR builds (including from forks) from
writing to the cache that master builds use, please untick "Save build
cache in Pull Request builds" in the AppVeyor project settings. Today,
a PR could plant ccache objects which a later master build "hits" and
links into its artifacts. This setting is not available in YAML.

Timings (local, MSYS2 on an 8-core machine, limited to 2 jobs like the AppVeyor VM)

7z on a 679 MB bundle, two threads:

7z level	time	size
default (-mx=5)	98 s	52 MB
-mx=3	10 s	76 MB
-mx=1	12 s	199 MB (switches to a weaker method)

make check:

	time	result
serial	279 s	all 16 tests passed
-j 2	184-207 s	all 16 tests passed in each of 7 serial/parallel runs

Testing

• Ran these lines through cmd.exe in MSYS2 with a temporary
  CCACHE_DIR and CCACHE_MAXSIZE=750MiB as in appveyor.yml:
  • the 02 line: reports max_size=750.0 MiB compression=true;
  • the 03 stats reset: "Statistics zeroed";
  • the 04 line: in a dir without ci_build.sh, the cleanup ran and
    exit code 127 was kept.
• 12-on_success.ps1 and the init logic: tested with stub AppVeyor
  cmdlets and variables, for a PR build, a master push and another
  branch.
• appveyor.yml parses, all files are ASCII, and the .bat files keep
  CRLF.
• NEWS bullet: passes the spellcheck.
• Not tested: an actual AppVeyor run, which this PR provides.
  • A PR build log should show no "Updating build cache..." line.
  • A failed master build should still show one.
  • Passed builds should show the ccache line in the GitHub
    notification.

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
and adjusted the changes (e.g. merged the on_success cleanup into the
PowerShell step), 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 d54560c is temporarily available: NUT-tarballs-PR-3790.zip.

[AppVeyorBot]

✅ Build nut 2.8.5.5827-master completed (commit d7fe499287 by @nmbro)

• artifacts
• ccache: 403 of 409 compilations were cache hits (98.5%); cache size 749 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-2.8.5.5827-master.7z

[jimklimov]

please untick "Save build cache in Pull Request builds" in the AppVeyor project settings

Done, thanks. I wonder if this also saves some time that can be given to PR job execution, so it is less limited (or not constrained at all) - meaning that 51-minute cut-off introduced recently to ensure that caches are saved consistently. I guess PRs can now use the full hour?

I also see many of those timed-out jobs clocking 51 minutes and a few seconds, so cache-saving on top of the scripted 51 minutes take under a minute. Maybe that allowance (now only on branch builds) can be bumped to 58 minutes or so? WDYT?

UPDATE: These proposals may go into a future PR, I am eager to merge the currently outstanding recipe fixes - so all other waiting PRs can benefit from the stable baseline.

[nmbro]

I will make a new branch for the timeout changes so that you can merge this one as is