zccache: Add version 1.13.21

[luhenry]

• Package: zccache
• Version: 1.13.21
• Source: https://github.com/zackees/zccache
• Docs: https://pypi.org/project/zccache/

Compiles the Rust zccache/zccache-fp binaries and three abi3 pyo3 extensions into one py3-none wheel. Upstream publishes no riscv64 wheel.

Mirrors the build/build-wheels/test-wheels jobs of upstream's release-auto.yml and the native-pytest job of python-tests.yml.

Differs from upstream

• Native cargo build in the manylinux image, not soldr/zig cross-compilation - upstream has no riscv64 lane
• dnf install protobuf-compiler - protoc-bin-vendored has no riscv64 binary
• Calls upstream's build_wheel() for one platform - its build-wheels command requires all six

Matrix: tests on cp312/cp313/cp314; the abi3 extensions cannot load on free-threaded cp314t.

Testing

• same as upstream, run against the installed wheel

License: Wheel ships zccache itself (AGPL-3.0-only); upstream's wheels carry no licence text, so 0003 adds it.

Patches

• 0001-protocol-fall-back-to-prost-build-s-own-protoc-looku.patch - Inappropriate [protoc-bin-vendored has no riscv64 binary]. build.rs panics without it. riscv64-only.
• 0002-deps-bump-crash-handler-to-0.8-for-riscv64-support.patch - To upstream. crash-context 0.6 has no riscv64 ucontext_t, so the build fails. Reproduces cross-compiling from x86.
• 0003-release_workflow-ship-LICENSE-in-every-wheel-s-dist-.patch - To upstream. Wheels carry no LICENSE without it. Reproduces everywhere.
• 0004-watcher-notify-main-thread-once-per-interrupt.patch - Backport [zackees/zccache@64b7e7c, released in 1.14.5]. A watcher predicate KeyboardInterrupt reached the main thread outside the tests' assertRaises and aborted the pytest session (exit 2). Reproduces on x86_64. Also applied in the test job, which checks the upstream tree out again.

[luhenry]

The failed check_commit_messages run (job 115916157215) is the known pull_request "opened" webhook stale-body-snapshot case, same as #3221-#3228: it evaluated the PR body before create_pull_request's auto-injected footer was stripped via update_pull_request. The edited-triggered re-run already passed cleanly (all three checks green). No fix needed — riscv64 build now queued.

[luhenry]

Test zccache 1.13.21 cp313 / cp314 (run 38582612676, jobs 116324636431 / 116324636408): a real upstream test defect, not an infra flake. Fixed in fe036c4 by backporting upstream's fix.

What the full logs show. Both legs fail the same way, in crates/zccache-watcher/python/tests. The python/tests suite passes on both (64 passed, 1 skipped).

cp314: 1.02s call python/tests/test_compat.py::WatcherTests::test_context_manager_and_resume_reset_state_cleanly
       ...
       !!!!!!!!!!!!!!!!!!!!!!!!!!!!!! KeyboardInterrupt !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
       <frozen codecs>:189: KeyboardInterrupt
       11 passed in 8.52s
cp313: ...
       !!!!!!!!!!!!!!!!!!!!!!!!!!!!!! KeyboardInterrupt !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
       /opt/python/cp313-cp313/lib/python3.13/pathlib/_abc.py:651: KeyboardInterrupt
       12 passed in 9.91s
##[error]Process completed with exit code 2.

The banner comes before N passed. pytest prints it in pytest_sessionfinish only when the session's exit status is INTERRUPTED, so the KeyboardInterrupt was raised inside the run-test loop, not in teardown after the session. The suite has 13 tests, so cp314 was cut off during test #12 and cp313 during test #13. Collection is alphabetical, so those are test_keyboard_interrupt_stop_logs_once_per_watcher (#12) and test_poll_predicate_keyboard_interrupt_propagates (#13). Both crash sites are inside Path.write_text(): 3.13 pathlib/_abc.py:651 is with self.open(mode='w', ...), and 3.14 codecs:189 is IncrementalEncoder.__init__, which open(mode='w') reaches. The podman orphan podman pause reaped at cleanup is the normal rootless-podman pause process and has nothing to do with this. No signal is involved: the interrupt is a _thread.interrupt_main() call from inside the process.

Root cause (upstream 1.13.21). Both tests use a notification_predicate that raises KeyboardInterrupt on the watcher's dispatch thread:

• FileWatcher._predicate_allows() catches it, calls handle_keyboard_interrupt(ke) (a real _thread.interrupt_main()), then re-raises. _dispatch_loop()'s except KeyboardInterrupt calls handle_keyboard_interrupt again, so one interrupt can reach the main thread twice.
• The tests don't mock interrupt_main, and they call file_path.write_text("b\n") before entering with self.assertRaises(KeyboardInterrupt):. NativeWatcher.poll_batch (crates/zccache-watcher-py/src/lib.rs) blocks in poll_timeout for up to poll_interval while holding the GIL, so the dispatch thread can starve the main thread. The main thread is then still inside write_text when the interrupt lands. That is outside assertRaises, so pytest treats it as a user Ctrl-C and aborts the whole session with exit code 2.

It isn't specific to riscv64 or podman. With the published x86_64 zccache==1.13.21 wheel and the upstream 1.13.21 suite, 23 of 50 local runs fail with exactly these two signatures (11 passed interrupted at test_keyboard_interrupt_behavior.py:83, 12 passed interrupted at :57). Tracing interrupt_main confirmed both calls come from the same watcher's dispatch thread (_predicate_allows:202 and _dispatch_loop:231).

Fix. Upstream already fixed this in zackees/zccache@64b7e7c, "fix(watcher): notify main thread once per interrupt (#1616)", which fixes zackees/zccache#1615 and first shipped in 1.14.5. It removes the duplicate notification from _predicate_allows in both copies of the watcher module. It also keeps interrupt_main mocked until the watcher has stopped and asserts it was called exactly once. Commit fe036c4:

• adds patches/zccache/1.13.21/0004-watcher-notify-main-thread-once-per-interrupt.patch (Upstream-Status: Backport, without upstream's tasks/todo.md hunk); 0001-0004 apply cleanly;
• adds the same python-wheels checkout and git apply step to test_wheel that build_wheel has. The test job checks the upstream tree out again, so without this the test-side half of the patch never applies (gotcha 501).

With the patch, x86_64 passes 60 of 60 runs (30 on cp313, 30 on cp314; 14 tests, including the new test_poll_predicate_keyboard_interrupt_notifies_main_once).

Next steps. The push re-triggers the workflow through pull_request: paths, which rebuilds the wheel with the library fix and reruns cp312/cp313/cp314. No manual re-run is needed. The PR stays a draft and stays watched.