Fix two flaky tests caused by shared fixture mutation

[perryqh]

Two tests fail intermittently on main today, from the same root cause. Independent of any other work in flight.

test	failure rate before
gitignore_test::test_respect_gitignore_can_be_disabled	2 in 20 idle · 3 in 6 under load
create_test::test_create_already_exists	~1 in 10

Cause

common::teardown() globs tests/fixtures/*/tmp/cache/packwerk and removes the cache for every fixture, not just the one the calling test used:

pub fn teardown() {
    glob::glob("tests/fixtures/*/tmp/cache/packwerk")

Tests within a binary run on parallel threads, so one test's teardown deletes state another test is actively using. create_test compounds it: delete_foobar/delete_foobaz remove whole pack directories from fixtures a sibling test is running against.

The specific window for the gitignore failure: pks writes a cache entry as create_dir_all(parent) then File::create. Losing the parent between those two calls is why it surfaces as EINVAL, not the ENOENT you would expect:

Error: Failed to check files: Failed to create cache file
".../app_with_gitignore_disabled/tmp/cache/packwerk/zeitwerk/<hash>":
Invalid argument (os error 22)

Two details worth knowing:

• The cache cannot be disabled to dodge this. Neither fixture sets a cache: key, yet a cache file is still written — the constant-resolver cache is not governed by that setting.
• Scoping teardown() to a single fixture would not have been enough. Four of the gitignore tests run against the same fixture and would still race each other. That is why this serializes instead of narrowing the glob.

Fix

#[serial] on the tests that shell out to pks — six in gitignore_test.rs, four in create_test.rs.

This follows existing convention rather than adding a mechanism. gitignore_test.rs already imported serial_test and already marked test_respects_global_gitignore that way; add_dependency_test.rs, add_constant_dependencies.rs, and update_test.rs all do the same. serial_test has been a dev-dependency all along, commented "Run specific tests in serial".

Each file gets a comment explaining why, including the observed error, so the attributes do not get stripped later as redundant.

Verification

	result
gitignore_test, 25 consecutive runs	25 / 0
create_test, 20 consecutive runs	20 / 0
full suite, 6 consecutive runs	258 passing / 0 failing each

cargo fmt --all -- --check and cargo clippy --all-targets --all-features clean.

The flakes were costing more than noise

cargo test stops at the first failing target, so a red gitignore_test truncated the run: 240 tests attempted instead of 258 — roughly 18 tests in later binaries silently never executed. A flaky test early in the sequence was quietly reducing coverage on exactly the runs where you would most want it.

Still latent, not fixed here

The real defect is that teardown(), delete_foobar*, and set_up_fixtures() all mutate global fixture state while their callers run in parallel. Serializing these two files removes the two failures I could reproduce; it does not remove the wire. Others calling these helpers with no serialization at all:

file	serialized	mutating helpers used
check_test.rs	0 / 26	teardown
check_unused_dependencies.rs	0 / 4	set_up_fixtures
folder_privacy_test.rs, visibility_test.rs, layer_violations_test.rs, validate_test.rs, …	0 / n	teardown

I ran check_test and check_unused_dependencies 12× each and they stayed green, so those look latent rather than live — but check_unused_dependencies rewriting shared package.yml files with four unserialized tests is the same wire, waiting.

Fixing it properly means narrowing teardown() to take the fixture it should clean, which touches every call site across ~15 files. That is a refactor, not a flake fix, so it is out of scope here.

🤖 Generated with Claude Code

[perryqh]

Superseded by #57, which fixes the same two flakes by giving each test its own fixture copy rather than serializing access to a shared one.

Isolation removes the shared state instead of coordinating access to it, so a future test cannot reintroduce the bug by forgetting an attribute. It also keeps the tests parallel (0.72s vs 0.83s), drops the #[serial] count from 10 to 1, and stops the suite leaving modified fixtures in the working tree.

Closing in favour of that. Diagnosis and evidence carry over unchanged.