[rafalmaciag]

Follow-up to #45 (rev-reach CD2: 9de418b9d, fbe88ab06).

m1: the lock directory on a shared box

• When this process creates the directory it is made 1777 (sticky), and each lock file 0666, whatever the umask. The modes are set after creation, which is umask-proof. A path another user owns cannot be chmodded; that is best effort.
• An existing directory is used as it is, which is the same as the foreign-owner case. So a stale same-user 0755 directory from an older version gives other users the named refusal below rather than a silent repair (rev-reach agreed).
• A lock file this user may not open (UnauthorizedAccessException: another user's, made before these modes) is a port to skip.
• A directory in which nothing can be locked (20 denied in a row) is refused, named, with chmod 1777 '<dir>' and "stale lock files of another user" as the causes. It is never an unhandled throw and never an endless search.
• Lock files are never unlinked. Dispose only closes the file, releasing the flock. A lock file removed while held lets the next opener create a new inode and hold it too, so "never tidy lock files while tests run" is written on the class. Under the sticky bit only the owner, the directory's owner or root can remove one.

Found by the row going red first: my own fix made the first 0555 stand-in vacuous. Acquire chmodded my own 0555 directory to 1777, and the refusal row saw no refusal. A foreign-owned directory cannot be chmodded, so the stand-in was not faithful. Hence "chmod only when this process creates it".

m2: both failures printed

A failed start whose cleanup also failed throws an AggregateException: the start failure first, the cleanup failure second. Both appear in ToString; Data, used before, is printed by neither xUnit nor ToString.

The note

The race row asserts HaveCount(2) before its only/unique claims, which hold vacuously over one result.

Rows and mutants (each mutant build-gated)

Row	Mutant → result
PortLockTests: modes 1777/0666	no chmod → red
PortLockTests: an unopenable lock file is skipped	the UnauthorizedAccessException catch removed → red
PortLockTests: a 0555 directory refused, named	the same catch removed → red (an unhandled throw surfaces)
start failure + cleanup failure → both in ToString	back to Data → red
the race row	one child only → red; only HaveCount(2) catches it

Each PortLockTests row uses its own directory and port range; the default directory is shared with every parallel class's stores.

Counts

Full MicroPlumberd.Tests at 6954f7b (with --blame-hang-timeout 4m): 200/207 (17 m 43 s). No hang and no host crash; 0 containers left. The 7 failures:

• The long-standing 6: GrpcApiTests ×3, ProcessManagerSourceGenerationTests.CommandTypes, SubscriptionRunnerStateTests ×2.
• Cron.CronTests.AnEnabledIntervalJob_IsScheduledByTheEngine_AndRuns: not this diff (cron, not Testing). ScheduledTotal was 0 for the full 60 s budget: never scheduled, not slow. A separate finding, reported to the lead:
  • the cron server subscribes JobExecutionProcessor from the end (FromRelativeStreamPosition.End), and its input loop waits 15 s;
  • CronTests: wait for the condition (bounded), this test's own evidence; the misnamed row renamed, a real manual-trigger row behind a fence #44 dropped a 1 s wait on the claim that the processor "folds the job's events whenever it subscribes", which this shows is false as stated.
  • Overlap: a docker-free unit class of another engineer ran beside this suite.

The head (cab1959) adds only comment and message text on top of 6954f7b; built, and PortLockTests 3/3.

🤖 Generated with Claude Code