Two leak paths: (a) AtomicFile.write renames the temp file onto the probe
path and THEN fsyncs the containing directory — if that last fsync throws,
the probe file already exists on disk but the old code returned false
without ever attempting removal; (b) if the explicit removeItem call itself
threw, there was no retry, so a transient File Provider removal failure left
the file behind permanently.
Fix: an unconditional `defer` now checks fileExists and retries removeItem
regardless of which branch returned early. The function only reports true
when the explicit write, fsync (inside AtomicFile.write), AND removal all
succeeded AND the file is confirmed gone afterward.
New test: a FileManager subclass whose removeItem(at:) throws on its first
call (AtomicFile.write itself never touches this injected FileManager — it
uses raw Darwin/POSIX calls, not FileManager, so this only intercepts the
explicit removal + the defer's retry) asserts the function returns false AND
no probe file remains — verified this actually needs the defer by
temporarily removing it and confirming the same test then fails with a
leftover ".redline-probe-<uuid>" file (see this branch's history for the
discarded revert). The directory-fsync failure path has no injectable seam
(raw fsync(2) on an already-open fd, not parameterized by any FileManager or
other substitutable dependency, and not reproducible via chmod or other
standard test techniques) — documented in the test rather than simulated.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012ZiTXPbPCSjzPVsfoweAbp
Claude Fable 5
e1f569cc3e