The Roto/Object Matte segment store is one per process (roto::set_seg_store),
and every Session::configure_disk_cache that opens a cache takes it over. The
roto and disk-cache tests already serialise on hold_roto_cache_test_lock(), but
many other engine tests build a Session with a settings store, which opens the
disk cache at the platform folder and takes the store; once that session drops,
the weak store is dead and object_matte_mattes_persist_in_the_disk_cache sees
no disk writes. It failed 3 of 3 in the full parallel suite here.
configure_disk_cache now takes that lock under cfg(test) before opening a
cache, so every such test runs apart from the roto tests. Re-entry on the same
thread is a no-op. Product code is unchanged: only the app's one session ever
opens a disk cache, so users cannot hit the race.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
iannitramandClaude Opus 5.5
c1d411b249