tests_complex::myanmar_syllables_are_single_clusters and
myanmar_in_a_latin_font_falls_back_to_a_regular_myanmar_face failed about
1 run in 4 of the full layout suite (cluster starts [0, 3, 6, 12, 18], "one
face for the word") but passed alone.
Root cause: a race in FontDb's system font fallback, not the tests. Tests
share FontDb::global(), and layout asks for a fallback per character from
several threads. A search marks a preferred family ("Padauk" for Myanmar) as
tried before loading it; a second search racing it skipped the family as
already tried, fell through to the coverage index, and loaded the first
cataloged file in path order that maps the character (Padauk-Bold.ttf on
its own). Characters of one syllable then fell back to different faces, so
the shaper split the syllable and the word's faces differed. Separately, a
search's answer was cached under the face count read after the search, so
an answer chosen before another thread's load could be cached as current.
Fix (crates/fonts):
- System fallback searches run one at a time (a `searching` lock). A search
that waited re-checks the loaded faces first, so it uses the family the
previous search loaded instead of loading another file.
- The fallback cache keys an answer by the face count it was chosen from,
and doesn't cache an answer from before a concurrent load.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
62bc097fab