sdf_text_emits_vector_ops_at_the_run_position laid glyphs into the process-
global text atlas, released the lock, then re-read the atlas in emit_text.
Another test growing or resetting the atlas in between shifted the uv keys,
so the lookup missed and the assert saw zero polylines — ~1 in 5 parallel
runs. Its own comment already noted it couldn't catch the one thing worth
catching (a mirrored corner mapping); the size-based tests and text_gpu's
vertex-order tests cover what remains. Not worth a global test mutex.
00639db678