The streams return end-of-input if they read no samples, but know that
they are fully buffered to at least the clipped end time. This helps to
detect the end of stream even if there are no new buffers after the end
of the clip (e.g. for sparse metadata tracks).
The race condition occurs because the buffered position is evaluated
after reading the sample. So between reading "no sample" and checking
the buffered position, the source may have loaded arbitrary amounts
of data. This may lead to a situation where the source has not read
all samples, reads NOTHING_READ (because the queue is empty) and then
immediately returns end-of-stream (because the buffered position
jumped forward), causing all remaining samples in the stream to be
skipped. This can fixed by moving the buffered position check to
before reading the sample, so that it never exceeds the buffered
position at the time of reading "no sample".
#minor-release
PiperOrigin-RevId: 548646464
c64d9fd6da