For sync-sample-only formats, we have an optimization to drop all buffers
with less than the start time when writing them to the queue.
For the same formats, if we set a new start time (=seek), we only seek
to the buffer at or before the start time. This means the first sample
in the queue is different depending on whether we seek to a start time
or set a start time and then write samples. This is inconsistent and
effectively means the first sample depends on a race condition between
the Loader thread (writing samples) and the playback thread (attempting
an initial seek in the already loaded samples).
The effect of this inconsistency is that we have to decode one sample
we don't need (and could have skipped) and that some tests become flaky
if the test setup runs into the mentioned race condition.
The fix is to change the SampleQueue seek method to also seek to
a sample at or after the specified time, to align the behavior to the
case where we write the same samples to an empty queue.
The change also clarifies the Javadoc of
MimeTypes.allSamplesAreSyncSamples to note that this should really only
return true if the samples have no "duration" that matters. Otherwise,
we could reasonably return true for most subtitle formats although it
would break subtitle display because we'd remove samples that start
before the seek time.
PiperOrigin-RevId: 547189941
5c8b142174