The prior algorithm summed up all Opus frame's number of samples and
calculated the expected timestamp from there. That expected timestamp
was compared to the Ogg page's granule position, and the discard
padding was calculated from that.
This lead to problems in several cases:
1. When the first packet's granule position was bigger than the number
of samples in the Opus frame, the specs say that the first sample's
timestamp is positive in that case. mkvmerge was wrongfully using 0
as the timestamp and thought discard padding was needed in such a
case.
2. When the first packet's granule position was smaller than the
number of samples in the Opus frame (meaning the first sample's
timestamp was negative), a technically invalid bitstream, mkvmerge
was wrongfully forcing the first packet's timestamp up to 0 and
inserting discard padding elements all over the place as its
internal calculations got confused.
The new algorithm bases the calculation on the last known granule
position special-casing for certain Ogg packet numbers:
1. For the first Ogg page the timestamps of all packets is calculated
by subtracting each packet's number of samples from the page's
granule position in reverse order. This may result in the first
packet's timestamp being positive, which mkvmerge now preserves.
mkvmerge won't calculate a discard padding for this page.
2. For all other pages each packet's timestamp is calculated by adding
each packet's number of samples to the previous Ogg page's end
timestamp as given by its granule position.
mkvmerge then calculates the discard padding for the last packet of
the page by comparing the calculated end timestamp of that packet
with the Ogg page's end timestamp as given by the granule
position.
As a workaround for certain rounding errors encountered in the
wild, a discard padding of one sample will be treated as if no
discard padding was present for all but the very last packet as
indicated by the "end of stream" bit.
Fixes#2280.