depacketize() decided whether the first OBU element of a payload was a
continuation of the previous packet's OBU by testing
OBUType.fromValue(...) == null. A continuation fragment's leading byte is
raw OBU payload data, so when its type bits happened to encode a valid
OBU type the fragment was parsed as a new OBU - and dropped when that
type was one not carried over RTP (temporal delimiter / tile list /
padding). Large frames fragmented across multiple packets were silently
truncated.

Detect continuation from the aggregation header Z bit
(firstPacketInFrame) on the first element of a payload instead, appending
its bytes to the prior OBU. A genuinely unrecognized header on a
non-continuation element now warns and skips its bytes, which also avoids
a possible NPE in the valid-OBU check when the element list is empty.

Verified by round-tripping a 16054-byte FRAME (14 fragments): previously
reassembled to 8872 bytes, now reconstructs the full 16054 bytes.

packetize(List) used the 2-byte LEB128 length branch for any fragment
>= 127 bytes, but 127 fits in a single LEB128 byte (0x7F). For an exactly
127-byte fragment LEB128.encode returned 0x7F and the 2-byte split wrote
a malformed length of [0x00, 0x7F]; the depacketizer then read a
zero-length fragment followed by a spurious second OBU, splitting one OBU
into two and corrupting the frame.

Use the 2-byte branch only for sizes > 127. Verified: a frame whose final
fragment is exactly 127 bytes now reconstructs to a single OBU of the
full length (previously two elements).

Caught by the byte-for-byte AV1 framing self-check added on the consumer
side.