Commit from kdkd fork 2c921c3

matroska/ParserUtils.java:

VINTs that were between 5 and 8 bytes were not being parsed correctly,
causing the parser to get out of sync with the stream.

To reproduce the original bug:

• Take any Matroska element whose size field is encoded with a 7‑byte VINT (legal per EBML). Example: a Cluster element (ID =
  0x1F43B675) with a size of 0x20 bytes encoded as:
  • ID bytes: 1F 43 B6 75
  • Size bytes (7‑byte VINT): 02 00 00 00 00 00 20
    • The leading byte 0x02 is 00000010b, which in EBML means “length = 7 bytes, value bits = 0x000000000020”.
• What the code does:
  • readVINT reads the first size byte fb = 0x02.
  • It sets len = fb >> 4, which is 0x0.
  • Because len is 0, it falls into the final else branch (“8‑byte VINT”), sets mask = MASK_BYTE_8, allocates an 8‑byte array,
    and reads 7 more bytes for this size.
  • The seven bytes it reads are the remaining 6 bytes of the size (00 00 00 00 00 20) plus the first byte of the element
    payload (because there were only 6 size bytes left). So the parser consumes one payload byte as if it were part of the
    size field.
• Consequences:
  • The computed size is wrong (it’s based on 8 bytes instead of 7, and includes one payload byte).
  • The stream position is now off by one byte. The next call to parseTag starts in the middle of the element payload, so
    subsequent IDs/sizes are garbage. This typically leads to decoding failures, incorrect sizes, and eventually EOF/NPE/assert
    failures.

Commit from kdkd fork 0621be8

matroska/ParserUtils.java:

skip() calls InputStream.skip(), which is allowed to return 0, for example
if reading from a named pipe or fifo file that's blocked or trying to
read past the end of the file. If skip isn't working, try reading one
byte to either unstick it or at least advance the file pointer so it
doesn't end up with an infinite loop.

An alternative would be to just give up if input.skip() returns 0 because
I'm not sure recovery is possible.

Commit from kdkd fork 1883e31

matroska/ParserUtils.java:

If inputStream.read has an error, it returns -1. Treat that like an error
instead of rewinding the read pointer by 1.

Commit from kdkd fork ced1aca

matroska/ParserUtils.java:

If parseTag fails it can return null. Intentionally throw a meaningful
exception instead of an NPE.