A long integer (one that doesn't fit in the 4-byte encoding) will now be encoded like this:

• initial header byte
• short integer (1-4 bytes) encoding the number of bytes of data and sign
• variable-length number of data bytes (absolute value of the integer) -- all bits are used

For example, a 32-bit integer can now always be encoded using at most 6 bytes (+ type tag).

This is optimized for size efficiency, not performance, since large integers are not expected to be a performance bottleneck. Having an efficient format makes it easier to improve performance in the future, however, without changing the encoding.

The header byte has a few unused bits which could be used to slightly improve efficiency, but I decided that it's not worth the extra complexity.