sets j to -1 in the previous release but to 255 in the upcoming release.

Every enumeration type implicitly provides the following predefined operator, where E is the enum type, and U is the underlying type of E: U operator –(E x, E y); This operator is evaluated exactly as (U)((U)x – (U)y)

That clearly means that the assignment above should have the same semantics as

int j = (byte)((byte)a-(byte)b));

C# defines only four built-in subtraction operators:

int operator –(int x, int y);
uint operator –(uint x, uint y);
long operator –(long x, long y);
ulong operator –(ulong x, ulong y);

There is an implicit conversion from byte to all four types, so we must select the best one.  According to section 7.4.2.3 the int version is the best one (because signed is preferable to unsigned and int goes to long but long does not go to int.)  So what we generate here is the equivalent of:

int j = (byte)((int)(byte)a-(int)(byte)b));

The conversions from E to byte to int will go off without a hitch, and the subtraction will result in an int set to -1.  That then gets cast to byte. What happens when we try to cast a computed-at-runtime integer to a byte? Section 7.5.12 says

For non-constant expressions (expressions that are evaluated at run-time) that are not enclosed by any checked or unchecked operators or statements, the default overflow checking context is unchecked unless external factors (such as compiler switches and execution environment configuration) call for checked evaluation.

Therefore this is an unchecked cast, and -1 goes to 255 as a byte. That then gets converted back to an int during the assignment.