The default CodeCache size calculation depends on number of Compiler threads available which depends on number of cores on machine.
AOT cache generation could be dome on machine with smaller number of cores and, as result, fewer compiler threads available.
Production run may run on bigger machine and as result CodeCache size could be bigger. Which will invalidate AOT code because short distance jumps/calls may not be valid any more.

I suggest to enforce CodeCache size during production run to the size during training/assembly.