If the address is calculated correctly for acall and incorrectly for lcall, for the same C code, the conclusions from evaluation of the compiler seem obvious ;)
Check 1) For limitations - lcall is 1 byte longer than acall, maybe that byte tips over the code size limit. 2) For timing conditions, lcall executes one cycle longer than acall, that may be the cycle that breaks sync with outside IO. 3) Exact manuals on the eval version limitations (due to the code size limit, there's no reason to use lcall, you can get everywhere within your eval-allowed address space with the short version. So maybe lcall is simply unsupported by the eval version just because there's no reason to use it?)

If neither of these occur and just that settings change break things, that might be a compiler bug.