I'm pretty sure that someone already did this, i.e. they forked CL-PPCRE for arbitrary sequences. But I can't remember the details right now. You'll probably find a link hidden in the mailing list archives.
I tried quite some time ago to change the RE-compilation into a macro, so that the _whole_ needed code would be visible to the compiler in one compile unit.
That should have enabled quite a few optimizations - starting from matching against a base-string, an (unsigned-byte 8) vector, any other sequence ...
But I didn't get that far ... I'd have had to reimplement most of the existing code, resp. convert everything to return forms. That got a bit messy, too.
So later on I decided that the expected performance-improvements would be reached faster by waiting for 18 months (to get the CPUs catch up) than trying to completely reinvent the wheel here.
Regards,
Phil