The cause was that the replace code necessarily used a PCRE internal
function to to template expansion.

The fix changes the code to use an SRE internal if SRE is used, and a
PCRE internal if SRE is used; in a way that should work with 1.5.2.

The solution can be sped up tremendously under the assumption that the
choice between sre and pre is not changed during the execution of the
program; especially replace-all will be slow.

But I'll leave that to someone else.