> After stepping through the code for that regex that fails, I concluded
> that the condition shouldn't depend on ctx->match_all at that point
> after all.

Tests are passed without this check. But I'm not sure it is not needed. At 
least without this check the code is not equivalent to the code before adding 
support for fullmatch(). So I prefer to left it as is.

> I thought I'd initialised it in all the places it's used.
> 
> I admit that I find the code a little hard to follow at times... :-(

Indeed, it is initialized in Modules/_sre.c, and it is always 0. Perhaps it 
will be more consistent to get rid of the match_all field in the SRE_STATE 
structure and pass it as argument.