The problem with the "old" way of building modules, is that it assumes that conditions that were true when Pike was compiled is still true as it causes headers generated by configure tests at compiletime to be used.
What Per is trying to point out, is that this doesn't work in every situation and isn't a good way of doing things. It's simply wrong to rely on tests done on another system. I wouldn't mind a cleanup operation on the whole build system, but that's a major task and I'm not sure anyone is willing to spend enough time on it.
