I gave it a brief spin and it's pretty nasty, particularly if you look at the internals. If you're doing serious numerical work, I'd go for Octave -- it links in to your platform's BLAS and LAPACK, along with a whole lot of other third party math libs. It also has a pretty clean interface for hooking in your own C++ extensions (see here).

You're right that Octave sucks for text manipulation, but I do alright using it for bioinformatics. The best way I've found is to preprocess the input in Perl, then have your Perl script write out its results in Octave's text save-file format (which is pretty self-explanatory). Inline::Octave doesn't seem to be worth the trouble -- it's almost as easy to just use a pipe.