In my experience as ex-academic, the above is all too true in too many computational fields (and many sub-fields of CS). So I ask what scientific value (or any kind of value, for that matter) is being generated by grinding out publications with results that won't be repeatable?
Lots of journals (at least in my biological sciences field) require you to submit your code along with your publication so that a) the editor can repeat your results to accept your publication and b) for other scientists to repeat your stuff. Biostatistics AFAIK has an editor responsible solely for reproduction of results.
Personally saying lots of my code is repeatedly re-used by colleagues who can't program, for example scripts that parse BLAST-output and get the best result for each query etc.
If you want a newer example, here's a very recent paper which looks at the results of a couple of RNA-Seq-papers and after being able to reproduce only 12% of the results, concludes that they all used the wrong assumptions in their statistics: http://www.plosgenetics.org/article/info%3Adoi%2F10.1371%2Fj...
Which shows that the editor really needs to know what's going on and why in each paper - I don't think every editor can know everything about his or her field.