I don't know first hand but I recall people stating that it can't work as the method won't be able to build the kmers that are long enough.

Stated for example in a blog post from Homologous: http://www.homolog.us/blogs/2012/10/10/multi-kmer-de-bruijn-graphs/

More relevant overall information on k and other parameters can be found in Titus Brown's blog:

• http://ivory.idyll.org/blog/the-k-parameter.html

In fact all pages tagged as assembly are worth consulting:

• http://ivory.idyll.org/blog/tag/assembly.html

I was curious about this because I use velvet a lot, so I tested it. There is no explicit warning from velveth, but you can tell there were no overlaps found by a couple of ways. First, look at the Roadmaps file. If you choose a k-mer size larger than your read lengths, then the Roadmaps found will be equal to the input sequence number. Another way would be to just run velvetg and take a look at the graph produced. If it runs rather quickly and ends with something like:

...
[155.488308] EMPTY GRAPH
Final graph has 0 nodes and n50 of 0, max 0, total 0, using 0/20198538 reads

Then you have a clear indication no overlaps were found for that hash length. Because read lengths vary, I think that all the sequences would have to be processed in order to warn about these conditions. Though, it would probably be helpful to warn about this after the pre-processing stage or fall back to a hash length shorter than the reads before working on the Roadmaps.