The approach I used for our 454 sequence reads was de novo assembly followed mapping the resulting contigs against a reference genome using nucmer. I assemble de novo first to prevent any inversions or repeats in any reference genome making their way into my build.

One question is what effect do inversions present in a reference genome, but not in your genome, have on comparative assembly? I imagine it depends on the weighting comparative assembly software gives to read-overlap or paired-read data versus reference genome alignment. A software suite like AMOS does allow you to combine different genome builds together so it's worth experimenting perhaps?

Both approaches have issues, and it is not possible to say if one is the better. De-novo assembly fails in repetitive regions. Mapping assembly miss segmental insertions and rearrangements. You can try both approaches and see how the results compare. To my knowledge there is no software that does this in one go - which is a shame.

It also depends on the sequencing depth of the newly sequenced genome. If you only have a low coverage (maybe <6x) the de-novo assembly quality will be dubious. I believe combining the two approaches would be the better strategy in this case.

If you have time, try to run a de novo assembly and then try to map it to a Reference.

Segemehl is one tool that helps out in finding insertions and deletions.

http://www.bioinf.uni-leipzig.de/Software/segemehl/