IMO, this might happen due to the error correction which assembly does - i.e. choosing a base for particular position by the majority of aligned reads - then reads with that base being different may not align back (also indels); another issue could be that reads are getting truncated within the assembly process based on the quality, but i am not an expert in Oases... You can investigate this issue further by lowering the alignment stringency when you map your reads back to assembled contigs.

Any read that spans two contigs won't map to any one contig, although the kmers that make up that read are still "used".

You might ask "why there are multiple contigs at all when there are reads that could connect them?" The answer is that those reads in fact map to multiple contigs because of repetitive regions. Any ambiguities introduced by repeats will force the assembler to keep those contigs separate. So the reads that don't map are likely those that border repeats.