Something else that could be happening is that contigs that are being collapsed into "heterozygous" groups. This would be a particular problem when a genome has a substantial amount of repeated sequence. Digging deep into the supplementary information of the SOAPdenovo2 paper, I've found the following information:
In SOAPdenovo2, heterozygous contig pairs are recognized by utilizing the information of contig depth and the locality of contig. The recognized heterozygous contig pairs should obey the following rules: 1) the similarity between contigs should be high enough, for example, ≥ 95%; 2) the depth of both contigs should be near half of the average depth or all contigs, complying Poisson distribution; 3) the two contigs should be located adjacently in a scaffold and have no relationship to each other inferred by paired-end reads information. The normal contigs neighboring the heterozygous regions, if they exist, could be connected to both of the heterozygous contig pairs (H1 and H2). Only the contig with relatively higher depth in a heterozygous contig pair were kept for scaffolding. The method reduces the influence of genome heterozygosity on final scaffold length. All heterozygous contig pairs were outputted to a file to facilitate further analysis. However, the trade-off of this method is that it might incorrectly remove paralogous contigs. This problem could be relieved by a gap-filling procedure while the removed copy of paralogous contigs would be represented by gaps during scaffolding.