At the moment, the write_vcf() function simply outputs the 0th allele from variant.alleles (i.e. the ancestral state) as the REF. If we have defined a reference sequence, then presumably we should be using the actual REF instead, if it exists.
Either way, the behaviour should be documented.