When you are processing paired-end data, singletons can arise in the merged output. So you’ve done nothing wrong. However, such singletons should be rare (and 15 total qualifies as rare in any normal sized study). In general I’d recommend just filtering them out as you did.
Singleton sequences can arise even though no singletons are called in the forward or reverse reads individually. When those F/R reads are merged, sometimes a read-pair can be corrected to a unique pair of F/R denoised sequences that are mergable. This is more likely in longer (less-overlapping) amplicons, but should be rare regardless. In my not-extensive testing, these merged singletons are as often an artefact as they are a real biological variant, hence the recommendation to filter them out if they exist in your data.