Actually you are right. Samtools manual says it is position, and htseq-count says it is by name. It's confusing.

From htseq-count manual:

For paired-end data, the alignment have to be sorted either by read name or by alignment position. If your data is not sorted, use the samtools sort function of samtools to sort it. Use this option, with name or pos for <order> to indicate how the input data has been sorted. The default is name.

If name is indicated, htseq-count expects all the alignments for the reads of a given read pair to appear in adjacent records in the input data. For pos, this is not expected; rather, read alignments whose mate alignment have not yet been seen are kept in a buffer in memory until the mate is found. While, strictly speaking, the latter will also work with unsorted data, sorting ensures that most alignment mates appear close to each other in the data and hence the buffer is much less likely to overflow.

From samtools specification:

Sort alignments by leftmost coordinates