I noticed that before, but not with RNAseq. It was a kind of artificial set up, where the sequence of interest was exactly 42 bases long, and so we sequenced for just 43 bases (so that there would be 42 bases with post-phasing data). And I noticed that a lot of reads had the last letter trimmed, but only if it was expected to be an A.

I remade the fastqs by turning off adapter trimming in the sample sheet, and that made all the reads 42 bases as they should be.

We were using a pretty old version of bcl2fastq for a while, so I can't vouch for the fact that newer versions will show this same behavior.

Yes, I'have seen that before as well (at least for DNA reads). It's a typical pattern for the read length +1th base, that base is near crap so it should (needs) to be removed anyway. That it's indicates it might be anything but A I'm not sure on (will need to check).

For more (related) info on this topic :

Why 50bp Illumina run produces 51bp long sequencs?

EDIT: I didn't even had to look for for an example: here is the fist random one I took from a dataset generated in 2018

Did you trim the reads? If so, maybe you used a trimming option where adaptors of length 1 are removed (like the -O 1 option in cutadapt). If your adaptor starts with an A, this will remove all A's in the last base of the read.