Maybe I should know better by now , but rather than sweeping it under the carpet I'd rather clear up the confusion:
I was under the impression that any form of replicate should not be merged. I thought that separate runs of the same sample, is effectively a technical replicate. Also 1 run may work better than another --> producing a batch effect
Some Illumina platforms have lanes on flowcells which are optically distinct but fluidically connected (NextSeq, NovaSeq). So one loads the same pool across the FC but the results may be reported on a per lane basis (unless collapsed into one file per sample during post-processing). So those can't really be considered technical replicates. If you were running the same sample on multiple FC then you could keep track of that by adding read groups if you want to be cautious.
Since you are sampling from the same library data produced by multiple runs should have similar distribution of reads (unless there are severe technical issues with run, e.g. air bubble in lane, unequal length of sequencing). Illumina's guidance has been that sequence produced by their platforms should be considered equivalent. If your definition of better run refers to yield of data then your downstream methods should be accounting for that (e.g. DESeq2) discrepancy.
Technical quality of sequencing long reached a point where there is no need to do technical replicates (and people have not been doing those for 5+ years).