For your dual-alignment step, I would consider aligning once against a concatenated H. sapiens / E. coli reference rather than doing two separate alignment runs. That is often “cleaner” (i.e., in terms of the assignment of what alignment to what reference) and it’s more efficient computationally, since it avoids the running / management of two independent alignment passes. Still, separate alignment like this is common, and in this particular setup (human and E. coli), there tends not to be much alignment ambiguity (in my anecdotal experience).

Also, you do not need to write SAM intermediates and then convert/sort afterwards; you can do data streaming from the aligner into Samtools, e.g., bowtie2 ${args_bt2} | samtools view ${args_view} | samtools sort ${args_sort}. (You can pipe directly to sort if not filtering with view.)

Btw, if you haven’t already, it may be useful to review the Henikoff Lab’s CUT&* processing materials at protocols.io. (I think much of the CUT&Tag data processing tutorial material is applicable here.)

In general, be careful regarding the assumption that one peak-caller setup is ideal for all of your CUT&RUN data. SEACR is commonly used, but the optimal parameters and/or caller itself may need to change depending on signal characteristics (e.g., sparse, broad, TF-like, etc.)