Could you please check the CI failure? It looks to me like it is unrelated to my changes (failure is while installing Bender). Otherwise, please let me know if you think it's due to my changes.

I'm not sure this PR fixes a symptom or the cause of the actual issue. Could you have a look at #389 to see if that helps fix your issue?

I am not sure if #389 fixes my issue. Also, I have unfortunately not saved that state of the design, so I cannot reproduce the issue anymore.

My issue was specific to the case where there is only a single bank, and the full strobe is zero. I know this is kind of weird, but this is what was received in my case (for some reason).
In that case, resp_valid and dead_response are both single-bit signals. If dead_response is 1, rvalid_o can remain high even if there is no request currently in flight. (This cannot happen unless the full strobe is zero, which make this such an edge case.)

I am not well-versed enough in your codebase to quickly write a minimum failing example.
I think it should be reproducible by inputting a large number of writes with a fully zero strobe, and then stop. You should observe spurious B responses in case.

Looking at it differently: How do you ensure that there are never more responses (on mem_stream_to_banks_detailed::rvalid_o) than there are requests (on mem_stream_to_banks_detailed::req_i)?

I don't think #389 actually fixes that issue, but is a different (unrelated) issue. As far as I can tell, the issue fixed by #389 can only occur if there is back-pressure from the memory bank, which was not the case for me (I tied mem_gnt_i high).

Ah, I see your point. Thanks for the additional explanation!