fix: Add exceptions for edge cases for best effort responses.

[stiegerc]

For best effort responses, it is expected that sometimes a second response arrives late. These cases are filtered by checking against the callback when inducting into the replicated state, but the canister could have been stopped or even removed in the mean time in which case we don't even make it to this check. Therefore these cases must be filtered elsewhere since they are expected and not critical errors.

[derlerd-dfinity]

Should we have a test for this case as well?

[stiegerc]

Should we have a test for this case as well?

I can add something for this specifically but you could argue that we don't have any unit tests for best effort anything at the moment. The pragmatic approach would seem to be adding tests for exceptions like this one rather than duplicating everything and then also higher level tests such as state machine tests (which is how I found this bug).

[alin-at-dfinity]

I can add something for this specifically but you could argue that we don't have any unit tests for best effort anything at the moment.

Indeed, we do not have any tests for best-effort messages in Message Routing proper. We do have quite a few of them lower down, in ic_replicated_state. I guess it may make sense to test most (all?) of the potential combinations (request / response, best-effort / guaranteed, canister running / stopping / stopped / deleted / no matching call context / whatever) as protection against regression. It could just be a bunch of for loops and a huge match statement.

Edit: Not necessarily for this PR, just stating my opinion.

[stiegerc]

There is a bit of confusion whether we should use best effort or best-effort. I went with best-effort since that occurs more often by a factor of 3 or so. I can drive-by standardize that in this PR if you want.