Note that the cause of discrepancies is Milo loading steps from a side channel (LogDog). This wouldn’t be a problem if Milo loaded steps from buildbucket: bug 850113

There (may?) still be an UI issue if the swarming task continues to run (Thereby keeping buildbucket in a "running" state), but all steps are "complete" (Since steps are sourced from Kitchen/Recipe).  It might not be a Milo issue, but can be still a source of inconsistency.

what would milo do in this case? something like "this build if wrapping up some internal stuff"?

There are various options that can be explored:

1. Milo or Buildbucket detects this and adds a magic step
2. Milo or Buildbucket detects this and changes the status to a magic status ("Running - Swarming cleanup")
3. Milo or Buildbucket detects this and adds an out-of-band message (ie butterbar)

In terms of user expectation, I think #1 makes sense, since as far as the user is concerned "something" in the infrastructure is happening, and "something" generally maps to steps.

since this is a UI-level problem, the solution should also be UI-level, which is Milo, not buildbucket.
I am prefer no magic. If we present this as a legit step, a user may expect to find it in other places, e.g. buildbucket RPC response or BigQuery. It wouldn't be there. We can render something below steps, but not pretending it is a step.

With a test-centric Milo and data from buildbucket, this isn't super important anymore.