This sounds like what is described here (where the “fix” was to manually adjust the timestamps to something more sensible):
I agree since if it was only the timestamp then the same files should be causing the error on every run, which isn’t the case for me, mr-fibble, etc.
What it does potentially tell us is that there are either multiple causes for that error or (more likely) that there are multiple failure points in the code that aren’t being caught until the metadata error check is reached.
Hopefully once your pull request reaches a canary version we’ll get enough additional detail to help narrow it down.