No, the problem is that the type completeness check has no memory across commits. So for each PR it needs to re-compute the type completeness of the head. so you have

1. commit A, merged after testing with pyright 1.1.372, type score 1.0
2. PR, comparing against commit A, recomputes with pyright 1.1.373, type score 0.976

If we wanted to catch these hases, we'd have to have something like codecov, i.e. a database for the type score per commit.
the upside of the current behavior is that a pyright update doesn't mark your PR as invalid even though your PR doesn't in itself decrease the type completeness …