Codecov reports wrong coverage diffs #3136
Description
Couldn't find a discussion on this problem. Sorry if it's a duplicate.
Expected behavior
Codecov CI should report coverage diffs that reflect the impact of modifications a new tests in a given PR.
Actual behavior
Nonsensical values are reported, which sometimes cause the build to fail due to a perceived drop in the overall project's coverage. An example is the recent PR #3135, where diff lines had 100% coverage but overall the project coverage dropped by 0.45%.
It is mathematically possible for a 100% covered diff to cause a drop if the overall number of project lines decreases; but in the case of #3135 the total number of lines also increased, so there's definitely something off.
I didn't explore much the API of codecov, but something that may be a hint is that for each PR multiple builds are reporting coverage values, and several with the same identifier (see the report for the latest commit into develop). Could that be a source of confusion when calculating diffs?