Running the tests locally does not seem to break anything, but there may be image comparisons tests that fails because of this

Codecov says your addition is never exercised in the tests; so it cannot break the tests 😄 .

What I meant was that we do not have any tests that relies on being able to pass an empty offsets (nor an empty paths). And, as it turned out, no in the additional ones on the CI (as I may not have all optional dependencies installed).

The problem I think is to know if it ever is useful to pass an empty offsets or if some valid input data would pass an empty offsets. Testing the function directly doesn't really make sense just to exercise the code path, but one would probably like to do it from an actual user-level API (even though this technically is one).

If we are not sure that somebody is relying on the current but ambiguous behavior, we should deprecate it instead of immediately erroring out. That gives people time to object. The issue has been standing since 2013, so there is no rush in turning it into an error.

Perhaps what this should really return is a null bbox (see discussion at #17115 (comment))? Although we can always change that later if we manage to resolve #17115 at some point.

Not sure why I put this as draft, but rebased and updated deprecation version.