The current test does not detect an addition that would end up between the order and 2^256, and very occasionally makes the unit tests fail.
Do the overflow detection in secp256k1_scalar_add instead, where it can trivially be exact, and use that. Add some bounds checking too.
Can you add a test that would have made the original error reliably reproducable?
You want me to add a test that would catch a non-existing bug in the tests?
Oh. derp. I hadn't read the change yet and I was concerned that I hadn't reproduced it here, and misread "the current test" as in the check for overflow.
ACK.