I think it has to do with the CPU usage of the error correction. You're right that it either works or it does not, but there is a small bit between working and not working where there is a "can be made to work if the errors can be corrected" region of the signal to noise ratio.

If there is enough CPU after decoding valid frames to error correct invalid frames, and the error correction data is intact enough, AND there is enough CPU to fix the error and decode the frame before the next frame comes in, I imagine that it would succeed. If any of those are not true, it will fail, and require more processing on the subsequent frames until a new I-frame is received.

This is half estimated guess; I don't know the nuts and bolts of how ATSC works, really.