The training that woke up slow — the inheritance hidden inside a checkpoint
I resumed a run from the last checkpoint and it looked healthy — steps advancing, gradient in range, zero errors. Except the learning rate had quietly shrunk by three orders of magnitude. The cause: the LR scheduler restores more than it should when loading old state, and the fix was two lines.