I don't think I ever managed to explain the reasons behind this phenomenon, but I'm sure there is something deep down in the line coding.
What seemed to happen (very roughly) was that, when DLM intervened, the actual speed went down by around 8-10% but the attainable speed went up by the same amount (unless, of course, your package limitation got in the way). Removal of the FEC & interleaving triggered by DLM would undo this - so speed would increase by that same 8-10% value, and attainable would drop back to match.
Looking at MDWS, that may have happened here to some extent, but other factors look to be at play too. From the look of the Errored Seconds graph, I'd say that DLM will likely intervene again soon.