Splitting into phrases before sending to the toplevel is sometimes incomplete
Typically if you edit in the middle then quickly press Ctrl-e, sometimes it seems that the phrase split is not complete, and the end of the buffer is sent in one blob to the toplevel. It's not a breaking bug, but doesn't look good.
The issue appears to be that we use the tokenisation information from ocp-indent, but obtained through Ace; and Ace does it in the background. I couldn't find a way in the Ace API to wait for it to be complete.