From #826 (comment)
the handling of --color-words is not quite what I had expected in v0.10.3 either. I suppose it makes sense given that delta is only passing along the raw values, but for a file with multiple hunks, I'm not getting either line numbers from delta or the hunk headers from the raw git.
It looks like the implementation in #807 isn't quite right: it's causing the hunk lines to be emitted raw by not handling them. But I think that what is needed is more complex: to transition into one of the hunk line states of the state machine, emit raw instead of the normal emission, and report the line as having been handled. I think that this is because we are not transitioning the state correctly