This is probably a problem that was introduced by the contextification of
patch.diff back in changeset dccb83241dd0.

In addition to the (restricted) working dir walk required to figure out what has
changed, we've started doing a full walk to generate a full manifest for the
working dir.

Can you do a quick test with hg revisions dccb83241dd0 (which should be bad) and
b4eaa68dea1b (which should be ok) to confirm this?