Try the following:

(setq unread-post-input-method-events '(?a ?b ?c))
(sit-for 0.1)

 (Note: This is not a test for the test, this is an actual problem
  with input methods and sit-for)

It will result in the unexpected behavior that events are processed as
« bca ».  IIUC, this is how it happens :

- Sit-for takes the `a' with its `read-event'
- Sit-for stores it back in `unread-command-events' _and it seems normal_:
  I think that users are expected to use that var, especially because
  the docstring says this is the FIRST input var processed.
- read_char is made three times: for the two first times it will take
  events from post-input, then for the third, from
  unread-command-events, because post-input is processed BEFORE
  unread-command-events. 

I think Handa's changes 2006-08-21 (fixing the docstring of post-input
that indicated that it was processed AFTER unread-command-events,
whilst it was the contrary) wasn't the good one.

I propose to revert the docstring and make the code respect both this
one and unread-command-events' one.