Thanks Raymond for the interesting use-case.

The original design of mode() was support only the basic form taught in 
secondary schools, namely a single unique mode for categorical data or 
discrete numerical data.

I think it is time to consider a richer interface to support more uses, 
such as estimating the mode of continuous numerical data, and the 
multi-mode case you bring up.

One reason I've been hesitant is that deciding what is the right 
behaviour is quite hard (or at least it has been for me). I think there 
are a few cases to consider:

- the existing behaviour (which may not be very useful?) which is to
  raise an exception unless the mode is unique; 

- would it be more useful to return an out-of-band value like a NAN 
  or None?

- the multi-mode case where you want some arbitrary(?) mode, possibly
  the left-most (smallest) for numeric data;

- the multi-mode case where you want all the modes.

I like Francis' suggestion to use an enum to select the behavioural, er, 
mode (pun intended). What do you think?