Guidelines for testing the ECN implementation in a router:

These are the things that I would test for if I was to test
an ECN implementation in a router.

This uses the following terminology for the four ECN codepoints in the
ECN field:
           0     0         Not-ECT
           0     1         ECT(1)
           1     0         ECT(0)
           1     1         CE

* Verify that CE is never set in an outgoing IP packet unless 
ECT(0) or ECT(1) was set in the incoming packet.  

* Verify the CE is in fact sometimes set, if packets are ECN-capable
and there is a moderate level of congestion.  (With no congestion,
no packets will be dropped or marked, and with very high congestion,
packets will be dropped instead of marked.  It is not a violation
of any standard *never* to set CE, but in this case the router
could not be said to be ECN-capable...)

* Verify that the router never changes the ECN field in any way
except from ECT(*) to CE.

* Verify that the router responds the same to ECT(1) as it does
to ECT(0).

* Verify that the router only sets CE in ECN-capable packets if
it in fact would have dropped those packets if they weren't 
ECN-capable.  

I can't think of anything else that it would be essential to verify.

One could also check for broken behavior (e.g., dropping packets
*because* the ECN field is non-zero, when the packets would not
have been dropped otherwise, would be clearly broken), but there
are so many possibilities for broken behavior that I can't imagine
checking for all of them.  

Verifying that someone's active queue management mechanism is
reasonable is another question, that I am not attempting to
answer here.