While investigating performance implications of RTO on ECN-enabled TCP flows, it was noted that the recovery from a lost retransmission - using the RTO timeout - is sent without the CWR flag in the TCP header (but the IP ECT codepoint is cleared).
Linux has a similar misbehavior, but there it appears that someone tied the signaling of TCP ECN header flags to the setting of the IP header ECT codepoints - which is not correct though.
dctcp has explicit code in its _cong_signal handler, to mark the RTO retransmission explicitly with the CWR flag.
Reviews and comments more than welcome
Fixed in base r357116.