I didn't do any deeper investigation on this problem yet, however as of Bogdan's call for bug reports before 1.6.3 here is what I experienced:
- upgraded from r6844 to r6972
- noticed problems with some clients not longer "reacheable" most of the time
- assured with xlog that nat_keepalive() has been called for those clients on their REGISTER requests
- restarted OpenSIPS to dump keepalive_state_file
- there has been no entry for those clients (manually checked just two of them)
Immediately reverted back to r6844 as this happened on a live system with a few thousand users, did no farther testing. Did a quick look to SVN history, but wasn't able to find anything suspicious nat_traversal-related change. Just a wild guess: could it have been caused by the bug fixed in r7008?