The failing duplex check was the Windows client's, run under Wine in
the network clients job (EXE=.exe). A Windows binary ends its lines
with CR LF, so '^got: s[1-4]$' matched none of its four 'got:' lines,
the check counted 0 of 4, and it failed on every run although the
client read all four, sent all 60, and the server printed 'done'
(its -msg trace shows the line and both KeyUpdates). The lines are
counted with the CRs taken out.
The Wine run of the duplex check now passes three of three; it failed
four of four before. d9b00b7's longer server tail was not the cause;
it costs nothing and stays.