ci: make the wedge hunt keep what it catches (#564)

[ohohoreilly]

ci: make the wedge hunt keep what it catches (#564)

The hunt reproduced #564 on its second dispatch and we learned nothing from
it. Three reasons, three fixes.

The round hung for 110 minutes and the job's own 120-minute cap cut it off
mid-evidence, so the artifact step never ran on the data that mattered. Bound
each round at 25 minutes instead: a round we kill is a round whose logs we
still have, and the loop moves on to the next sample rather than the whole
batch dying on one.

jhm's output was block-buffered, so whatever it wrote after the last flush
went down with the cancelled job. Run it under stdbuf -oL.

And #568's join warning -- the line that names the stalled replication loop --
is orlyi syslog landing in the CHILD's log file under /tmp/import_repl_test_*.
The fixture only surfaces those tails when it fails its own assertions, so a
round killed from outside discards them. Archive the scratch dirs
unconditionally.

Note a hung round is now itself recorded as a catch. With #571 in, jhm no
longer hangs reaping a test that left grandchildren holding stdout, so a round
that still overruns means something genuinely stuck -- worth stopping on.