test(teardown): bound the holder-readiness wait and fail loudly on timeout (#1123)

[fujibee]

test_teardown_reap.bats hung the CI shard it landed in: 419 of 435 tests
completed, then thirty minutes of silence and a terminated job leaving xargs
and several bash orphans (#1123).

The cause is not the reaper. _start_holder ran inside a command substitution
and started a holder that never closed stdout and stderr, so the capturing pipe
never reached EOF and _start_holder itself did not return.

• Close the holder's standard streams to /dev/null.
• Extract the readiness wait into _wait_for_holder_ready with an explicit
  bound, non-zero when the bound is reached.
• The reaper's own six-second bound now returns non-zero as well; teardown keeps
  the failure, still removes the directory, and returns non-zero at the end, so a
  reaper that gives up is visible rather than silent.

Measured before and after, isolating the stall: before the change the file
produced no output for sixty seconds and was killed; after it, no timeout, five
of its six tests pass and the sixth fails only because ps is refused in the
sandbox that ran it. Regression tests short-circuit each bound and go red when it
is reached.

Based on #1072, which this makes landable.