fix: SIGPIPE false negatives in pipefail shell tests

[mkoura]

Problem

A regression run wedged with 8 of 9 cluster instances stuck in
respin_in_progress, each start-cluster burning its full 7200s
tx-firehose wait and then retrying, forever - while tx-firehose was
demonstrably submitting transactions the whole time.

The wait was:

if tail -n 100 "$logfile" | grep -q "TxFirehose.Submit.Success"; then

grep -q exits on the first match and closes the pipe while tail may
still have a pending write(). tail dies on SIGPIPE with 141, and
pipefail promotes that to the pipeline's exit status, so the test reads
as "no match" even though the pattern is right there. Reproduced against
the live log:

$ tail -n 100 .../tx-firehose.stderr | wc -c
21691
$ bash -c 'set -euo pipefail; if tail -n 100 .../tx-firehose.stderr | grep -q "TxFirehose.Submit.Success"; then echo MATCH; else echo "NOMATCH rc=$?"; fi'
NOMATCH rc=141

The perverse part: a healthy tx tool fills the last 100 log lines with
the very pattern being matched, so the match lands on line 1, leaving
tail with the most left to write. The healthier the tool, the likelier
the false negative.

This is a race on syscall interleaving, not a size threshold - the same
construct on the same file returned 141 and later 0. Output size only
shifts the odds.

Changes

Most of the affected code is in cardonnay, fixed there and released as
0.4.4 (IntersectMBO/cardonnay#160): five sites including all three
_wait_for_tx_*_tx helpers, plus check_spend_success, which had the
inverted failure mode - 141 negated by ! returned "inputs spent" while
they were not. This PR bumps the dependency to pull that in.

Three sites in this repo:

• cluster_scripts/testnets/start-cluster - the replay wait is a
  while loop, so 141 ended it early and skipped waiting for the
  relay node to finish replaying, silently. Dropping -q is enough:
  plain grep scans all of its input, so tail always finishes
  writing. Verified against /bin/grep (GNU grep 3.11) rather than the
  shell's ugrep shim, in case of a /dev/null-implies--q
  optimisation - there is none.
• runner/runc.sh - a single awk replaces sed | head, so the
  assignment sees awk's own status and an unreadable .git still aborts
  under errexit.
• scripts/postgres-start.sh - drops the pipeline entirely. A quoted
  [[ == ]] right-hand side is already a literal match, which is all
  grep -F was there for, so this loses a fork too.

Verification

• Every changed predicate exercised across match / no-match / producer-
  failure / large-input-with-early-match; the replay wait is 0/15 false
  negatives on a 2.4MB input where the old form was 10/10.
• runc.sh: output identical to sed | head on this worktree's real
  .git, on a two-gitdir file, and on a file with no gitdir: line.
• postgres-start.sh: matches old behaviour on all cases including
  pg_dir=/var/tmp/pg*x, where * stays literal and /var/tmp/pgYx
  correctly does not match.
• ./ai_run.sh make lint clean.
• cardonnay v0.4.4 is exactly v0.4.3 plus the one commit.