fix(tests): widen _wait_pidfile's window and name what it saw (#595)

[Masashi-Ono0611]

Closes the last of #595's four sites.

What this fixes

_wait_pidfile in tests/test_watch.bats polled up to 3s for a relaunched
watcher's pidfile to record the new pid. A relaunch is a real fork +
lock-acquire + SIGTERM-the-predecessor + self-write before the pidfile
reflects it, and 3s could lose that race on a loaded CI runner — the flake
#595 caught on PR #436's macos-latest 4/4 shard, at _wait_pidfile "$pf" "$w2".

Of #595's four sites, three turned out to already be fixed independently
since I filed it: marker-gc and codex-monitor by #606, the launcher
re-registration race by #615. This PR is the fourth and last — _wait_pidfile
itself, still on its original budget.

What changed

• Widened the poll window from 3s to 10s, matching the launcher suite's
  wait_for_child_count budget (test_codex_bridge_launcher.bats) rather
  than inventing a new number.
• On timeout, the helper now reports the pidfile path, the wanted pid, and
  what it last read instead of that — bats suite flakes across the process-lifecycle tests: main itself fails ~half its runs, a different test each time #595 asked for a failure message that
  can tell "never arrived" from "arrived as something else"; a bare
  assertion failure at that line couldn't.

What this deliberately doesn't do

#595 also suggested hoisting a shared predicate-wait helper across all four
sites. I didn't do that here: wait_for_child_count is a different
predicate shape (process count, not pidfile content) with no shared caller
today, and unifying the pattern is a design call rather than something this
one-site fix implies. Happy to take that separately if it's still wanted now
that three of the four sites are gone.

Verification

• bats tests/test_watch.bats: 27/27 pass, before and after.
• The target test individually: pass.
• The timeout path's message format tested standalone against a forced
  -unreachable pid — confirmed it reports the actual stale value read from
  the pidfile, not a generic failure.