test: make the midpoint budget case event-driven and drop a duplicate suite

[elkaix]

Fixes the CI failure on main at 37b53fd (run 31552000770), where two suites went red.
Neither was a product defect — the identical content had passed as PR #5's check minutes
earlier.

liveness.sh — t14_midpoint_warning_continues_same_job

The case required the job to complete inside a window of (ceiling/2, ceiling), and then
spent seven serialized poll cycles crossing it — each spawning a node process — against
an 8-second ceiling. On a loaded runner the accumulated poll cost passed the ceiling, the
loop cancelled at its deadline exactly as designed, and the case reported rc=11 want 0.

This is intermittent, not new: the same case failed on c2f2ea9 nine days ago and passed on
b3adebc. Enlarging the numbers cannot fix it — the window stays proportionally
(ceiling/2, ceiling) under any values.

Fix: make completion causally follow the warning instead of racing it. The fixture
already supports MAESTRO_TEST_JOB_TERMINAL_FLAG (tests/job-lock.sh uses it), which
reports the job as running until a named file exists. The case now waits for the halfway
warning, then creates the file, so completion is one poll cycle after the warning rather
than seven. The ceiling is a single variable that the assertion pattern derives from, so
the two cannot drift. All four assertions are unchanged.

Counterexample, made deterministic with the fixture's MAESTRO_TEST_STATUS_HANG (delays
each status call, i.e. "a loaded runner" on demand):

	Before	After
MAESTRO_TEST_STATUS_HANG=1	rc=11 want 0 — the exact CI symptom	passes

shared-git-dir.sh — duplicate lease suite

It re-ran the entire lease suite that tests/run.sh had already run and passed in the
same job, piping it through tail -1 so a failure inside it was undiagnosable — which is
why the original CI log could not say which case broke. No added coverage, ~55s of CI time,
and a doubled chance that any timing-sensitive lease case fails the build. Every other check
in the file stays.

42.87s → 2.07s.

Wait bounds (preventive, labelled as such)

Nine event waits go from 5s to 30s across liveness.sh, lease.sh, and
provenance-edge.sh. These fix nothing observed — they are hardening. Each loop breaks the
moment its event arrives, so a larger bound costs nothing on a healthy host and removes a
false negative on a loaded one; a genuinely absent event still fails, just later. Only loops
that break on success and report "never happened" were widened; no settling delay before a
negative assertion was found.

t19 — deliberately unchanged

t19 fails only in the Codex sandbox, with
MAESTRO_LOCK: process start identity unavailable for pid=.... There ps -o lstart= returns
nothing, so the contender fails closed instead of entering the wait and never prints the
string the case polls for. No bound can fix that. It passes on the GitHub runner
(PASS waiting contender TERM never cancels the active owner in the very run being fixed)
and on developer hosts.

Verification

Full suite under a fabricated empty HOME with /bin/bash 3.2 — what the runner actually
provides, and the environment in which bash tests/lease.sh alone would have given a false
green:

SUMMARY: 18 passed, 0 failed