test: cover two guard branches and make the third pragma reason precise

[lesnik512]

What

Three # pragma: no cover sites had vague reasons. I checked each against the code:

Site	Verdict	Change
subscriber/usecase.py _run_with_reconnect, if client is None: return ("defensive teardown race")	Reachable. TestOutboxBroker._patch_broker puts back the broker's original client on exit, which is None for a broker built without an engine, and the loops read the client again on every iteration.	Pragma removed. New test test_run_with_reconnect_returns_when_client_is_gone checks that a running loop with no client returns without opening resources or entering inner.
subscriber/config.py OutboxSubscriberConfig.__post_init__, if parent_post_init is not None ("defensive: base has none today")	Not reachable with the pinned faststream 0.7.x, which gives SubscriberUsecaseConfig no __post_init__. The guard is there so upstream init logic keeps running if faststream adds some.	Pragma removed. New test test_subscriber_config_runs_upstream_post_init adds a __post_init__ to SubscriberUsecaseConfig with monkeypatch and checks that registering a subscriber calls it.
testing.py _sync_dispatch, if fake_row is None: return ("defensive: feed just returned this id")	Reachable only in a contrived case, so the pragma stays with a precise reason. In single publish, nothing is awaited between feed() and the lookup. In batch publish, every row is fed before any is dispatched, so the handler of an earlier row does run first. That handler could remove a later row only by reaching into the fake: a lease-guarded delete needs a token the row does not have until its own dispatch, and cancel_timer needs a timer_id, which batch rows never have.	Reason rewritten: "only if an earlier handler in this batch deleted the row off the fake".

Evidence

• Both new tests fail when the guarded code is removed. Without the client is None return, the loop hits AttributeError: 'NoneType' object has no attribute 'engine' and goes into reconnect backoff. Without parent_post_init(), the mock is called 0 times.
• Removal paths in FakeOutboxClient (delete_with_lease, delete_batch_with_lease, cancel_timer) all require a matching acquired_token, or a timer_id with no lease.

Checks (local)

• just install: no lockfile change
• just lint-ci: pass
• just test-ci against a postgres:17 container (compose credentials, non-default host port): 631 passed, coverage 100.00%. The 89 integration tests ran; none were skipped.

[github-actions]

Benchmark gate

✅ gate passed

scenario	msg/s	delete/msg	WALrec/msg	WALB/msg	fpi	upd	del	dead_tup
consumer/w1/b10	952	1.000	6.79	915	242	5000	5000	10000
consumer/w1/b100	948	1.000	6.84	920	243	5000	5000	10000
consumer/w2/b10	1099	1.000	6.81	968	242	5000	5000	10000
consumer/w2/b100	1300	1.000	6.73	955	243	5000	5000	10000
consumer/w4/b10	1010	1.000	6.77	973	261	5000	5000	10000
consumer/w4/b100	1489	1.000	6.85	1031	244	5000	5000	10000
consumer/w1/b100/tfbs100	5176	0.010	6.07	1116	243	5000	5000	10000
producer/w1/b100	2107	0.000	3.04	584	0	0	0	0

Gated (fails the build): delete_calls + tuple counters (upd/del/ins) + the producer's insert_calls, exact; select_calls within +2; wal_records within a 10% band. msg/s, WAL bytes and total calls are informational (timing/FPI noise).