3 Nit fixes first (cheap), 2 Low test additions (sync Bulkhead Hypothesis
suite, sync on_error BaseException tests), release notes + PR. Test-only,
no production changes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

`test_no_slot_leak_after_drain` asserted `bulkhead._sem._value ==
max_concurrent` after a gather of n_requests. That assertion read a
CPython implementation detail of asyncio.Semaphore — non-portable across
interpreters and brittle against any future asyncio refactor.

Replace with a behavioral check: after gather completes, submit
max_concurrent fresh acquires under a tight acquire_timeout (0.05s). If
any slot leaked, the post-drain gather would block past the timeout and
BulkheadFullError would surface as a deterministic failure.

The `_acquire_timeout` override is still a private-attribute touch
(per-instance test-local config, not a CPython internal), so the
`# noqa: SLF001` annotation stays.

Closes audit Nit finding (test_bulkhead_props.py:113) from
planning/audit/2026-06-07-deep-audit.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ted total

`assert len(budget._deposits) > 0` would have passed even with the
internal lock removed and the deque corrupted, as long as ONE survivor
remained. The audit's recommended tighten is to assert the exact
expected total — computable from the test constants.

After the 0.8.3 deposit-hoist, deposits count requests (not attempts),
so `(N_SYNC_THREADS * N_OPS_PER_THREAD) + N_ASYNC_TASKS` = (4*50)+20 = 220
is the exact expected. Budget TTL is 60.0s so no purge fires within the
sub-second test runtime.

If the actual count diverges, the test now surfaces it.

Closes audit Nit finding (test_threading_with_shared_budget.py:77) from
planning/audit/2026-06-07-deep-audit.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… test

`test_async_client_accepts_explicit_decoder_without_pydantic` covered the
AsyncClient escape hatch (explicit `decoder=` bypasses the pydantic
fail-fast). The sync Client has the same `_default_pydantic_decoder()`
gate at the same point in `__init__`, but no test asserted the parallel
behavior.

Add `test_sync_client_accepts_explicit_decoder_without_pydantic` as the
sync mirror; hoist `_FakeDecoder` to module top to keep the two tests DRY.

Closes audit Nit finding (test_optional_extras_pydantic_missing.py:41)
from planning/audit/2026-06-07-deep-audit.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds tests/test_bulkhead_sync_props.py mirroring the async file's three
property tests, using threading.Thread + ThreadPoolExecutor + a
lock-guarded shared counter instead of asyncio.gather:

1. in_flight_never_exceeds_max_concurrent (under any thread interleaving)
2. fail_fast_rejects_when_at_capacity (acquire_timeout=0 raises BulkheadFullError)
3. no_slot_leak_after_drain (behavioral check, no _value peek)

Strategies match the async file: max_concurrent ∈ [1, 8],
n_requests ∈ [1, 32], delay ∈ [0.001, 0.005] s. Async file is unchanged;
the new file establishes sync/async parity for the resilience
property-test surface.

Closes audit Low finding (test_bulkhead_props.py:1, sync gap) from
planning/audit/2026-06-07-deep-audit.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

test_middleware.py::test_on_error_lets_cancelled_propagate pinned the
async invariant: on_error catches Exception, NOT BaseException, so
asyncio.CancelledError propagates through compose_async unchanged.

The sync `on_error` decorator has the same `except Exception` clause but
no test asserted the parallel KeyboardInterrupt / SystemExit invariant.
Add two sync mirror tests so a future regression that widens the catch
to BaseException surfaces immediately.

Closes audit Low finding (test_middleware.py:162 sync parity gap) from
planning/audit/2026-06-07-deep-audit.md.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Five test-only fixes close the last actionable audit findings: behavioral
drain check, exact threading-budget deposit total, sync Client decoder-
escape peer, sync Bulkhead Hypothesis suite, sync on_error BaseException
propagation tests. No production code change, no behavior change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>