test: move TTL expiry on a controlled clock instead of sleeping

[allen0099]

Summary

Closes #185

Part of #123. About 22 s of the ~25.5 s suite (without live servers) was spent sleeping until TTLs ran out. The tests now move a clock instead.

The clock fixture (tests/conftest.py)

• clock points every time read behind a TTL or expiry check at a settable Clock:
  • time.time() in the memory backend;
  • datetime.now() in session.manager, session.models, state.manager and state.models.
• It patches module references only, and no library code changes.
• clock.advance(seconds) moves it.
• await clock.wait(seconds, backend) advances it and also sleeps for real when backend is not the memory backend. It is for the tests parametrized over memory and a live Redis, whose server keeps its own time.

Converted tests

Test	Was	Now
test_memory.py: ttl_expiry, cleanup, get_evicts_the_expired_entry_it_skips; test_proxybackend.py::test_backend_cleanup	1.05–2 s sleep	clock.advance
test_memory.py: cleanup_task_impl, write_only_use_starts_the_cleanup_task	2 s sleep for a 1 s sweep interval	clock.advance; the instance's cleanup_interval is set to 10 ms with monkeypatch, and the test polls the dict for up to 1 s of real time
test_cache_manager.py: 6 TTL tests; state/test_manager.py::test_get_state_metadata_after_expire	1.1–1.2 s sleep	clock.advance, or clock.wait for the tests also parametrized with Redis
session/test_jwt.py::test_jwt_expiration_enforced	1.2 s sleep	PyJWT checks exp against the real time, so the session is created 5 s in the past on the clock. Its backend entry and the session itself are still current on the clock, so only the token check can reject it.
backends/test_memcached.py::test_memcached_tries_a_failed_server_again_after_the_dead_timeout	1.1 s sleep	pymemcache times the dead timeout with the real clock, so the test sets _DEAD_TIMEOUT to 0.05 s and sleeps 0.1 s. A separate assertion pins the documented value of 1 s.
session/test_models.py::test_session_renew, session/test_manager.py::test_sliding_expiration	10 ms sleep "to ensure time passes"	clock.advance(1). test_session_renew now asserts > instead of >=.

Sleeps that stay

• The Redis-parametrized variants above, through clock.wait.
• The live-server tests that rely on real server-side expiry: test_redis.py:316 (1.5 s), and test_memcached.py:460 (2 s) and :790 (2.1 s). Memcached's expiry has one-second resolution.
• Short sleeps that are not waiting for a TTL:
  • race windows (time.sleep(0.05) in the proxy tests);
  • lock contention (test_lock.py);
  • test_cache_hit.py's slow handler;
  • a 10 ms yield in test_cache_manager.py.

Tests

Runtime (uv run pytest, without live servers): 25.5 s → 3.6 s, with 845 passed and 190 skipped both before and after. The suite also passed three times in random order.

Mutation checks, full suite, with each mutation applied on its own:

Mutation	Converted tests that fail
Memory entries never expire (_is_live returns True)	all 5 converted memory tests, test_backend_cleanup, and all 6 CacheManager TTL tests (25 tests fail in total)
The sweep removes nothing	cleanup_task_impl, write_only_use_starts_the_cleanup_task, test_backend_cleanup
CacheManager ignores default_ttl	test_add_uses_default_ttl, test_default_ttl_used_when_not_specified
Session.renew keeps the old expiry	test_session_renew
Sliding renewal never happens	test_sliding_expiration, plus 7 other sliding tests
JWT verify_exp is False	test_jwt_expiration_enforced
The backend does not pass dead_timeout to pymemcache	the dead-timeout test
State expiry: memory expiry and _is_past disabled together	test_get_state_metadata_after_expire[MemoryBackend]

On the last row: that state test, like before this PR, passes if only one of the two expiry paths is broken, because either one hides the state.

Other checks

• mypy --strict on the changed test files: clean.
• pre-commit: passes.
• uv lock --check: passes.

Live servers were not run by me.

CHANGELOG

None: test-only.