[robkooper]

What changed

Tests only, no production code changes. They pin the exact end_date instant (end_date == utcnow()) in all four places that check a model's end date:

• _resolve_single_access in lumen/services/llm.py: blocked at NOW, allowed at NOW + 1µs. There is also an end-to-end check that get_model_access_status / get_model_access block at the exact instant.
• ModelConfig.active in lumen/models/model_config.py: covers both the Python property and the SQL expression. It is inactive at NOW and active at NOW + 1s. The test uses 1 s because database sub-second precision varies.
• ModelSnapshot.active in lumen/services/model_cache.py: resolve() hides the model at NOW and returns it at NOW - 1µs.

Every test patches the module-level utcnow of the module under test to a fixed naive-UTC 2030-01-01 12:00:00, so no test reads the real clock. The PR also adds one CHANGELOG bullet.

How I tested it

• uv run pytest tests/unit/test_llm_access.py tests/unit/test_model_cache.py: pass.
• Mutation checks, run locally and not committed. Each mutation made at least one new test fail:
  • llm.py:172: <= changed to <
  • model_config.py:73: > changed to >=
  • model_config.py:78: > changed to >=
  • model_cache.py:102: > changed to >=
• Full uv run pytest passes locally. The PostgreSQL and Redis tests were skipped locally because no database services were available; CI runs them.

Fixes #311

[robkooper]

How to test
On a disposable local instance with a development user and dummy backend configured as described in the README, give a public test model a future UTC End Date in Admin → Configuration. Confirm an ordinary user can see and use it, then set the End Date to the past and save. After configuration reload, it should disappear from model listings and requests for it should be rejected. This PR only adds regression coverage, so existing behavior should stay the same.