Model catalogue update

The catalogue decides what the agency routes to in production. A wrong id costs
every call that lands on it, so the rules below are not style.

1. Check the retired list first. RETIRED in tests/test_nvidia_default_model.py
   and the removed … / (404) / (410) comments in config/models.yaml name ids
   that failed live on this account. Do not re-add one. A vendor page saying the model
   exists is not evidence it answers here: #1695 (Nemotron 3.5 Lightning) and #1697
   (Llama 3.3 Nemotron Super 49B) both re-added retired ids from vendor docs and went red.

2. Never duplicate a key. Search the dict or YAML for the id before adding it;
   a duplicate silently overwrites the earlier entry (#1689).

3. Unprobed means explicit-only. A new id gets priority: 99, stays out of
   PROVIDER_CANDIDATES and every preset, and its cost entry says it is an estimate
   until a probe (catalogue-probe.yml) has seen HTTP 200 with a real completion.

4. Mirror both configs. config/models.yaml and brain_config.py must agree;
   test_yaml_presets_mirror_brain_config checks it.

5. Prove it. Run, and paste the summary line of:

   pytest -q tests/test_one_model_catalogue.py tests/test_nvidia_default_model.py tests/test_model_router.py

The pre-PR gate (agent/pr_gate.py) runs these suites whenever model config changes,
so an agent PR that breaks rule 1 or 2 does not open. This skill is the advisory half;
the gate is the deterministic half.