fix: annotate runinfra model set for the type budget, pin 27B cache pricing in tests

The 27B cached input price was verified live after the original test pinned its absence; the test now pins the verified rate. The new provider set line carries a mutable-ok reason like the gate requires.

litellm/__init__.py

@@ -647,7 +647,7 @@ dashscope_models: Set = set()
 moonshot_models: Set = set()
 publicai_models: Set = set()
 darkbloom_models: Set = set()
-runinfra_models: Set = set()
+runinfra_models: Set = set()  # mutable-ok: provider model registry, appended by add_known_models like every sibling set
 v0_models: Set = set()
 morph_models: Set = set()
 lambda_ai_models: Set = set()

tests/test_litellm/llms/openai_like/test_json_providers.py

@@ -425,7 +425,7 @@ class TestRuninfra:
                 None,
             ),
             "runinfra/Inferact/Qwen3.8-2.4T-A95B-NVFP4": (2e-06, 6e-06, 2e-07),
-            "runinfra/Qwen/Qwen3.8-27B": (1e-07, 4e-07, None),
+            "runinfra/Qwen/Qwen3.8-27B": (1e-07, 4e-07, 1e-08),
         }
         for model, (input_cost, output_cost, cache_read_cost) in expected_models.items():
             assert model in model_cost