Apply same fix to GPT-5.4 OpenAI direct provider

GPT-5.4 has the same /v1/chat/completions tools+reasoning_effort 400 as
GPT-5.5 — kept in lockstep. Verified test_tools_all_built_in_models now
skips for both gpt_5_4-openai and gpt_5_5-openai. The OpenRouter routes
and Pro/Mini/Nano variants were tested and pass, so they're left alone.

https://claude.ai/code/session_01WSZjtD8K2X51BBJgaq2c5M

Author: claude

libs/core/kiln_ai/adapters/ml_model_list.py

@@ -567,6 +567,10 @@ class KilnModel(BaseModel):
                 structured_output_mode=StructuredOutputMode.json_schema,
                 available_thinking_levels=GPT_5_4_OPENAI_THINKING_LEVELS,
                 default_thinking_level="none",
+                # OpenAI rejects reasoning_effort + tools on /v1/chat/completions
+                # for gpt-5.4 direct. Disable function calling until Kiln routes
+                # these models to /v1/responses. The OpenRouter route is unaffected.
+                supports_function_calling=False,
                 suggested_for_evals=True,
                 suggested_for_data_gen=True,
                 supports_doc_extraction=True,