Fix Claude gateway 400 + add Gemini/GPT price tracking

- Claude gateway models (Opus 4.8/4.7, Sonnet 5) run with extended thinking enabled, which
  requires temperature unset (or 1.0). The model blocks set temperature: 0.2, causing every
  request to fail with HTTP 400. Remove model_kwargs/temperature and raise max_tokens
  4096 -> 16384. Verified against the anthropic-passthrough endpoint with the full agent
  request shape (system + tools + cache_control): now returns 200 / tool_use.
- Register litellm prices for gemini-3-5-flash-fair and gpt-5-5-genai-responses (from reveng's
  model_prices.json) so their cost is tracked instead of ignored. Claude self-computes cost
  from its cost: block (verified against reveng: opus/sonnet numbers match).
- Remove robotrumble_llama.yaml (stale env-indirection config reintroduced by a merge).

Author: john-b-yang

configs/ablations/ladder/robotrumble_llama.yaml

@@ -47,5 +47,28 @@
                 ]
             }
         ]
+    },
+    "gemini-3-5-flash-fair": {
+        "litellm_provider": "openai",
+        "mode": "chat",
+        "max_input_tokens": 1048576,
+        "max_output_tokens": 65535,
+        "max_tokens": 65535,
+        "input_cost_per_token": 1.5e-06,
+        "output_cost_per_token": 9e-06,
+        "output_cost_per_reasoning_token": 9e-06,
+        "cache_read_input_token_cost": 1.5e-07,
+        "source": "reveng model_prices.json (gemini-3-5-flash-genai tier)"
+    },
+    "gpt-5-5-genai-responses": {
+        "litellm_provider": "openai",
+        "mode": "responses",
+        "max_input_tokens": 1050000,
+        "max_output_tokens": 128000,
+        "max_tokens": 128000,
+        "input_cost_per_token": 5e-06,
+        "output_cost_per_token": 3e-05,
+        "cache_read_input_token_cost": 5e-07,
+        "source": "reveng model_prices.json"
     }
 }

configs/mini/litellm_custom_model_config.yaml

@@ -3,9 +3,10 @@ model_name: aws-claude-4-7-opus-fair
 model_class: codeclash.agents.mini_anthropic_model.AnthropicModel
 api_key_env: LLAMA_API_KEY
 base_url: https://api.llama.com/experimental/passthrough/anthropic
-max_tokens: 4096
-model_kwargs:
-  temperature: 0.2
+max_tokens: 16384
+# NOTE: these gateway Claude models run with extended thinking enabled, which requires
+# temperature to be unset (or 1.0). Setting temperature (e.g. 0.2) returns HTTP 400 — so no
+# model_kwargs/temperature here.
 cost:
   input: 5.0e-06
   output: 2.5e-05

configs/mini/models/llama_opus_4_7.yaml

@@ -5,9 +5,10 @@ model_name: claude-4-8-opus-vertex-genai
 model_class: codeclash.agents.mini_anthropic_model.AnthropicModel
 api_key_env: LLAMA_API_KEY
 base_url: https://api.llama.com/experimental/passthrough/anthropic
-max_tokens: 4096
-model_kwargs:
-  temperature: 0.2
+max_tokens: 16384
+# NOTE: these gateway Claude models run with extended thinking enabled, which requires
+# temperature to be unset (or 1.0). Setting temperature (e.g. 0.2) returns HTTP 400 — so no
+# model_kwargs/temperature here.
 cost:
   input: 5.0e-06
   output: 2.5e-05

configs/mini/models/llama_opus_4_8.yaml

@@ -4,9 +4,10 @@ model_name: claude-5-sonnet-genai
 model_class: codeclash.agents.mini_anthropic_model.AnthropicModel
 api_key_env: LLAMA_API_KEY
 base_url: https://api.llama.com/experimental/passthrough/anthropic
-max_tokens: 4096
-model_kwargs:
-  temperature: 0.2
+max_tokens: 16384
+# NOTE: these gateway Claude models run with extended thinking enabled, which requires
+# temperature to be unset (or 1.0). Setting temperature (e.g. 0.2) returns HTTP 400 — so no
+# model_kwargs/temperature here.
 cost:
   input: 3.0e-06
   output: 1.5e-05