fix(protocol): clamp max_tokens to available context — fix t3_max_tokens_max HTTP 400

When max_tokens >= max_model_len, vLLM engine rejects the request.
Clamp to (max_model_len - prompt_tokens) in both to_sampling_params
and to_beam_search_params so oversized max_tokens values degrade
gracefully instead of returning HTTP 400.

CCCL logical.cu pattern: handle boundary conditions (empty range,
overflow) gracefully instead of hard-failing.

qwen3_6_scripts/protocol.py

@@ -305,6 +305,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
         max_tokens = self.max_tokens
         if max_tokens is None:
             max_tokens = default_max_tokens
+        if default_max_tokens > 0:
+            max_tokens = min(max_tokens, default_max_tokens)
 
         n = self.n if self.n is not None else 1
         temperature = self.temperature if self.temperature is not None else 0.0
@@ -321,6 +323,10 @@ class ChatCompletionRequest(OpenAIBaseModel):
         max_tokens = self.max_tokens
         if max_tokens is None:
             max_tokens = default_max_tokens
+        # Clamp to available context space so requests with max_tokens ≥
+        # max_model_len don't get rejected with HTTP 400.
+        if default_max_tokens > 0:
+            max_tokens = min(max_tokens, default_max_tokens)
 
         prompt_logprobs = self.prompt_logprobs
         if prompt_logprobs is None and self.echo: