fix(openclaw): clamp GPT-5.4 output tokens

[kywch]

Summary

Clamp OpenClaw GPT-5.4 output tokens to the provider limit.

Root cause

OpenClaw sent max_tokens=171802 while GPT-5.4 accepts at most 128000. The ACP path uses internal alias benchflow-openai-gpt-5.4, so raw gpt-5.4 matching alone misses live requests.

Changes

• Cap raw and ACP-alias GPT-5.4 IDs at 128000.
• Replace missing, malformed, or oversized configured values.
• Preserve explicit numeric values at or below the limit.

Scope

OpenClaw + GPT-5.4 only. No global model cap, routing change, or OpenClaw runtime-pin change. #977 is adjacent only.

Validation

• 47 focused tests passed.
• git diff --check passed.

[Galius5136]

I think there is one proxy-alias path still uncovered here. I checked this on the current PR head. These two spellings resolve to the same upstream model (openai/gpt-5.4), but produce different proxy aliases: gpt-5.4 -> benchflow-gpt-5.4, while openai/gpt-5.4 -> benchflow-openai-gpt-5.4.

With the current _MODEL_MAX_TOKENS table, _default_max_tokens(safe_model_alias("gpt-5.4")) returns None, while _default_max_tokens(safe_model_alias("openai/gpt-5.4")) returns 128000. So if session/set_model receives the proxy alias derived from a bare --model gpt-5.4, the clamp is skipped even though both spellings resolve to the same openai/gpt-5.4 upstream model.

I also couldn't find a test exercising the new _default_max_tokens logic. A small regression test that derives the alias through safe_model_alias() for both spellings would pin this without hardcoding the alias-generation behavior.

[kywch]

@Galius5136 thanks for looking into it! I made it more general and added a small test.

[Galius5136]

@kywch confirmed, both spellings now hit the cap:

_default_max_tokens(safe_model_alias("gpt-5.4")) -> 128000

_default_max_tokens(safe_model_alias("openai/gpt-5.4")) -> 128000

The new parametrized test also fails on 3210f5c4 for the bare gpt-5.4 case and passes on def35f88, so it does pin the regression rather than just restating the fix.

Thanks for the quick turnaround!