OpenAI client silently ignores `MaxCompletionTokens`

[maragubot]

The OpenAI chat client never applies req.MaxCompletionTokens. In clients/openai/chat_complete.go, the openai.ChatCompletionNewParams struct (around line 274) sets Model, Messages, ToolChoice, Temperature, ReasoningEffort, and ResponseFormat, but never sets a max-tokens field. So when a caller sets MaxCompletionTokens on the request:

• it is not sent to the OpenAI API, so the requested output cap has no effect, and
• the ai.max_completion_tokens span attribute is not recorded (unlike the Anthropic and Google clients).

By contrast:

Client	Sends max to API?	Records ai.max_completion_tokens?
Anthropic	Always (default 16384)	Always
Google	Only when the request sets it	Only when the request sets it
OpenAI	Never	Never

Expected

The OpenAI client should set MaxCompletionTokens (the Chat Completions max_completion_tokens field) on the params when req.MaxCompletionTokens != nil, and record the ai.max_completion_tokens span attribute, matching Google's conditional behaviour.

Notes

Found during the observability review on PR #288. The docs/observability.md catalogue added there documents the current (pre-fix) behaviour accurately; this issue tracks the functional fix.