SDK follow-ups when server PR interfaze#218 (OpenAI spec-compat) deploys #14

Tracking the SDK changes that should land once InterfazeAI/interfaze#218 is merged and deployed — not before, since they're correct against the current server. From an audit of interfaze-py against #218.

Already handled (deploy-independent), for context:

• Streaming json_object fence leak — fixed in fix: strip json_object fence on streaming (aggregated) path #13.
• Role tolerance / non-stream fence strip / logprobs handling — kept as backward-compat (no-op post-deploy).
• models.list() already parses today (openai-python is lenient; owned_by just fills in post-deploy) — no change needed.

Flip when #218 deploys

• max_completion_tokens — #218 honors it as an alias for max_tokens (explicit max_tokens still wins).
  • Update README "Good to know": the bullet "Use max_tokens (not max_completion_tokens) to bound output" becomes stale.
  • Drop the known-limitation note on ChatInterfaze (feat: ChatInterfaze — in-repo LangChain chat model (#5) #12) about ChatOpenAI renaming max_tokens→max_completion_tokens — it stops being a silent no-op.
• models.retrieve() — #218 adds GET /v1/models/{id}; today it raises NotFoundError (HTML 404 page), works once deployed. Add a live-QA check. (Mocked unit tests for the OpenAI-shaped list/retrieve responses can land anytime.)

Not changing (intentional server behavior)

• Chat completions accept any model string (echoes interfaze-beta); post-#218 models.retrieve("gpt-4") will 404 while chat.create(model="gpt-4") still echoes — server-side asymmetry flagged as intentional in #218.