docs: front-load API contract on landing and api reference

[brunobuddy]

Summary

The cloud base URL, auth header, key prefix, and manifest/auto model identifier are spread across multiple pages. An LLM helping a user integrate has to chain through several files just to assemble a working request — and the most common hallucination ("can I pass gpt-4o-mini directly?") is never explicitly refuted.

This is a surgical contract-clarity pass on two files. No marketing/positioning changes, no restructure.

• introduction.mdx — add a Quick reference table at the top: cloud base URL, self-hosted base URL, auth header, key acquisition link, the only accepted request model, both endpoint paths, and a note that no client SDK is needed (use openai / @anthropic-ai/sdk with baseURL set).
• reference/api.mdx —
  • Make the cloud curl the primary sample; self-hosted follows. The previous sample only showed http://localhost:2099, which silently breaks anyone integrating with the cloud.
  • Add a Model identifier callout next to chat completions: manifest/auto is the only accepted value; provider-specific names like gpt-4o-mini or claude-sonnet-4 are rejected. This closes the highest-frequency hallucination.
  • Expand the Authentication section into the three explicit steps (sign up → create agent → pass mnfst_<key>), with the M003 link surfaced.

Test plan

• Mintlify preview renders the Quick reference table on /introduction above the feature cards
• The new cloud curl block on /reference/api is the first sample under "Chat completions"
• The M003 link resolves to /errors/M003
• /llms-full.txt now contains the Quick reference (it's pulled from introduction.mdx)