fix: ground AI analysis in SDK documentation

[nicknisi]

Summary

The AI analysis in workos doctor was generating inaccurate findings because it had no access to actual SDK documentation or project files — it was guessing based on training data and hallucinating SDK options, config properties, and "anti-patterns" that were actually correct code.

Before

• AI only received metadata (framework name, SDK version, env type)
• Hardcoded "correct patterns" for Next.js in the prompt — didn't scale
• Generated findings like "AuthKitProvider needs 'use client'" (wrong), "middleware matcher too broad" (wrong), "missing error handling in handleAuth" (wrong)

After

• Fetches the SDK's README from GitHub at runtime — the actual source of truth for correct patterns, imports, and configuration
• Includes project files (middleware, layout, page, callback route, env) in the prompt so the AI sees real code
• Includes .env.local with redacted secrets so the AI can verify env var names and redirect URIs
• Task reframed: "compare code against documentation, report deviations" instead of "guess what's wrong"
• Scales to all SDKs — 13 package-to-repo mappings covering all AuthKit JS SDKs
• Zero maintenance — always fetches latest README, no hardcoded patterns to update

Files Changed

• src/doctor/agent-prompt.ts — Async README fetch, project file collection, env redaction, rewritten prompt
• src/doctor/checks/ai-analysis.ts — Await async prompt builder
• src/doctor/checks/ai-analysis.spec.ts — Async test updates
• src/doctor/index.ts — Pass installDir to analysis context