MAHAi-Movie Mind: Pre-Launch Review

We are preparing the MAHAi-Movie Mind app for public launch. It is a React frontend with a Python/FastAPI backend that proxies to an AWS Bedrock Claude model, with a retrieval knowledge base. An authorized security and safety review found issues we must fix before promoting the app for personalized health-record use.

Ground rules:
- Work on a new branch. Do not run destructive commands or touch production data.
- Before any large change, show me a short plan and wait for approval.
- Treat all uploaded content (labs, bills, records) as sensitive health data.
- Add or update tests for each fix. Keep the app deployable at every step.

Please start by reading the codebase and confirming where each item below lives, then propose a prioritized plan.

PRE-LAUNCH BLOCKERS (do these first):

1. Individual user accounts. Replace the single shared HTTP Basic password with per-user authentication tied to the course purchase. Prefer a hosted auth provider (e.g. Cognito, Auth0, Clerk, or Supabase Auth) over building our own. Add: sessions with real server-side logout, password reset, MFA for admins, per-user IDs used for all data ownership and audit, and the ability to revoke or suspend one user.

2. Authorization on every request. Derive the current user from the server-side session, never from the client. Add ownership checks so a user can only read their own conversations, uploads, and results. Add an automated test proving user A cannot access user B's data.

3. Rate limiting and cost controls on /api/chat and /api/upload. Enforce server-side: requests per minute per user, global concurrency limit, max message-history length, max attachment size and extracted-text size, output-token cap, and a daily spend cap per user. Return HTTP 429 with Retry-After when throttled. Add AWS Budget and Bedrock quota alarms.

4. Health-safety guardrails outside the model prompt. Add a server-side layer that: detects medical emergencies and self-harm and returns escalation guidance; blocks or heavily gates prescription-tapering, pediatric-dosing, and personalized-dosage outputs; and labels uncertainty. Remove or clinically rewrite the built-in prompt templates that request tapering plans, child dosing, and specific supplement brands/doses. Add a regression test suite covering emergency, pediatric, pregnancy, psychiatric, medication-interaction, and oncology scenarios.

5. Privacy and data-handling. Add a plain-language privacy notice and a just-in-time consent step before the first upload. Implement per-user data export and deletion. Confirm and document whether AWS Bedrock model-invocation logging is enabled, and where request/response data and application logs are stored and for how long. Add Cache-Control: no-store to all endpoints that return health data.

6. Fix the misleading "save my Health CEO File" flow. Either implement real authenticated, encrypted per-user storage with retention/deletion, or change the copy to state clearly that files and chats exist only in the current session, show which attachments are active, confirm before clearing a conversation, and offer a local export.

NEXT (after the blockers):

7. Add security headers on all responses: HSTS, a Content-Security-Policy including frame-ancestors 'none', X-Frame-Options, X-Content-Type-Options: nosniff, Referrer-Policy, and Permissions-Policy.
8. Make the public health-check endpoint return only a generic status (no provider/model details); normalize production API errors; return human-readable citation titles instead of internal filenames.
9. Add preview-before-send for prompt templates and a file-review step before upload.
10. Run a WCAG 2.2 AA accessibility pass: accessible names for the upload and send buttons, a live region for streamed responses, keyboard/focus testing, and larger mobile safety text.
11. Self-host fonts and load any third-party video only after user action or consent; set a strict referrer policy.

For each item, tell me: the files you will change, the approach, the tests you will add, and anything that needs a decision from me (for example, which auth provider). Do not start coding item 1 until I approve the plan.