Replace ChatGPT help with streaming AI assistant conversations

[iHiD]

Shifts the editor's AI help to the same model as Jiki's llm-chat-proxy: the browser talks directly to a Cloudflare Worker (exercism/llm-chat-proxy) which streams Socratic-mentor responses from Gemini, replacing the old fire-and-forget ChatGPT proxy + SPI callback + ActionCable round-trip.

Entitlement model

• Insiders: unlimited conversations on every exercise.
• Everyone else: one free exercise conversation, ever, pinned to the first exercise they start a conversation on.
• Uniform cost caps enforced by the worker (10 msgs/min burst, 100/day, 500/month per user), matching Jiki.

Rails

• Solution::AssistantConversation model (JSON messages column, unique per solution) + Mandate commands mirroring Jiki's: CreateConversationToken (1h HS256 JWT bound to the solution), CheckUserAccess, FindOrCreate, AddUserMessage, AddAssistantMessage, AddMessage, VerifyHMAC, VerifyConversationToken.
• POST /api/v2/solutions/:uuid/assistant_conversation mints tokens, gated by Cloudflare Turnstile (Captcha::VerifyTurnstileToken, fail-open on CF outage). user_messages / assistant_messages persist the transcript; assistant messages must carry a valid HMAC signature from the worker (payload includes the solution uuid and a 10-minute freshness window, closing Jiki's replay gap).
• GET /api/v2/solutions/:uuid/assistant_context serves instructions + tests to the worker, authenticated with the same conversation JWT.

Front-end

• New editor/AssistantChat module ported from Jiki: streaming chat panel with token caching (one Turnstile pass per session), transparent refresh on token expiry, usage warnings at 90%, cap/upsell states, and free-conversation upsell for non-insiders.
• Replaces the ChatGPT panel/dialog/model-picker; Stuck? button now routes to the assistant tab for everyone.

Removed

The entire legacy pipeline: Submission::AI::ChatGPT::*, the ai_help endpoint, /spi/chatgpt_responses, Submission::AIHelpRecordsChannel, ResetChatGPTUsageJob + schedule, chatgpt usage counters, and the ChatGptFeedback components. The submission_ai_help_records table is left in place (nothing reads it). The OpenAI training-data/tagging pipeline is untouched.

Requires before deploy

• exercism-config: assistant_chat_jwt_secret, assistant_chat_hmac_secret (secrets, shared with the worker) and assistant_chat_url (config; prod https://chat.exercism.org, local http://local.exercism.io:3027).
• The worker deployed from exercism/llm-chat-proxy (see its PR).

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJjLSVfCaDvJQM3oATNcL6