Admin Hardening

QA READY — Daryle AI DevOps Harness

What changed: The bot had no automated way to detect when it
stops responding — failures only appeared via user complaints.
Now a daily cron actually tests the bot, model errors get
logged, and models can be swapped from the admin panel without
a deploy.

What to test:

1. 4-provider status cards
   Before: Only OpenAI and Pinecone had status cards
   After: Anthropic and Gemini cards also visible
   Steps: Admin → Service Health. Count cards at top — 4
   expected. Click "Run Check Now" — all 4 update.
2. LLM Failure Events panel
   Before: No visibility into AI model errors
   After: Panel shows errors grouped by type; empty state
   when system is healthy
   Steps: Admin → Service Health → scroll below donut charts.
   "LLM Failure Events — Last 24h" section visible.
3. Model Configuration panel
   Before: Changing AI model = code deploy
   After: Change model from admin, live within 60 seconds
   Steps: Admin → Service Health → scroll to bottom. Six rows
   (grounded/fast/claude/pro/chatgpt/gemini) with
   editable fields and Save buttons.
4. Chat modal model info
   Before: Chat modal showed messages only
   After: Footer shows model name, response time, fallback
   Steps: Admin → Chats → open any chat. Footer line shows
   model and latency (only for chats with log data).
5. Daily canary (deploy-gated)
   Before: No automated health check
   After: GitHub Actions fires 6am + 6pm UTC, tests each
   route, emails alert on 2nd consecutive failure
   Steps: GitHub → Actions → Health Canary → Run workflow.
   Edge case: Requires SUPABASE_FUNCTIONS_URL +
   SUPABASE_SERVICE_ROLE_KEY in repo secrets first.