Server 'unreachable' in the browser but healthy in curl: check CORS per origin
A frontend reported the API as unreachable while curl /health returned 200. The cause was CORS on the plain GET: FastAPI CORSMiddleware sends Access-Control-Allow-Origin only for whitelisted origins, so from any other origin (alias domains like a second *.vercel.app name, Vercel preview URLs, localhost) the browser blocks the response. fetch() then rejects with the same TypeError as a network failure, so retry and wake-up logic can't tell it apart. Diagnose with curl -D - <api>/health -H 'Origin: <site>' | grep -i access-control-allow-origin for each origin the app is served from. A frontend-only fix that removes the problem entirely: proxy the API through the frontend's own origin, using a vercel.json rewrite /api/(.*) -> <api>/$1 plus a Vite dev server.proxy, so the browser never makes a cross-origin call.
Follow-up after implementing the proxy: vercel.json rewrite { source: '/api/:path*', destination: 'https://<api>/:path*' } placed before the SPA catch-all, plus Vite server.proxy { '/api': { target, changeOrigin: true, rewrite: p => p.replace(/^\/api/, '') } }. Two gotchas. (1) A stale VITE_API_URL env var in Vercel still baked the absolute URL into the build and bypassed the proxy. The Vercel CLI refuses 'env rm' without human approval, so I renamed the override the code reads (VITE_API_DIRECT_URL) and defaulted to '/api'. (2) Preview deployments sit behind Vercel Authentication (curl gets a 302), so test them with 'vercel curl <url> -- -s', and grep the built JS for the API base to confirm which URL was compiled in.