## Summary
- Fixes periodic production crashes (undici ECONNREFUSED ::1) by
bounding server cache size/lifetime and hardening server HTTP client.
### Root cause
- React server cache (cache(...)) held axios responses indefinitely
across many parameter combinations, causing slow memory growth until the
Next.js app router worker was OOM-killed. The main server then failed
IPC to the worker (ECONNREFUSED ::1:<port>).
### Changes
- `app/lib/data.ts`: Replace unbounded cache(...) with unstable_cache
and explicit keys; TTLs: 60s for teams/detail/favorites/user, 300s for
meta (jobs/skills/accessories/raids/version).
- `app/lib/api-utils.ts`: Add shared Axios instance with 15s timeout and
keepAlive http/https agents; apply to GET/POST/PUT/DELETE helpers.
- `pages/api/preview/[shortcode].ts`: Remove duplicate handler to dedupe
route; retain the .tsx variant using `NEXT_PUBLIC_SIERO_API_URL`.
### Notes
- Build currently has pre-existing app/pages route duplication errors;
out of scope here but unrelated to this fix.
- Ensure `NEXT_PUBLIC_SIERO_API_URL` and `NEXT_PUBLIC_SIERO_OAUTH_URL`
are set on Railway.
### Risk/impact
- Low risk; behavior is unchanged aside from bounded caching and
resilient HTTP.
- Cache TTLs can be tuned later if needed.
### Test plan
- Verify saved/teams/user pages load and revalidate after TTL.
- Validate API routes still proxy correctly; timeouts occur after ~15s
for hung upstreams.
- Monitor memory over several days; expect stable usage without steady
growth.
Justin Edmund
426645813e