[SandObserver]

Summary

• Add BATCH_MS (2500ms) and use it for the per-item requests in the /api/badges and /api/health fan-outs, so one hung service no longer holds the whole batch for PING_MS.
• Keep PING_MS for /api/badge-test and the version check, which a user runs by hand and waits on.
• Add readBadgeUpdate to badge-logic.js, which reads one item's entry as a value or a failure rather than trusting value alone.
• Keep an item's last known badge value when its own fetch failed and mark it stale through a new activityStale input to computeBadgeVisual, reusing the existing status.stale string.
• Document the two deadlines and the per-item failure handling in docs/frontend.md.

Note on the shape of this fix. The tradeoff was between tightening what exists and building a real per-item state machine with server-side result caching. I took the smaller one: the cost of a dead service is a timeout that was simply too long for a background poll, and the misleading zero was information the server already sent and the frontend threw away. Caching results server-side would mean long-lived state in a project that deliberately has none, for a problem that turned out to be two small ones.

Testing
All green, apart from the Docker build smoke test, which needs a local Docker daemon. Added 2 route tests, including one proving a hung service no longer costs the batch its old deadline, and 5 frontend tests for reading a failed item and marking it stale.

Checklist

• Stays within the project constraints: one container, no runtime dependencies, no build step (see CONTRIBUTING.md).
• Tests added or updated for new logic, and the full suite, lint, and typechecks pass.
• Docs updated if behavior or configuration changed.
• CHANGELOG.md updated under Unreleased if the change is user-visible.