Adds HealthResponse carrying inflight count, total requests, 1/5/15-min
EWMA load averages, uptime, and NumCPU so a scheduler can pick the least
busy worker. A background sampler updates the load averages every 5s,
stopped by a new idempotent Close method. Work endpoints (/collect,
/evaluate, /report) are wrapped with a trackWork middleware; /health
and /definition are excluded so polling traffic does not pollute the
signal.
Errors from provider.Collect() and json.Marshal were returned with HTTP
200, making failures invisible to monitoring, proxies, and clients that
check status codes. Return 500 Internal Server Error so HTTP-level
tooling can detect failures without parsing the response body.