forge_http_request_duration_seconds only grows, so rate() and histogram_quantile() read it correctly #4013

Pillar: (0) Trust and calm.

After #3985, GET /metrics builds forge_http_request_duration_seconds from route_timing rows for the current and previous hour (src/metrics/forge.ts). Each hour, when the older hour leaves the window, _bucket, _count and _sum drop. Prometheus reads a drop as a counter reset, so rate() over-counts and histogram_quantile(rate(...)) spikes every hour. The window is also 1–2 hours, not the "last hour" its HELP text says.

Done:

• [ ] The family's counts only grow for the life of the process, for example in-process totals kept alongside createRouteTiming().record(). A restart resets them, as Prometheus expects.
• [ ] HELP and docs/METRICS.md describe it correctly. The doc also says 404, not 401, for a request without credentials.
• [ ] A test fails if a value drops across an hour boundary (inject now).

Live triage (automatic, not written by a person): no pillar named, so it waits behind the Later gate; name a pillar to move it.