The reply-path stall detector

This is the most BNG-specific health concept in the system and it exists nowhere
in the raw telemetry plane. The sampler compares the access port's receive rate
(CPE toward BNG) against its transmit rate and tracks consecutive ticks in
which the emit ratio falls below threshold while the receive rate is above a
load floor. Any healthy tick resets the counter. The requirement for consecutive
samples is what prevents a single slow tick from flapping an alarm.

The resulting state — rx_pps, tx_pps, ratio, stall_samples — is published
on GET /api/bng/health and as the bng_reply_path_* metric family. It detects
a class of fault that neither the controller nor a simple interface counter can
see: subscriber traffic arriving and being accepted, but replies not making it
back out.

Access Model and Conventions

A small number of conventions apply across the entire interface. They are stated
once here and not repeated per endpoint.