fix(httpapi): log dependency state on transition, not every probe

[crypto-a]

A kubelet polls /readyz every few seconds for the life of a pod, so
logging steady state there emitted the same warning thousands of times a
day per replica. One service with an unconfigured cache produced 290
identical lines in minutes across three replicas.

That is not a louder signal, it is a quieter one: the line that matters,
a dependency that just broke, is buried under identical copies of a line
that has been true since boot, and the log store bills for every copy.

Keyed by dependency and holding the last state logged, so a flap logs
both edges while a dependency unconfigured since startup logs once. The
recovery line is new and deliberate: without it a log shows only half of
an incident, the moment it broke and never the moment it came back.