Clear all attempts for key.
Account-grain keys only. Called after a successful auth to forgive a
user's own typos, which is safe because the key *is* the account being
attacked — clearing it requires that account's credential and can only
widen the budget against an account the caller already holds.
Never call this on an IP-aggregate key. That bucket is the
distributed-spray backstop, and success on one account must not zero the
budget shared by every other account reachable from the same IP;
otherwise an attacker interleaves their own logins (or, under open signup,
throwaway account creations needing no credential at all) with guesses
against arbitrary victims and sprays indefinitely from one address.
The only success that records on an IP bucket is an open signup (charged
like a failure); no success clears one, so with no reset each is a
monotone-within-window counter that only time clears.
Because they are monotone, the IP buckets are one instance per auth
surface (login_ip_rate_limiter, signup_ip_rate_limiter,
bootstrap_ip_rate_limiter) rather than one shared across all of them.
A single instance meant a failure on any surface spent the budget that
bounds guessing on every other one, and one caller's exhaustion denied
four routes at once — a cost that used to be masked by the reset this
doc forbids.
Attempts still in flight on key (reserved by concurrent requests) are
kept: a reset forgives what was recorded, not what is still running.