φ-accrual failure detector — public API.

The detector emits a continuous suspicion value φ per monitored node. φ grows monotonically with time since the last heartbeat and is calibrated against observed inter-arrival statistics (EWMA mean and variance). See PhiAccrual.Core for the math and README.md for the positioning (observability-grade, not decision-grade in v1).

Typical usage (bring-your-own-signal)

# Track a node with default estimator settings.
PhiAccrual.track(:node_a@host)

# Call observe/2 from wherever you already receive cross-node
# traffic — GenServer replies, :pg broadcasts, your own pings.
PhiAccrual.observe(:node_a@host)

# Query φ at any time.
PhiAccrual.phi(:node_a@host)
#=> {:ok, 1.82, :steady}

observe/1 auto-tracks unknown nodes with default settings — call track/2 first if you want custom per-node estimator options.

The monitored key is an Erlang node() when fed by the bundled DistributionPing source, but the detector treats the key opaquely and accepts any term — see detector_key/0 for the contract used by non-distribution transport companion packages.

Overload behaviour

observe/2 is bounded: if the target estimator's mailbox exceeds :shed_threshold (default 10_000) the sample is dropped and a [:phi_accrual, :overload, :shed] telemetry event is emitted. For failure detection you don't need every heartbeat, you need enough heartbeats — if you're shedding, your heartbeat rate is too high for your EWMA α settings. Tune α before raising the threshold.