aion-server 0.31.0

//! Server half of the liminal connection dead-man switch: the liveness probe.
//!
//! # The gap this closes
//!
//! The server's connection lease ([`HeartbeatTracker`]) is advanced by frames a
//! worker SENDS, and an idle worker sends nothing. So on 2026-07-29 a healthy,
//! connected, idle worker's lease expired 37 seconds after its last activity —
//! "idle worker connection lease expired; worker deregistered" — the worker was
//! never told, kept believing it was connected, and the next dispatch parked
//! forever on a queue nobody was serving. The same absent mechanism cost a live
//! run 28 minutes of mutual blindness when a link died mid-activity: the server
//! pushed into a socket nobody was reading and the worker blocked on a socket
//! nobody was writing.
//!
//! # The probe
//!
//! [`LivenessProbe`] pushes a [`LivenessPing`] to every liminal-connected worker
//! on a fixed cadence and waits for the correlated [`LivenessPong`]. One
//! exchange proves both legs:
//!
//! - The PONG proves the worker is alive to the server, and its arrival advances
//!   the worker's connection lease — so an idle-but-alive connection is no
//!   longer "idle" at the lease layer and the idle expiry cannot fire on it.
//!   That is the structural death of the first failure above.
//! - The PING's arrival proves the server is alive to the worker, whose own
//!   dead-man switch (`aion-worker`'s `liminal_liveness`) declares the link dead
//!   when pings stop. That is the death of the second.
//!
//! A ping that is not answered inside the cadence is logged LOUDLY with the
//! worker's identity and queue, and withdraws the worker's DISPATCH
//! ELIGIBILITY once the silence outlasts the window.
//!
//! # Why eligibility, and not the connection lease
//!
//! This module used to claim that an unanswered ping let "a genuinely dead
//! worker's lease run down" so the expiry sweep would reap it. **That was
//! false, and run `dfd2117c` proved it**: the server could not push to a worker
//! for fifteen minutes and the worker never lost its lease, because the
//! worker-side liveness pump beats from a background task and keeps refreshing
//! it whatever the worker's serve loop is doing.
//!
//! Two different facts were collapsed into one lease:
//!
//! - **the worker process is alive** — proven by anything the worker sends,
//!   pump included, on the worker-to-server direction;
//! - **the server can reach the worker's dispatch path** — proven only by an
//!   answered ping, on the server-to-worker direction.
//!
//! Only the second is a dispatch precondition, and
//! [`LiminalWorkerDelivery::push_payload_with_deadline`] already says why the
//! ping is the only thing that can prove it: it rides "the exact path a
//! dispatch would take, not a parallel one that could be healthy while the real
//! one is not." The pump is exactly such a parallel channel, so it must not
//! feed the fact the ping exists to establish.
//!
//! So the probe advances a SEPARATE reachability clock
//! ([`HeartbeatTracker::record_dispatch_reachability`]) and publishes an
//! eligibility verdict the dispatch selector honours. The connection lease and
//! its expiry sweep are untouched and still own process liveness — there is
//! still deliberately no second reaper, and an unreachable worker is excluded
//! from dispatch rather than torn down.
//!
//! # No knobs
//!
//! Both timings are DERIVED from the operator's existing
//! `worker.heartbeat_window` — the one place they already declared what silence
//! means:
//!
//! - the ping cadence is [`sweep_interval`] of that window (a quarter of it,
//!   clamped to `[1s, window]`), the identical derivation the expiry sweeper
//!   uses, so a healthy connection is refreshed four times per window and the
//!   idle lease has no chance to expire;
//! - the window the worker is told to expect is the heartbeat window ITSELF, so
//!   both ends of the link declare death on exactly the same operator contract.
//!
//! Nothing here is separately configurable, and the worker holds no copy of the
//! window: it is carried on every ping.
//!
//! # Both transports, one probation (#197)
//!
//! This module is named for the transport it was built against, but
//! [`LivenessProbe`] is no longer liminal-only. A round enumerates liminal
//! connections AND gRPC-delivered registrations
//! ([`grpc_liveness`](super::grpc_liveness)), pings each over ITS OWN
//! transport, and feeds every answer into the SAME
//! [`HeartbeatTracker`] probation and the SAME
//! [`publish_reachability_verdict`](LivenessProbe::publish_reachability_verdict)
//! — one [`DISPATCH_PROBATION_PINGS`], one eligibility set, one set of
//! announcement transitions, no per-transport constant and no exemption.
//!
//! Only the wire differs: a liminal worker is pinged with the serde
//! [`LivenessPing`] on its connection, a gRPC worker with the protobuf
//! [`ProtoLivenessPing`](aion_proto::ProtoLivenessPing) on its task stream. The
//! fact each establishes is identical, which is why the verdict is.
//!
//! The accepted cost of riding the task stream is that stream death is not a
//! separate signal: a closed stream simply makes the NEXT ping unanswerable, so
//! withdrawal latency is bounded by the probe cadence rather than being
//! instantaneous. That is deliberate — a special-case fast path for stream
//! death would be a second, softer route into the eligibility verdict, and the
//! registry's own disconnect teardown already owns deregistration.
//!
//! # A verdict reaches only the transports this probe CARRIES (#25)
//!
//! The probation is transport-blind; the verdict's REACH is not. A probe
//! publishes an exclusion for a worker only if it holds a wire on which that
//! worker could have been asked — a liminal notifier for a liminal worker, a
//! gRPC answer registry for a gRPC one ([`LivenessProbe::covers`]). Without
//! that scope the verdict was formed from the whole of
//! [`HeartbeatTracker::unreachable_workers`], so a worker of an uncarried
//! transport could never bank an answer, never serve its probation, and never
//! leave the exclusion set: registered, healthy, and withheld from dispatch for
//! the life of the process. That was reported live against a server whose probe
//! covered liminal only, and it is what [`LivenessProbe::new`] would still do to
//! a gRPC fleet if the reach were left unscoped.
//!
//! Coverage is STRUCTURAL — which wires the probe holds — and never the set of
//! targets a round happened to enumerate. The difference is load-bearing: a
//! liminal worker whose connection has closed appears in no round's target list,
//! and it is exactly the worker that must stay excluded. Scoping by enumeration
//! would re-admit it, which is a worse defect than the one being cured.

use std::collections::BTreeMap;
use std::sync::Arc;
use std::time::{Duration, Instant};

use aion_core::WorkerTransport;
use serde::{Deserialize, Serialize};
use tokio::sync::watch;
use tracing::{info, warn};

use super::grpc_liveness::{GrpcLivenessTarget, GrpcLivenessWaiters, ping_grpc_worker};
use super::heartbeat::{
    DISPATCH_PROBATION_PINGS, DispatchExclusion, ExcludedWorker, HeartbeatTracker, sweep_interval,
};
use super::liminal_transport::{LiminalConnectionNotifier, LiminalWorkerDelivery};
use super::liveness::{PingFailure, ProbedTransport};
use super::registry::{ConnectedWorkerRegistry, WorkerId};

/// Wire liveness ping the server pushes on an established liminal connection.
///
/// Field-for-field mirror of `aion-worker`'s `liminal_liveness::LivenessPing`
/// (same serde field names), the same cross-crate contract the
/// dispatch/response and intervention pairs pin. `liveness_ping` is also the
/// worker's demux discriminator: no other pushed frame carries it, and a ping
/// carries none of the fields a dispatch or intervention requires.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct LivenessPing {
    /// Monotonic ping sequence within this connection, echoed on the answer.
    pub liveness_ping: u64,
    /// How long the worker may hear NOTHING on this connection before it must
    /// declare the link dead — this server's `worker.heartbeat_window`, carried
    /// on the wire so the worker never holds a second copy of it.
    pub silence_window_ms: u64,
}

/// Wire answer the worker replies with, echoing the ping's sequence.
///
/// Field-for-field mirror of `aion-worker`'s `liminal_liveness::LivenessPong`.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]
pub struct LivenessPong {
    /// The sequence of the ping being answered, echoed verbatim.
    pub liveness_pong: u64,
}

/// One connection the probe pings on a round: the connection pid, the worker it
/// registered, and the push leg to reach it.
#[derive(Clone, Debug)]
pub struct LivenessTarget {
    /// Liminal connection process id the worker is addressed on.
    pub pid: u64,
    /// Registry identity of the worker that registered on this connection.
    pub worker_id: WorkerId,
    /// Push leg used to deliver the ping and await its answer.
    pub delivery: LiminalWorkerDelivery,
}

/// The production driver of the liminal connection dead-man switch.
///
/// Shares the server's shutdown watch, so it drains with the transports exactly
/// like [`HeartbeatSweeper`](super::HeartbeatSweeper) and the outbox dispatcher.
pub struct LivenessProbe {
    /// The liminal connection census, or `None` on a boot that hosts no liminal
    /// listener. `None` is not "no liveness": the gRPC half below still runs.
    notifier: Option<Arc<LiminalConnectionNotifier>>,
    /// The gRPC answer-correlation registry, or `None` when this probe has no
    /// channel to receive gRPC answers on.
    ///
    /// Structural rather than a flag: a probe with no correlation registry
    /// cannot hear a gRPC answer, so it must not push gRPC pings either — every
    /// one would time out and hold healthy workers off dispatch. The production
    /// wiring
    /// ([`ServerState::spawn_liminal_liveness_probe`](crate::ServerState::spawn_liminal_liveness_probe))
    /// always supplies one; [`LivenessProbe::new`] is the liminal-only
    /// construction for callers with no gRPC fleet.
    grpc: Option<GrpcLivenessWaiters>,
    tracker: HeartbeatTracker,
    registry: ConnectedWorkerRegistry,
    cadence: Duration,
    silence_window: Duration,
}

impl std::fmt::Debug for LivenessProbe {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        formatter
            .debug_struct("LivenessProbe")
            .field("cadence", &self.cadence)
            .field("silence_window", &self.silence_window)
            .finish_non_exhaustive()
    }
}

impl LivenessProbe {
    /// Build a probe over the notifier that owns the liminal connections, the
    /// shared liveness tracker whose leases a pong refreshes, and the registry
    /// the WARN lines resolve a worker's queue through.
    ///
    /// Both timings derive from `heartbeat_window` (see the module docs); there
    /// is no separate configuration surface.
    /// This construction probes the LIMINAL transport only: it carries no gRPC
    /// answer-correlation registry, so it does not enumerate gRPC deliveries.
    /// Use [`Self::across_transports`] for the production probe.
    #[must_use]
    pub fn new(
        notifier: Arc<LiminalConnectionNotifier>,
        tracker: HeartbeatTracker,
        registry: ConnectedWorkerRegistry,
        heartbeat_window: Duration,
    ) -> Self {
        Self::across_transports(Some(notifier), None, tracker, registry, heartbeat_window)
    }

    /// Build the probe that covers EVERY transport a worker may be delivered
    /// over (#197).
    ///
    /// `notifier` is `None` on a boot with no liminal listener; `grpc` is the
    /// correlation registry the gRPC stream handler delivers answers into, and
    /// is `None` only for a probe that must not push gRPC pings it could never
    /// hear the answers to.
    ///
    /// There is exactly ONE probe per server. Two would each publish a whole
    /// eligibility set over the other's — the verdict is a replacement, not a
    /// merge — so the last writer would silently erase the other transport's
    /// findings every cadence.
    #[must_use]
    pub fn across_transports(
        notifier: Option<Arc<LiminalConnectionNotifier>>,
        grpc: Option<GrpcLivenessWaiters>,
        tracker: HeartbeatTracker,
        registry: ConnectedWorkerRegistry,
        heartbeat_window: Duration,
    ) -> Self {
        Self {
            notifier,
            grpc,
            tracker,
            registry,
            cadence: sweep_interval(heartbeat_window),
            silence_window: heartbeat_window,
        }
    }

    /// The interval between probe rounds.
    #[must_use]
    pub const fn cadence(&self) -> Duration {
        self.cadence
    }

    /// The silence window this probe declares to every worker it pings.
    #[must_use]
    pub const fn silence_window(&self) -> Duration {
        self.silence_window
    }

    /// Run the probe until `shutdown` flips to `true`.
    ///
    /// Rounds never overlap: each tick's pings are awaited to completion (each
    /// bounded by the cadence) before the next round starts, and a missed tick
    /// is skipped rather than queued.
    ///
    /// That bounds the concurrent WAITS to one per connection. It was once
    /// claimed to bound the outstanding PUSHES to one as well, "so it can never
    /// crowd out real dispatches against liminal's per-connection pending-push
    /// cap." **That claim was false and the failure it denied is exactly what
    /// happened** on run `dfd2117c`: a push slot is not released by the caller
    /// giving up, only by a consumed reply, a deadline expiry, or a connection
    /// close. Awaiting a round to completion ends the wait, not the slot. So
    /// abandoning one unanswered no-deadline ping per round leaked one slot per
    /// round — 32 of them, then total refusal of every push on that connection.
    ///
    /// The bound is now real because
    /// [`LiminalWorkerDelivery::push_payload_with_deadline`] attaches the
    /// cadence as the push's own reply deadline, so an unanswered ping's slot
    /// expires and RELEASES its cap admission instead of accumulating.
    pub async fn run(self, mut shutdown: watch::Receiver<bool>) {
        info!(
            cadence_ms = self.cadence.as_millis(),
            silence_window_ms = self.silence_window.as_millis(),
            "worker liveness probe started (liminal + grpc)"
        );
        let mut sequence = 0_u64;
        let mut ticks = tokio::time::interval(self.cadence);
        ticks.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
        loop {
            tokio::select! {
                _ = ticks.tick() => {
                    if *shutdown.borrow() {
                        break;
                    }
                    sequence = sequence.saturating_add(1);
                    self.probe_once(sequence).await;
                }
                changed = shutdown.changed() => {
                    // A receive error means every sender dropped; treat that as
                    // a shutdown request rather than spinning.
                    if changed.is_err() || *shutdown.borrow() {
                        break;
                    }
                }
            }
        }
        info!("worker liveness probe stopped");
    }

    /// Ping every reachable worker once — liminal connections and gRPC task
    /// streams alike, concurrently — apply each answer to the worker's dispatch
    /// probation, and publish ONE verdict for the round.
    ///
    /// The two transports are pinged over their own wires, CONCURRENTLY with
    /// each other, and are otherwise indistinguishable from here down: the same
    /// probation, the same tracker, the same single
    /// [`Self::publish_reachability_verdict`] call.
    ///
    /// EVERY round publishes, including one that found no target to ping. This
    /// module used to skip publication for an empty round, reasoning that
    /// "publishing an empty exclusion set would be indistinguishable from
    /// clearing one". They are the same act, and skipping it is what left a
    /// stale verdict standing: a verdict is published by REPLACEMENT, so a round
    /// that publishes nothing leaves the previous round's exclusions in force
    /// over a fleet this round measured afresh. The set is derived from the
    /// tracker's standing evidence rather than from this round's targets, so it
    /// is well defined whether or not anything was pinged (#25).
    ///
    /// Public because it is the whole unit of work: [`Self::run`] is a timer
    /// around it and adds nothing but the schedule. Driving rounds directly is
    /// how a test states a probation in ROUNDS rather than in wall-clock
    /// sleeps, which is the difference between asserting the law and asserting
    /// this machine's timing.
    ///
    /// `sequence` must advance between rounds: it is the echo a worker's answer
    /// is matched against, and repeating one would let a previous round's late
    /// answer satisfy this round's ping.
    pub async fn probe_once(&self, sequence: u64) {
        let liminal = self
            .notifier
            .as_ref()
            .map_or_else(Vec::new, |notifier| notifier.liveness_targets());
        let grpc = self.grpc_targets();
        let silence_window_ms = u64::try_from(self.silence_window.as_millis()).unwrap_or(u64::MAX);
        // Joined, not sequenced. Each half bounds its own pings by the cadence,
        // so awaiting them in turn would let a round of unanswered gRPC pings
        // push the NEXT liminal ping a full cadence late — stretching the gap
        // between a healthy liminal connection's refreshes toward three
        // quarters of the heartbeat window and eroding the four-refreshes-per-
        // window margin this module's timings are derived to guarantee. The
        // halves share no state until their answers are applied, so there is
        // nothing to serialize.
        tokio::join!(
            self.probe_liminal(sequence, silence_window_ms, liminal),
            self.probe_grpc(sequence, silence_window_ms, grpc),
        );
        self.publish_reachability_verdict(sequence);
    }

    /// This round's gRPC targets, or none when this probe carries no answer
    /// channel (see [`Self::grpc`]). A registry read failure is LOUD and yields
    /// no targets: the alternative is pinging a fleet the probe cannot describe.
    fn grpc_targets(&self) -> Vec<GrpcLivenessTarget> {
        let Some(_) = self.grpc.as_ref() else {
            return Vec::new();
        };
        match self.registry.grpc_liveness_targets() {
            Ok(targets) => targets,
            Err(error) => {
                warn!(
                    %error,
                    "could not enumerate gRPC-delivered workers for this liveness round; they \
                     go unprobed and will therefore not clear their dispatch probation"
                );
                Vec::new()
            }
        }
    }

    /// Push the serde ping to every liminal connection concurrently.
    async fn probe_liminal(
        &self,
        sequence: u64,
        silence_window_ms: u64,
        targets: Vec<LivenessTarget>,
    ) {
        if targets.is_empty() {
            return;
        }
        let ping = LivenessPing {
            liveness_ping: sequence,
            silence_window_ms,
        };
        let payload = match serde_json::to_vec(&ping) {
            Ok(payload) => payload,
            Err(error) => {
                // Structurally unreachable (two integers), but a probe that
                // cannot encode its own ping must say so rather than silently
                // stop being a dead-man switch.
                warn!(%error, "liminal liveness probe could not encode its ping; skipping round");
                return;
            }
        };
        let deadline = self.cadence;
        let answers = targets.into_iter().map(|target| {
            let payload = payload.clone();
            async move {
                let outcome = tokio::task::spawn_blocking(move || {
                    ping_one(&target.delivery, payload, deadline)
                })
                .await;
                (target.pid, target.worker_id, outcome)
            }
        });
        for (pid, worker_id, outcome) in futures::future::join_all(answers).await {
            self.apply_answer(pid, worker_id, sequence, outcome);
        }
    }

    /// Push the protobuf ping down every gRPC task stream concurrently.
    async fn probe_grpc(
        &self,
        sequence: u64,
        silence_window_ms: u64,
        targets: Vec<GrpcLivenessTarget>,
    ) {
        let (Some(waiters), false) = (self.grpc.as_ref(), targets.is_empty()) else {
            return;
        };
        let deadline = self.cadence;
        let answers = targets.iter().map(|target| async move {
            let outcome = ping_grpc_worker(
                waiters,
                target,
                aion_proto::ProtoLivenessPing {
                    liveness_ping: sequence,
                    silence_window_ms,
                },
                deadline,
            )
            .await;
            (target.worker_id, outcome)
        });
        for (worker_id, outcome) in futures::future::join_all(answers).await {
            self.apply_grpc_answer(worker_id, sequence, outcome);
        }
    }

    /// Publish this round's reachability verdict to the registry, so dispatch
    /// selection skips workers the server cannot reach.
    ///
    /// Runs after every round, including rounds where every ping succeeded —
    /// that is what RESTORES eligibility to a worker whose pings have started
    /// answering again. Recovery must not need a separate trigger.
    ///
    /// This is the half that makes the switch able to fire at all. A worker's
    /// connection lease is refreshed by anything it sends, including its own
    /// background liveness pump, so a worker whose dispatch path is completely
    /// dead can look perfectly alive indefinitely. Reachability is tracked
    /// separately and only an answered ping advances it, so the pump can no
    /// longer hold dispatch eligibility open against failing pings.
    ///
    /// # What this says out loud, and why it changed
    ///
    /// Until 2026-08-05 this announced exactly one transition — the withdrawal
    /// — in one sentence that was FALSE in the commonest case. Every healthy
    /// worker start logged `WITHDRAWING DISPATCH ELIGIBILITY … the server has
    /// not been able to reach its dispatch path within the window`, because
    /// registration opens an unserved probation and the first round after it
    /// always finds the probation unserved. At that instant the server had
    /// reached the worker — one answer was already banked — and the window had
    /// nothing to do with it. The remedy sentence was wrong too: it promised
    /// eligibility back after "one ping", when
    /// [`DISPATCH_PROBATION_PINGS`](super::heartbeat::DISPATCH_PROBATION_PINGS)
    /// consecutive answers are required and one was already in hand.
    ///
    /// The restoration a few seconds later was silent, so an operator saw the
    /// alarm and never the all-clear. On Tom's server that read as a broken
    /// worker and was reported to him as a caveat on a fix that was in fact
    /// working. An alarm that fires on every ordinary connect carries no
    /// information; a resolution nobody announces cannot cancel it.
    ///
    /// So the three transitions are now distinguished and all three are said:
    /// the probation opening (ordinary, INFO), the loss of eligibility that was
    /// actually held (an incident, WARN), and the recovery (INFO).
    ///
    /// # SATURATION IS NO LONGER MEASURED AS SILENCE
    ///
    /// This WARN used to carry a known false red, and it is struck: a gRPC
    /// worker running at its full concurrency stopped reading its task stream,
    /// because the receive loop awaited a concurrency permit before reading the
    /// next frame. A ping already on the wire went unanswered until a permit
    /// freed, so a worker saturated for longer than one probe cadence looked
    /// silent while it was in fact working, and this line said its dispatch
    /// path was unreachable.
    ///
    /// Both halves of that are now false, and by two independent changes:
    ///
    /// - The worker's admission never parks its receive loop. The permit is
    ///   taken without awaiting, so a saturated worker keeps reading frames and
    ///   answers this probe's ping while its work runs. SATURATION is no longer
    ///   a way to produce an unanswered ping.
    ///
    ///   Not the same as saying an unanswered ping now means only one thing, and
    ///   it does not: [`PingFailure::Unaskable`] still fires when a session's
    ///   push channel is full, and a worker whose runtime is occupied by
    ///   blocking handlers can still miss one. Those are different conditions
    ///   with their own causes; what changed is that a worker running exactly as
    ///   many activities as it advertised is no longer one of them.
    /// - The server does not push a worker past the concurrency it advertised
    ///   at registration. Selection counts each worker's in-flight dispatches
    ///   against its own number, so the condition that produced the false red
    ///   is not reached by the ordinary route at all.
    ///
    /// The old note also carried a bounding claim — "in-flight work is
    /// untouched throughout" — which was true of THIS verdict and false of the
    /// composed system. Eligibility withdrawal touched nothing, but the
    /// heartbeat sweep read the same silence and DEREGISTERED the worker,
    /// surfacing its live work as transport losses onto a queue the census then
    /// reported as having no worker connected. That gap is closed at the sweep:
    /// deregistration now requires an explicit connected fact to be absent, not
    /// merely for noise to have stopped.
    ///
    /// What is unchanged is this verdict itself. Withdrawing eligibility from a
    /// worker the server cannot prove it reaches is correct, and it remains the
    /// resting state for one: ineligible, still registered, still holding its
    /// in-flight work, restored after its probation the moment it answers.
    fn publish_reachability_verdict(&self, sequence: u64) {
        let now = Instant::now();
        let measured = match self.tracker.unreachable_workers(now) {
            Ok(workers) => workers,
            Err(error) => {
                warn!(
                    %error,
                    liveness_ping = sequence,
                    "could not read worker reachability; leaving the previous dispatch \
                     eligibility verdict in place rather than guessing"
                );
                return;
            }
        };
        let Some(unreachable) = self.within_reach(&measured) else {
            warn!(
                liveness_ping = sequence,
                "could not read which transport each unreachable worker is delivered over, so \
                 this round cannot tell an exclusion it is entitled to publish from one it has \
                 no wire to test; the previous verdict stands unchanged"
            );
            return;
        };
        // 🔴 The exclusion travels WITH the worker id into the published
        // verdict. It used to be dropped here — `.map(|e| e.worker_id)` into a
        // bare set — which left the registry unable to tell a worker serving
        // its opening probation from one whose reachability was lost. Those are
        // the two facts `DispatchExclusion`'s own documentation calls facts "an
        // operator must be able to tell apart", and collapsing them is what let
        // a fan-out row park on a non-clearing exclusion with nothing published
        // about why. The value was already in hand; only this line threw it out.
        let excluded_now: BTreeMap<WorkerId, DispatchExclusion> = unreachable
            .iter()
            .map(|excluded| (excluded.worker_id, excluded.exclusion))
            .collect();
        // The previous verdict is what makes a transition a transition. If it
        // cannot be read the announcements are SKIPPED rather than guessed —
        // assuming an empty previous set would re-announce every standing
        // exclusion as though it had just happened. The verdict itself still
        // publishes below: gating dispatch is the load-bearing half, and it
        // must not be dropped because the narration failed.
        match self.registry.dispatch_ineligible() {
            Ok(previously_excluded) => {
                self.announce_transitions(
                    sequence,
                    &previously_excluded,
                    &unreachable,
                    &excluded_now,
                );
            }
            Err(error) => warn!(
                %error,
                liveness_ping = sequence,
                "could not read the published liminal dispatch eligibility set; this round's \
                 eligibility changes go UNANNOUNCED, though the verdict itself is still published"
            ),
        }
        if let Err(error) = self.registry.set_dispatch_ineligible(excluded_now) {
            warn!(
                %error,
                liveness_ping = sequence,
                "could not publish worker dispatch eligibility; selection keeps the \
                 previous verdict"
            );
        }
    }

    /// Narrow the tracker's evidence to the workers this probe is ENTITLED to
    /// judge: those delivered over a transport it carries a wire for (#25).
    ///
    /// A worker of an uncarried transport is dropped — not because it is
    /// reachable (nothing here knows that) but because this probe has no way to
    /// find out, and an exclusion it could never lift is a permanent withdrawal
    /// dressed as a measurement. A worker that has left the registry is dropped
    /// for the same reason: it has no transport at all, selection can no longer
    /// see it, and naming it would put a fabricated fact in an operator-facing
    /// set.
    ///
    /// `None` when the registry cannot be read: the caller must then publish
    /// NOTHING. Falling back to the unscoped set would reinstate the defect at
    /// exactly the moment the server is least able to explain itself.
    fn within_reach(&self, measured: &[ExcludedWorker]) -> Option<Vec<ExcludedWorker>> {
        let transports = self
            .registry
            .transports_of(measured.iter().map(|excluded| excluded.worker_id))
            .map_err(|error| warn!(%error, "could not read worker transports"))
            .ok()?;
        Some(
            measured
                .iter()
                .filter(|excluded| {
                    transports
                        .get(&excluded.worker_id)
                        .is_some_and(|transport| self.covers(*transport))
                })
                .copied()
                .collect(),
        )
    }

    /// Whether this probe holds a wire on which a worker of `transport` could be
    /// asked its liveness question.
    ///
    /// Structural, and deliberately not a function of how many targets a round
    /// enumerated: a liminal worker whose connection has closed is enumerated by
    /// nothing and is precisely the worker a liminal-carrying probe must keep
    /// excluded.
    const fn covers(&self, transport: WorkerTransport) -> bool {
        match transport {
            WorkerTransport::Grpc => self.grpc.is_some(),
            WorkerTransport::Liminal => self.notifier.is_some(),
        }
    }

    /// Say what changed this round — and only what changed, so a persistently
    /// excluded worker does not re-log every cadence.
    fn announce_transitions(
        &self,
        sequence: u64,
        previously_excluded: &BTreeMap<WorkerId, DispatchExclusion>,
        unreachable: &[ExcludedWorker],
        excluded_now: &BTreeMap<WorkerId, DispatchExclusion>,
    ) {
        for excluded in unreachable {
            let Some(announcement) = transition(
                previously_excluded.contains_key(&excluded.worker_id),
                Some(excluded.exclusion),
            ) else {
                continue;
            };
            self.say(sequence, excluded.worker_id, announcement);
        }
        for worker_id in previously_excluded
            .keys()
            .filter(|worker_id| !excluded_now.contains_key(worker_id))
        {
            // A worker that DEPARTED is not a worker that recovered. The
            // registry read is the discriminator: an entry no longer in it left
            // the fleet, and announcing its recovery would be a fabrication.
            let Some(task_queue) = self.task_queue_of(*worker_id) else {
                continue;
            };
            let Some(announcement) = transition(true, None) else {
                continue;
            };
            self.say_with_queue(sequence, *worker_id, &task_queue, announcement);
        }
    }

    /// Emit one announcement, resolving the worker's queue for the line.
    fn say(&self, sequence: u64, worker_id: WorkerId, announcement: Announcement) {
        let task_queue = self
            .task_queue_of(worker_id)
            .unwrap_or_else(|| "<unregistered>".to_owned());
        self.say_with_queue(sequence, worker_id, &task_queue, announcement);
    }

    /// Emit one announcement against an already-resolved queue.
    fn say_with_queue(
        &self,
        sequence: u64,
        worker_id: WorkerId,
        task_queue: &str,
        announcement: Announcement,
    ) {
        let transport = self.transport_of(worker_id);
        match announcement {
            Announcement::ProbationOpened { answers } => info!(
                worker_id = worker_id.value(),
                task_queue,
                transport,
                liveness_ping = sequence,
                answers_banked = answers,
                answers_required = DISPATCH_PROBATION_PINGS,
                "worker is SERVING ITS DISPATCH PROBATION: it has answered {answers} of \
                 {DISPATCH_PROBATION_PINGS} consecutive liveness pings since it connected, and is \
                 not selected for dispatch until the run is complete. This is the ordinary cost of \
                 connecting — every healthy worker start passes through it, on every transport — \
                 not a fault, and not a statement that anything is unreachable"
            ),
            Announcement::EligibilityWithdrawn => warn!(
                worker_id = worker_id.value(),
                task_queue,
                transport,
                liveness_ping = sequence,
                answers_required = DISPATCH_PROBATION_PINGS,
                silence_window_ms = self.silence_window.as_millis(),
                "WITHDRAWING DISPATCH ELIGIBILITY from worker: it had PROVED its dispatch path \
                 reachable on this connection and the server can no longer prove it — either a \
                 liveness ping failed or the last proof aged out of the window. It stays \
                 registered and keeps its in-flight work, and becomes eligible again after \
                 {DISPATCH_PROBATION_PINGS} consecutive answered pings"
            ),
            Announcement::EligibilityRestored => info!(
                worker_id = worker_id.value(),
                task_queue,
                transport,
                liveness_ping = sequence,
                "DISPATCH ELIGIBILITY RESTORED to worker: it has answered a full run of \
                 consecutive liveness pings, so the server can again prove it reaches this \
                 worker's dispatch path. Dispatch selection includes it from now"
            ),
        }
    }

    /// The transport a worker is delivered over, for the announcement line.
    ///
    /// `"<unregistered>"` when the worker has left the registry or the registry
    /// cannot be read — a departed worker's transport is genuinely unknowable
    /// here, and naming one would be a fabrication.
    fn transport_of(&self, worker_id: WorkerId) -> &'static str {
        match self.registry.worker_by_id(worker_id) {
            Ok(Some(handle)) => handle.delivery().transport().name(),
            Ok(None) | Err(_) => "<unregistered>",
        }
    }

    /// Apply one connection's ping outcome: a correct answer advances dispatch
    /// reachability, anything else is logged LOUDLY, proves nothing, and RESETS
    /// the probation — proof of reachability must be a consecutive run.
    fn apply_answer(
        &self,
        pid: u64,
        worker_id: WorkerId,
        sequence: u64,
        outcome: Result<Result<LivenessPong, PingFailure>, tokio::task::JoinError>,
    ) {
        let failure = match outcome {
            Err(join_error) => {
                PingFailure::Unanswered(format!("ping task failed to run: {join_error}"))
            }
            Ok(Err(failure)) => failure,
            Ok(Ok(pong)) if pong.liveness_pong != sequence => PingFailure::Unanswered(format!(
                "worker answered with mismatched sequence {}",
                pong.liveness_pong
            )),
            Ok(Ok(_)) => {
                self.record_reachable(ProbedTransport::Liminal { pid }, worker_id);
                return;
            }
        };
        self.record_unreachable(
            ProbedTransport::Liminal { pid },
            worker_id,
            sequence,
            failure,
        );
    }

    /// The gRPC counterpart of [`Self::apply_answer`] (#197).
    ///
    /// Deliberately the same two outcomes fed into the same two recorders: the
    /// only gRPC-specific step is that a mismatched echo cannot reach here at
    /// all, because the correlation registry
    /// ([`GrpcLivenessWaiters::answer`](super::grpc_liveness::GrpcLivenessWaiters::answer))
    /// refuses to match one — so an echo that arrives is by construction the
    /// sequence that was sent.
    fn apply_grpc_answer(
        &self,
        worker_id: WorkerId,
        sequence: u64,
        outcome: Result<u64, PingFailure>,
    ) {
        match outcome {
            Ok(echoed) if echoed == sequence => {
                self.record_reachable(ProbedTransport::Grpc, worker_id);
            }
            Ok(echoed) => self.record_unreachable(
                ProbedTransport::Grpc,
                worker_id,
                sequence,
                PingFailure::Unanswered(format!(
                    "worker answered with mismatched sequence {echoed}"
                )),
            ),
            Err(failure) => {
                self.record_unreachable(ProbedTransport::Grpc, worker_id, sequence, failure);
            }
        }
    }

    /// Bank one answered ping: the ONE thing that proves the server can reach
    /// this worker's DISPATCH path.
    ///
    /// Only an answer advances reachability — an inbound frame proves the
    /// opposite direction and cannot stand in for it. A `false` return from the
    /// tracker means the worker was already deregistered (an answer racing a
    /// reap); an answer must never resurrect it.
    fn record_reachable(&self, transport: ProbedTransport, worker_id: WorkerId) {
        if let Err(error) = self
            .tracker
            .record_dispatch_reachability(worker_id, Instant::now())
        {
            warn!(
                %error,
                connection_pid = transport.pid(),
                transport = transport.name(),
                worker_id = worker_id.value(),
                "failed to record worker dispatch reachability from a liveness answer"
            );
        }
    }

    /// Record one FAILED ping and say which of the two failures it was.
    ///
    /// A probation is CONSECUTIVE, so any failure resets it to zero. Without
    /// this the counter would be cumulative, and a link that answers one probe
    /// in three would still accrue its way to eligibility and then flap in and
    /// out of it forever — which is exactly the defect the probation exists to
    /// stop. Both failure classes reset: whether we could not ask or the worker
    /// did not answer, the run of answers is broken either way. `Ok(false)`
    /// means the worker was already deregistered; nothing to reset.
    fn record_unreachable(
        &self,
        transport: ProbedTransport,
        worker_id: WorkerId,
        sequence: u64,
        failure: PingFailure,
    ) {
        if let Err(error) = self.tracker.record_dispatch_unreachable(worker_id) {
            warn!(
                %error,
                connection_pid = transport.pid(),
                transport = transport.name(),
                worker_id = worker_id.value(),
                "failed to reset worker dispatch probation after a failed liveness ping; \
                 its eligibility may outlive the proof that earned it"
            );
        }
        // The two failures are DIFFERENT FACTS and the operator must be able to
        // tell them apart: one is about the worker, the other is about us.
        let task_queue = self
            .task_queue_of(worker_id)
            .unwrap_or_else(|| "<unregistered>".to_owned());
        let transport_name = transport.name();
        match failure {
            PingFailure::Unaskable(reason) => warn!(
                connection_pid = transport.pid(),
                transport = transport_name,
                worker_id = worker_id.value(),
                task_queue = %task_queue,
                liveness_ping = sequence,
                reason = %reason,
                silence_window_ms = self.silence_window.as_millis(),
                "THE SERVER COULD NOT ASK this {transport_name} worker for liveness — the push \
                 itself was refused, so nothing was sent and the worker has no idea it was probed. \
                 A dispatch would be refused by the same channel for the same reason. This says \
                 nothing about whether the worker is healthy; it says this server cannot currently \
                 reach it. Its dispatch eligibility is withdrawn NOW and it must answer a full run \
                 of consecutive pings to earn it back"
            ),
            PingFailure::Unanswered(reason) => warn!(
                connection_pid = transport.pid(),
                transport = transport_name,
                worker_id = worker_id.value(),
                task_queue = %task_queue,
                liveness_ping = sequence,
                reason = %reason,
                silence_window_ms = self.silence_window.as_millis(),
                "{transport_name} worker did not answer its liveness ping; the server could not \
                 prove it can reach this worker's dispatch path, so its dispatch eligibility is \
                 withdrawn NOW and it must answer a full run of consecutive pings to earn it back \
                 — an unanswered probe is direct evidence about the push leg, not mere silence. \
                 NOTE: the worker's connection lease may still be fresh — its liveness pump beats \
                 from a background task and keeps proving the process is alive — so do NOT expect \
                 an expiry sweep to reap it"
            ),
        }
    }

    /// The task queue a worker is registered on, for the WARN line. `None` when
    /// the worker is no longer in the registry (already reaped).
    fn task_queue_of(&self, worker_id: WorkerId) -> Option<String> {
        self.registry
            .worker_by_id(worker_id)
            .ok()
            .flatten()
            .map(|handle| handle.task_queue().to_owned())
    }
}

/// Push one ping and block for its correlated answer, bounded by `deadline`.
///
/// Runs on a blocking thread: the liminal push/await pair is thread-based, not
/// async. Every failure is rendered as a reason string, because at this layer
/// the distinction that matters is "answered" vs "did not answer" — the typed
/// error text rides into the WARN verbatim.
/// A change in one worker's dispatch standing that the operator must be told
/// about — never the standing state itself, so a persistently excluded worker
/// does not re-log every cadence.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum Announcement {
    /// A freshly connected worker began serving its probation. Ordinary.
    ProbationOpened {
        /// Consecutive answers banked when the probation was announced.
        answers: u32,
    },
    /// A worker that HELD eligibility lost it. An incident.
    EligibilityWithdrawn,
    /// A worker that was excluded is dispatchable again.
    EligibilityRestored,
}

/// The whole truth table for one worker's standing between two rounds.
///
/// Pure and total on purpose: the four inputs are exhaustively enumerated in
/// the tests, which is the only way to be sure the alarming half and the
/// reassuring half are both reachable. The old code had no such function — the
/// decision was inlined and only ever produced one of the three lines, so the
/// missing two were invisible.
const fn transition(
    was_excluded: bool,
    now_excluded: Option<DispatchExclusion>,
) -> Option<Announcement> {
    match (was_excluded, now_excluded) {
        // Newly excluded. WHICH exclusion decides whether this is news.
        (false, Some(DispatchExclusion::OpeningProbation { answers })) => {
            Some(Announcement::ProbationOpened { answers })
        }
        (false, Some(DispatchExclusion::ReachabilityLost)) => {
            Some(Announcement::EligibilityWithdrawn)
        }
        // Left the exclusion set: the all-clear.
        (true, None) => Some(Announcement::EligibilityRestored),
        // No CHANGE in standing, by either route: a worker still excluded (its
        // exclusion was announced when it began, and repeating it every cadence
        // is how a log stops being read), or one that was eligible and stayed
        // eligible. Both are silence, for different reasons.
        (true, Some(_)) | (false, None) => None,
    }
}

fn ping_one(
    delivery: &LiminalWorkerDelivery,
    payload: Vec<u8>,
    deadline: Duration,
) -> Result<LivenessPong, PingFailure> {
    // The deadline is attached to the PUSH, not just to the wait. Without it the
    // reply slot is reclaimed only by a consumed reply or a connection close, so
    // abandoning an unanswered ping every cadence leaks one slot per round until
    // the connection's push cap is exhausted and nothing — ping, dispatch or
    // intervention — can be pushed to that worker again.
    let awaiter = delivery
        .push_payload_with_deadline(payload, deadline)
        .map_err(|error| PingFailure::Unaskable(error.to_string()))?;
    let reply = awaiter
        .receive(deadline)
        .map_err(|error| PingFailure::Unanswered(format!("no answer arrived: {error}")))?;
    serde_json::from_slice(&reply)
        .map_err(|error| PingFailure::Unanswered(format!("answer could not be decoded: {error}")))
}

#[cfg(test)]
mod tests {
    use crate::namespace::NamespaceGuard;
    use std::sync::Arc;
    use std::time::{Duration, Instant};

    use super::super::heartbeat::{DISPATCH_PROBATION_PINGS, HeartbeatTracker};
    use super::super::liminal_transport::LiminalConnectionNotifier;
    use super::super::registry::{ConnectedWorkerRegistry, WorkerId};
    use super::{
        Announcement, DispatchExclusion, LivenessPing, LivenessPong, LivenessProbe, PingFailure,
        WorkerTransport, transition,
    };

    /// The WHOLE truth table, enumerated. Four inputs, and every one of them is
    /// asserted here — which is the only way to be sure both the alarming and
    /// the reassuring outcomes are reachable.
    ///
    /// The version of this logic that shipped until 2026-08-05 could emit
    /// exactly one of these lines. The other two were not wrong; they did not
    /// exist, so an operator saw the withdrawal on every healthy worker start
    /// and never saw the recovery that followed seconds later.
    /// The runtime a test notifier's admission is bound to. These tests never
    /// register a worker through the notifier, so the runtime is only ever
    /// named, never driven.
    fn test_admission_runtime() -> std::io::Result<tokio::runtime::Runtime> {
        tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
    }

    #[test]
    fn every_standing_change_has_exactly_one_announcement() {
        assert_eq!(
            transition(
                false,
                Some(DispatchExclusion::OpeningProbation { answers: 1 })
            ),
            Some(Announcement::ProbationOpened { answers: 1 }),
            "a fresh connection serving its probation is ORDINARY and must not be announced as a \
             worker the server cannot reach"
        );
        assert_eq!(
            transition(false, Some(DispatchExclusion::ReachabilityLost)),
            Some(Announcement::EligibilityWithdrawn),
            "losing eligibility that was actually held is the incident the WARN exists for"
        );
        assert_eq!(
            transition(true, None),
            Some(Announcement::EligibilityRestored),
            "the all-clear must be said out loud — an alarm nobody cancels is read as ongoing"
        );
        assert_eq!(
            transition(true, Some(DispatchExclusion::ReachabilityLost)),
            None,
            "a standing exclusion must not re-log every cadence"
        );
        assert_eq!(
            transition(
                true,
                Some(DispatchExclusion::OpeningProbation { answers: 0 })
            ),
            None,
            "including while a still-excluded worker is still serving its probation"
        );
        assert_eq!(
            transition(false, None),
            None,
            "an eligible worker that stayed eligible is not news"
        );
    }

    /// THE WHOLE COVERAGE TABLE, enumerated (#25). Two transports by two
    /// wire-holdings, and every cell asserted.
    ///
    /// Coverage is what decides whether this probe may publish an exclusion for
    /// a worker at all, so a cell that silently answered `true` would hand a
    /// permanent withdrawal to a fleet the probe has no way to ask — the whole
    /// injury. Enumerating it here is the only way to be sure the deny half and
    /// the admit half are BOTH reachable: a `covers` that returned `true`
    /// unconditionally is exactly the pre-fix behaviour, and it satisfies any
    /// test that only checks the admitting direction.
    #[test]
    fn a_probe_covers_exactly_the_transports_it_holds_a_wire_for() -> std::io::Result<()> {
        let admission_runtime = test_admission_runtime()?;
        let window = Duration::from_secs(30);
        let registry = ConnectedWorkerRegistry::default();
        let tracker = HeartbeatTracker::new(window);
        let notifier = || {
            Arc::new(
                LiminalConnectionNotifier::new(registry.clone()).with_admission(
                    NamespaceGuard::shared_engine(),
                    false,
                    admission_runtime.handle().clone(),
                ),
            )
        };
        let build = |liminal: bool, grpc: bool| {
            LivenessProbe::across_transports(
                liminal.then(notifier),
                grpc.then(super::GrpcLivenessWaiters::new),
                tracker.clone(),
                registry.clone(),
                window,
            )
        };

        let both = build(true, true);
        assert!(both.covers(WorkerTransport::Liminal));
        assert!(both.covers(WorkerTransport::Grpc));

        let liminal_only = build(true, false);
        assert!(liminal_only.covers(WorkerTransport::Liminal));
        assert!(
            !liminal_only.covers(WorkerTransport::Grpc),
            "a probe holding no gRPC answer registry cannot hear a gRPC answer, so it must not \
             judge a gRPC worker: the exclusion it would publish could never be lifted"
        );

        let grpc_only = build(false, true);
        assert!(
            !grpc_only.covers(WorkerTransport::Liminal),
            "and the same in the other direction — the reach follows the wire, not the fleet"
        );
        assert!(grpc_only.covers(WorkerTransport::Grpc));

        // Not constructible in production (the composition root always supplies
        // at least the liminal notifier), but the fourth cell of a truth table
        // is not optional: a probe with no wire at all judges nobody.
        let neither = build(false, false);
        assert!(!neither.covers(WorkerTransport::Liminal));
        assert!(!neither.covers(WorkerTransport::Grpc));
        Ok(())
    }

    /// The ping/pong pair round-trips with stable field names — the cross-crate
    /// wire contract with `aion-worker`'s mirror of these types. A drift here is
    /// a wire break, so the exact JSON is pinned.
    #[test]
    fn the_liveness_pair_round_trips_through_json() -> Result<(), serde_json::Error> {
        let ping = LivenessPing {
            liveness_ping: 7,
            silence_window_ms: 30_000,
        };
        let encoded = serde_json::to_string(&ping)?;
        assert_eq!(encoded, r#"{"liveness_ping":7,"silence_window_ms":30000}"#);
        assert_eq!(serde_json::from_str::<LivenessPing>(&encoded)?, ping);

        let answer = LivenessPong { liveness_pong: 7 };
        let encoded = serde_json::to_string(&answer)?;
        assert_eq!(encoded, r#"{"liveness_pong":7}"#);
        assert_eq!(serde_json::from_str::<LivenessPong>(&encoded)?, answer);
        Ok(())
    }

    /// A ping decodes as NEITHER of the other two frames that share the push
    /// channel, and neither of them decodes as a ping — the demux contract the
    /// worker's serve loop relies on.
    #[test]
    fn a_liveness_ping_is_disjoint_from_the_other_pushed_frames() -> Result<(), serde_json::Error> {
        let ping = serde_json::to_vec(&LivenessPing {
            liveness_ping: 1,
            silence_window_ms: 1_000,
        })?;
        assert!(
            serde_json::from_slice::<super::super::liminal_transport::DispatchRequest>(&ping)
                .is_err(),
            "a liveness ping must never decode as a dispatch"
        );
        assert!(
            serde_json::from_slice::<super::super::liminal_transport::InterventionRequest>(&ping)
                .is_err(),
            "a liveness ping must never decode as an intervention"
        );
        Ok(())
    }

    /// Both timings derive from the operator's heartbeat window, with no
    /// separate knob: the cadence is the sweeper's quarter-window derivation and
    /// the declared silence window is the heartbeat window itself.
    #[test]
    fn probe_timings_derive_from_the_heartbeat_window() {
        let window = Duration::from_secs(30);
        assert_eq!(super::sweep_interval(window), Duration::from_millis(7_500));
        // The declared window IS the operator's window; the cadence divides it,
        // so a healthy connection is refreshed four times per window.
        assert!(super::sweep_interval(window) * 4 <= window);
    }

    /// 🔴 THE WIRING PIN. The probation lives in the tracker, but only the probe
    /// can tell it a ping FAILED — and a probe that recorded successes and
    /// dropped failures would compile, log its WARN lines exactly as it does
    /// now, and leave the probation permanently unreset. The eligibility bug
    /// would be silently back with every test in `heartbeat` still green,
    /// because those tests drive the tracker directly and never go through this
    /// seam.
    ///
    /// Every failure class is asserted individually. The fourth arm — a
    /// `JoinError` from the ping task — converges on the same
    /// `PingFailure::Unanswered` path these two take, one line above the reset.
    #[test]
    fn every_failed_probe_class_withdraws_dispatch_eligibility() -> std::io::Result<()> {
        let admission_runtime = test_admission_runtime()?;
        let window = Duration::from_secs(30);
        let worker = WorkerId::from_value(1);
        let start = Instant::now();

        // Each class gets its own probe and its own served probation, so a class
        // cannot pass by inheriting the withdrawal an earlier class performed.
        let failures = [
            (
                "the push was refused, so nothing was even asked",
                Ok(Err(PingFailure::Unaskable("push refused".to_owned()))),
            ),
            (
                "the ping was sent and no answer came back",
                Ok(Err(PingFailure::Unanswered("no answer arrived".to_owned()))),
            ),
            (
                "an answer came back carrying the wrong sequence",
                Ok(Ok(LivenessPong { liveness_pong: 99 })),
            ),
        ];

        for (class, outcome) in failures {
            let registry = ConnectedWorkerRegistry::default();
            let tracker = HeartbeatTracker::new(window);
            let probe = LivenessProbe::new(
                Arc::new(
                    LiminalConnectionNotifier::new(registry.clone()).with_admission(
                        NamespaceGuard::shared_engine(),
                        false,
                        admission_runtime.handle().clone(),
                    ),
                ),
                tracker.clone(),
                registry,
                window,
            );

            assert!(
                tracker.register_connection(worker, start).is_ok(),
                "tracker registration must succeed for {class}"
            );
            for _ in 0..DISPATCH_PROBATION_PINGS {
                assert!(
                    tracker
                        .record_dispatch_reachability(worker, start)
                        .is_ok_and(|tracked| tracked),
                    "the worker serves its probation before {class}"
                );
            }
            assert!(
                tracker
                    .is_dispatch_reachable(worker, start)
                    .is_ok_and(|reachable| reachable),
                "precondition: the worker is eligible before {class}"
            );

            // Sequence 1 is the ping that was sent; the mismatched-answer case
            // deliberately answers 99.
            probe.apply_answer(7, worker, 1, outcome);

            assert!(
                tracker
                    .is_dispatch_reachable(worker, start)
                    .is_ok_and(|reachable| !reachable),
                "the probe must withdraw dispatch eligibility when {class} — otherwise the \
                 probation never resets and a one-way link keeps its eligibility forever"
            );
        }
        Ok(())
    }

    /// The control for the pin above: the SUCCESS path through the same seam
    /// must keep eligibility. Without it, a probe that withdrew eligibility on
    /// every outcome — including healthy answers — would satisfy every
    /// assertion above and strand every worker on the fleet.
    #[test]
    fn an_answered_probe_keeps_dispatch_eligibility_through_the_same_seam() -> std::io::Result<()> {
        let admission_runtime = test_admission_runtime()?;
        let window = Duration::from_secs(30);
        let worker = WorkerId::from_value(1);
        let start = Instant::now();
        let registry = ConnectedWorkerRegistry::default();
        let tracker = HeartbeatTracker::new(window);
        let probe = LivenessProbe::new(
            Arc::new(
                LiminalConnectionNotifier::new(registry.clone()).with_admission(
                    NamespaceGuard::shared_engine(),
                    false,
                    admission_runtime.handle().clone(),
                ),
            ),
            tracker.clone(),
            registry,
            window,
        );

        assert!(tracker.register_connection(worker, start).is_ok());
        for sequence in 1..=u64::from(DISPATCH_PROBATION_PINGS) {
            probe.apply_answer(
                7,
                worker,
                sequence,
                Ok(Ok(LivenessPong {
                    liveness_pong: sequence,
                })),
            );
        }

        assert!(
            tracker
                .is_dispatch_reachable(worker, start)
                .is_ok_and(|reachable| reachable),
            "answered probes must EARN eligibility through the probe seam, not merely fail to \
             withdraw it"
        );
        Ok(())
    }
}