onlyne-client 2.0.0

use super::*;

use super::outbound::store_ack;
use super::projection::stored_task_state;
use super::state::{
    DispatchInner, DispatchState, forget_tools_binding, has_attached_transport, session_exited,
    slot_key_serving_task,
};
use super::transport::{held_read_only, names_session};

/// Reason a session that stopped answering is closed with.
///
/// The wire word the ledger and `onlyne sessions` read for a death this client
/// judged: the reconnect sweep stamps it on the refusal ack that buries the
/// session's delivery, and an operator's own
/// `repair fail --reason session_dead` writes the same word.
pub const SESSION_DEAD: &str = "session_dead";

/// Retire one session. The stored tuple decides whether a live resource
/// remains to close, and the caller's reason reaches the backend unchanged, so an
/// operator cancel stops reporting itself as a completion.
pub fn on_recycled(
    state: &DispatchState,
    task_id: &str,
    reason: crate::backend::CloseReason,
) -> Result<()> {
    let mut inner = state.inner.lock();
    release_locked(&mut inner, task_id, Some(reason))
}

/// Why the task of one session ended, as the retirement reason the task table
/// records.
///
/// The task's own row is the only place a verdict lives: the session tuple says
/// nothing about how its work ended, and an open task — `pending`, or no row at
/// all, which is the same reading — has no reason to retire anything.
pub(super) fn stored_close_reason(
    inner: &DispatchInner,
    task_id: &str,
) -> Option<crate::backend::CloseReason> {
    match stored_task_state(inner, task_id) {
        TaskState::Pending => None,
        TaskState::Done => Some(crate::backend::CloseReason::Completed),
        // A blocked delivery leaves the work owed, which is what a fault reason
        // names here, exactly as it does for a failed one.
        TaskState::Failed | TaskState::Blocked => Some(crate::backend::CloseReason::Fault),
        TaskState::Cancelled => Some(crate::backend::CloseReason::Cancelled),
    }
}

/// The reason one session the reconnect grace retires is closed with, read from
/// what this client holds rather than from a settle that never came.
///
/// The id is the session's own (`slot.session.task_id`), never the task binding
/// beside it: the sweep is what feeds that session's row and closes the resource
/// its agent was holding, and the two ids part company exactly where a slot's
/// binding is not the session it was born for. A session still owing work closes
/// as that task's own record reads: a `done` task is a `Completed`, a
/// `cancelled` one is a `Cancelled`, and a `failed` task — like one that never
/// settled at all — is a `Fault`, because the work was still owed when the agent
/// left.
fn grace_close_reason(inner: &DispatchInner, task_id: &str) -> crate::backend::CloseReason {
    match stored_task_state(inner, task_id) {
        TaskState::Done => crate::backend::CloseReason::Completed,
        TaskState::Pending | TaskState::Failed | TaskState::Blocked => {
            crate::backend::CloseReason::Fault
        }
        TaskState::Cancelled => crate::backend::CloseReason::Cancelled,
    }
}

/// Whether one slot is due for the reconnect grace to take it.
///
/// The window as it always was: the connection that would have sent this
/// session's next heartbeat has ended, and the window runs from the moment it
/// left. A slot a live connection serves is not this arm's to end, because an
/// attached transport is the one thing that says the agent is still reachable.
fn dropped_past_window(
    inner: &DispatchInner,
    key: &str,
    slot: &SessionSlot,
    now: Instant,
    window: Duration,
) -> bool {
    slot.dropped_at.is_some_and(|dropped| {
        !has_attached_transport(inner, key, slot)
            && now
                .checked_duration_since(dropped)
                .is_some_and(|away| away >= window)
    })
}

/// Whether one session's own agent has gone quiet on a socket that is still up.
///
/// The window's other door, and it exists precisely because an attached
/// transport — the one thing the arm above trusts — can lie. A socket that is
/// still up proves the connection survived; it says nothing about the agent
/// behind it, and a plugin whose event loop is blocked holds its socket and
/// stops beating. Nothing the client reads before this could see that: no socket
/// ends, so no drop clock ever starts, and the session keeps its slot, its
/// projected row and its host resource for as long as the client runs.
///
/// So the reading is taken off the frames themselves. A session whose task is
/// still bound and unsettled and whose last accepted frame is older than the
/// protocol's heartbeat interval by [`HEARTBEAT_SILENCE_MARGIN`] has no agent
/// behind its socket.
///
/// Why no task bound is excluded: the plugin stops its heartbeat loop with the
/// last task it was given, so a task-free session that has gone quiet is an
/// agent waiting for work by design — the ordinary shape between deliveries.
/// Sweeping it would retire the very connection the next payload is staged onto.
/// The unsettled check beside the binding says the same thing about work that
/// already landed: a settled task has nothing left for its session to answer,
/// and `release_locked` has already given the binding back.
///
/// Why the drop clock's arm never reads this stamp: the two are mutually
/// exclusive by construction. A re-mount clears `dropped_at`, so a session that
/// came back is judged by its beat alone, and a session whose connection went
/// away is judged by the clock alone and never by a stamp its agent can no
/// longer refresh.
fn silent_past_window(inner: &DispatchInner, key: &str, slot: &SessionSlot, now: Instant) -> bool {
    if slot.dropped_at.is_some() || !has_attached_transport(inner, key, slot) {
        return false;
    }
    let Some(task_id) = slot.task_id.as_deref() else {
        return false;
    };
    if stored_task_state(inner, task_id) != TaskState::Pending {
        return false;
    }
    let quiet = HEARTBEAT_INTERVAL * HEARTBEAT_SILENCE_MARGIN;
    slot.last_beat.is_some_and(|beat| {
        now.checked_duration_since(beat)
            .is_some_and(|away| away >= quiet)
    })
}

/// Stop holding a session's connection as its transport.
///
/// The act a socket ending performs, run here on the client's own verdict
/// instead: a session judged dead by its silence still holds its socket, and the
/// connection is not going to end on its own while the agent behind it is
/// blocked. The binding goes now, so the retirement below runs as the ordinary
/// one, and a frame from that connection afterwards is refused the way every
/// frame from a connection no session answers for is — it is served no state,
/// which is the same door a stale reporter already comes to.
fn unbind_transports(inner: &mut DispatchInner, key: &str, slot: &SessionSlot) {
    inner
        .transports
        .retain(|served, _| !names_session(key, slot, served));
}

/// One backend close a retirement owed, carried off the dispatch lock.
///
/// A close is a host round trip — a pane kill, a terminal close, an agent reap —
/// and it can take seconds. The dispatch lock is the one lock every adapter frame,
/// every report, and every slot of this role queues behind, so a sweep that
/// retires more than one session at a time must not spend that lock on the hosts
/// it is calling. A retirement therefore records what it took down here and runs
/// the closes once it is off the lock.
pub(super) struct PendingClose {
    pub(super) backend: Arc<dyn SessionBackend>,
    pub(super) session: SessionRef,
    pub(super) reason: crate::backend::CloseReason,
}

/// Run the closes a retirement collected.
///
/// Four of the five callers hold the dispatch lock and can put it down first — the
/// two sweeps, the goodbye path, and the merged-handoff retire — and each does,
/// because a close that blocks must not hold every adapter frame of this role
/// behind it. The fifth is `release_locked`, which runs inside a caller that
/// already holds the lock for the store work above it and cannot leave; it takes
/// the close where it stands, which is the behavior that path has always had.
///
/// A failed close is reported and not propagated. The slot is gone from this
/// client's books either way and its row already reads closed, so the caller has
/// nothing left to undo; an error that travelled back would replace the answer
/// the caller is waiting for — which sessions left, and therefore must be
/// published — with a failure to say so.
pub(super) fn close_retired(pending: Vec<PendingClose>) {
    for close in pending {
        if let Err(error) = close.backend.close(&close.session, close.reason, false) {
            tracing::warn!(
                task = %close.session.task_id,
                backend = %close.session.backend,
                resource = %close.session.backend_ref,
                error = %error,
                "session resource retirement failed"
            );
        }
    }
}

/// Whether this role's scope keeps one session alive after its delivery settles.
///
/// `oneshot` is the rule the client has always had: the session served its one
/// delivery and the slot is done with it. A `task` or `role` session outlives
/// the delivery that opened it, which is the whole point of the scope, so its
/// slot stays serving nothing until the scope sends it the next delivery or
/// `idle_close` releases its process.
///
/// A read-only slot is never kept: it is a connection that came back for a task
/// a newer session already took, and it owns no delivery of this role.
pub(super) fn keeps_idle(inner: &DispatchInner, key: &str) -> bool {
    inner
        .sessions
        .get(key)
        .is_some_and(|slot| slot.keeps_idle && !slot.read_only)
}

/// Retire one task-free session after its transport set becomes empty.
///
/// This is the `oneshot` ending, and the ending of every session whose scope
/// does not keep it: its resource closes because the agent able to run another
/// task in it has left. [`keeps_idle`] answers for the scopes that keep a
/// session between deliveries, and those never reach here with work behind them.
///
/// The idle slot releases its backend resource because the agent able to run
/// another task in it has left. An attached transport keeps the resource because
/// that agent remains reachable. The dispatch lock serializes the final transport
/// check, reference refresh, lifecycle projection, and slot removal with adapter
/// binding. The backend close is deliberately NOT one of them: it is recorded in
/// `pending` for the caller to run once it is off the lock, so a sweep of sessions
/// cannot hold every frame of this role behind a host round trip. See
/// [`close_retired`] for the one path that cannot put the lock down.
pub(super) fn retire_idle_locked(
    inner: &mut DispatchInner,
    key: &str,
    reason: crate::backend::CloseReason,
    pending: &mut Vec<PendingClose>,
) -> bool {
    let Some(slot) = inner.sessions.get(key) else {
        return false;
    };
    if slot.task_id.is_some() || has_attached_transport(inner, key, slot) {
        return false;
    }

    let original = slot.session.clone();
    let task_id = original.task_id.clone();
    let resource = inner
        .store
        .get_session(&task_id)
        .ok()
        .flatten()
        .map(|row| row.resource_state)
        .unwrap_or_else(|| "detached".to_string());
    if resource != "detached" && resource != "closed" {
        let session = match inner.backend.attach(&original) {
            Ok(refreshed) => {
                if refreshed != original {
                    inner.bridge.track_live(refreshed.clone());
                    if let Some(slot) = inner.sessions.get_mut(key) {
                        slot.session = refreshed.clone();
                    }
                }
                refreshed
            }
            Err(_) => original,
        };
        tracing::info!(
            task = %task_id,
            backend = %session.backend,
            resource = %session.backend_ref,
            ?reason,
            "retiring idle session resource"
        );
        if let Err(error) = feed_resource_closed(&inner.bridge, &inner.store, &task_id) {
            tracing::warn!(
                task = %task_id,
                backend = %session.backend,
                resource = %session.backend_ref,
                error = %error,
                "session resource close projection failed"
            );
        }
        pending.push(PendingClose {
            backend: Arc::clone(&inner.backend),
            session,
            reason,
        });
    }
    if reason == crate::backend::CloseReason::Completed {
        if let Err(error) = feed_agent_gone(&inner.bridge, &inner.store, &task_id) {
            tracing::warn!(
                task = %task_id,
                error = %error,
                "agent-gone projection failed for a completed session"
            );
        }
    }
    inner.bridge.untrack_live(&task_id);
    forget_tools_binding(inner, key);
    inner.sessions.remove(key);
    true
}

/// Give one session's task slot back. Settled tasks enter idle retirement, and
/// explicit reasons drive the control-close path.
pub(super) fn release_locked(
    inner: &mut DispatchInner,
    task_id: &str,
    reason: Option<crate::backend::CloseReason>,
) -> Result<()> {
    let resource = inner
        .store
        .get_session(task_id)?
        .map(|row| row.resource_state)
        .unwrap_or_else(|| "detached".to_string());
    if let Some((key, slot)) = slot_key_serving_task(inner, task_id)
        .and_then(|key| inner.sessions.get(&key).map(|slot| (key, slot.clone())))
    {
        if let Some(reason) = reason {
            if resource != "detached" && resource != "closed" {
                feed_resource_closed(&inner.bridge, &inner.store, task_id)?;
                inner.backend.close(&slot.session, reason, false)?;
            }
            // The agent goes with the resource: this path closes a session whose work an
            // operator ended or whose backend faulted, and the slot below leaves the map in
            // the same breath. Without this feed the tuple keeps the agent phase its last
            // beat reported, and `project` answers `working` for a `cancelled` or `failed`
            // task whenever the agent is not `Gone` — so the mirrored row read `working` for
            // a session the client had already closed, which is what `onlyne sessions` and
            // the board showed beside a ledger row that had settled. The reconnect sweep
            // feeds the same event for the same ending.
            if let Err(error) = feed_agent_gone(&inner.bridge, &inner.store, task_id) {
                tracing::warn!(
                    task = %task_id,
                    error = %error,
                    "agent-gone projection failed for a closed session"
                );
            }
            inner.bridge.untrack_live(task_id);
            // The delivery row this session is still holding is answered here,
            // while the slot that holds its handle is still in the map. The close
            // takes the slot, and the handle goes with it: a row this client
            // never answered is a row the server still reads as owed, so the
            // pull that would have taken it passes and the release of a session
            // the server judges gone hands it back to the queue — which
            // dispatches the task a second time and runs it again, as the live
            // run showed for a task whose work the close had already ended.
            //
            // The handle is taken, so the row is answered once: the control
            // settle watchdog and the reconnect sweep refuse the same handle
            // the same way, and whichever of the three runs first spends it and
            // the others find nothing left to answer.
            let held = inner
                .sessions
                .get_mut(&key)
                .and_then(|slot| slot.msg_id.take());
            if let Some(msg_id) = held {
                store_ack(
                    inner,
                    AckArgs {
                        msg_id,
                        op_id: None,
                        accepted: false,
                        reason: Some(close_refusal(reason).to_string()),
                    },
                );
            }
            forget_tools_binding(inner, &key);
            inner.sessions.remove(&key);
        } else {
            let session_id = inner
                .sessions
                .get(&key)
                .map(|slot| slot.session.task_id.clone())
                .unwrap_or_else(|| task_id.to_string());
            let keeps = keeps_idle(inner, &key);
            let now = Instant::now();
            if let Some(session) = inner.sessions.get_mut(&key) {
                session.task_id = None;
                session.ready = false;
                session.idle_since = Some(now);
            }
            // The binding goes with the delivery. A session that serves several
            // deliveries in turn serves none between them, and both the hello
            // claim and the mirrored row read the open binding to say which
            // delivery a session is on.
            if let Err(error) = inner.store.release_binding(&session_id, task_id) {
                tracing::warn!(
                        session = %session_id,
                        task = %task_id,
                        error = %error,
                        "a settled session's binding was not handed back"
                );
            }
            if keeps {
                // A `task` or `role` session outlives the delivery that opened
                // it: that is what the scope is for. It stays idle, with its
                // process and its row, until the scope sends it the next
                // delivery or `idle_close` releases the process.
                tracing::info!(
                        session = %session_id,
                        task = %task_id,
                        "delivery settled; the session stays idle for its scope's next one"
                );
            } else {
                let mut pending = Vec::new();
                retire_idle_locked(
                    inner,
                    &key,
                    crate::backend::CloseReason::Completed,
                    &mut pending,
                );
                // This function answers to its caller's lock, which is already held
                // across the store work above, so the close it owes cannot wait for an
                // unlock this scope cannot perform. Running it here is the same
                // under-lock call the path has always made; the sweeps that CAN leave
                // the lock — `retire_dropped_ghosts`, `reclaim_exited_resources`,
                // `close_all` — are the ones that now defer it.
                close_retired(pending);
            }
        }
    }
    inner.stall.forget(task_id);
    if reason.is_some() && resource == "detached" {
        inner
            .store
            .note_alert(format!("session recycled {task_id}"));
    }
    Ok(())
}

/// The operator's word a control close stands for, as the refusal that answers
/// the row its session was still holding.
///
/// The words are the ones the settle fallback already writes for the same
/// commands ([`ControlWord::refusal`]), so a row reads the same thing whichever
/// door refused it, and each names the command that was given rather than the
/// verdict that command left behind.
fn close_refusal(reason: crate::backend::CloseReason) -> &'static str {
    match reason {
        // The close an operator's `cancel` runs: the `ControlOp::Cancel` arm of
        // `on_control`, which is also how the server asks for `repair close` and
        // `repair fail` to reach this client.
        crate::backend::CloseReason::Cancelled => ControlWord::Cancel.refusal(),
        // The close an operator's `recycle` runs: the `ControlOp::Recycle` arm
        // of `on_control`.
        crate::backend::CloseReason::Operator => ControlWord::Recycle.refusal(),
        // No control command reaches this branch with another reason: a
        // `completed`, `fault` or `replaced` close retires an idle slot, and a
        // `shutdown` close runs `close_all`, neither of which comes through
        // here. The word is the server's own for a close that named no command
        // — `repair close` without a `--reason` writes it on the task's row.
        _ => "operator close",
    }
}

/// Close every live session's resource with `reason` and forget the slots.
///
/// This is the shutdown path: a stopped client must not leave resources behind
/// that only it can address, and each backend's own record of the resource —
/// the Orca tab map included — ends with the session. `budget` bounds the whole
/// sweep, because an operator's SIGTERM must not turn into a hang while a slow
/// backend CLI exits; whatever the budget cuts off is reported and dropped
/// anyway.
///
/// The order is deliberate: every slot is taken off this client's books under the
/// dispatch lock — untracked and removed, so nothing still reads a live session —
/// and the closes it collected run once the lock is let go. A shutdown that
/// closed each pane while holding that lock spent the whole budget on the host,
/// with the lock no adapter frame could reach.
pub fn close_all(state: &DispatchState, reason: crate::backend::CloseReason, budget: Duration) {
    let started = Instant::now();
    let pending: Vec<PendingClose> = {
        let mut inner = state.inner.lock();
        let sessions: Vec<(String, SessionRef)> = inner
            .sessions
            .iter()
            .map(|(key, slot)| (key.clone(), slot.session.clone()))
            .collect();
        let mut pending = Vec::with_capacity(sessions.len());
        for (key, session) in sessions {
            inner.bridge.untrack_live(&session.task_id);
            forget_tools_binding(&mut inner, &key);
            inner.sessions.remove(&key);
            pending.push(PendingClose {
                backend: Arc::clone(&inner.backend),
                session,
                reason,
            });
        }
        pending
    };
    for close in pending {
        if started.elapsed() > budget {
            tracing::warn!(
                task = %close.session.task_id,
                "shutdown close budget reached; the resource is left behind"
            );
            continue;
        }
        if let Err(error) = close.backend.close(&close.session, close.reason, false) {
            tracing::warn!(
                task = %close.session.task_id,
                error = %error,
                "session close failed during shutdown"
            );
        }
    }
}

/// Which way a session's window closed.
///
/// Two arms reach the same verdict through different facts, and an operator reading one
/// aggregate log line cannot tell them apart: one means the connection ended and the agent
/// stayed away, the other means the connection is still up while nothing the client accepts
/// arrives over it. The words exist so the log can name which reading retired a session.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum RetirementArm {
    /// The plugin connection ended and `[client] reconnect_grace_secs` expired.
    Dropped,
    /// The connection stayed up while the session went quiet past the heartbeat window.
    Silent,
}

impl RetirementArm {
    pub fn word(self) -> &'static str {
        match self {
            Self::Dropped => "reconnect_grace",
            Self::Silent => "heartbeat_silence",
        }
    }
}

/// One session the sweep retired, with what it read on the way.
pub struct Retired {
    /// The retired session's own id, which a client-held session shares with its task.
    pub session_id: String,
    /// The arm that decided it.
    pub arm: RetirementArm,
    /// Seconds since this session's last accepted frame.
    pub quiet_secs: u64,
    /// Seconds since its connection ended, when one did.
    pub away_secs: Option<u64>,
}

/// How long a session has gone without a frame this client accepted.
fn quiet_secs(slot: &SessionSlot, now: Instant) -> u64 {
    slot.last_beat
        .map(|beat| now.saturating_duration_since(beat).as_secs())
        .unwrap_or(u64::MAX)
}

/// How long a session's connection has been gone, for a session still waiting on one.
fn away_secs(slot: &SessionSlot, now: Instant) -> Option<u64> {
    slot.dropped_at
        .map(|left| now.saturating_duration_since(left).as_secs())
}

impl DispatchState {
    /// Retire tracked resources whose stored lifecycle has reached `Exited`.
    ///
    /// The periodic readiness tick calls this after completed work becomes an
    /// idle slot. Task-free sessions with an attached transport stay bound to
    /// their host resource, and task-free sessions whose agent has left release it.
    ///
    /// The session ids come back because the retirement wrote each one of those
    /// rows and the server mirrors only what this client reports: the resource
    /// close and, for a completed session, the agent's exit both moved the row
    /// this tick found, and a publish cannot run under this lock. The caller is
    /// handed what to publish, the same answer [`DispatchState::retire_dropped_ghosts`]
    /// gives the sweep above.
    pub fn reclaim_exited_resources(&self) -> Vec<String> {
        let mut inner = self.inner.lock();
        let candidates: Vec<(String, crate::backend::CloseReason)> = inner
            .sessions
            .iter()
            .filter(|(key, slot)| {
                // A suspended session is not an exited one. Its process is gone
                // *because* it was released, its conversation is what the
                // family's next delivery is for, and the row it published says
                // `idle` for exactly that reason — `binding_task_state` answers
                // `Pending` for a slot that serves nothing and whose scope keeps
                // it, so the two derivations of "is this over" disagree here.
                //
                // This sweep asks the other one, and it asks about the *work*:
                // a suspended session's delivery is finished, so it reads
                // `Exited` and the slot is retired within a tick. The family's
                // next delivery then finds nothing to resume and opens a second
                // conversation for one chain — which is the failure this filter
                // existed to prevent, reached through the sweep that was meant
                // to clean up after a session that was already gone.
                !slot.suspended
                    && slot.task_id.is_none()
                    && session_exited(&inner, &slot.session.task_id)
                    && !has_attached_transport(&inner, key, slot)
            })
            .filter_map(|(key, slot)| {
                stored_close_reason(&inner, &slot.session.task_id)
                    .map(|reason| (key.clone(), reason))
            })
            .collect();
        let mut retired: Vec<String> = Vec::new();
        let mut pending: Vec<PendingClose> = Vec::new();
        for (key, reason) in candidates {
            // The id that travels is the session's own, the one whose row the
            // retirement below is about to write.
            let Some(task_id) = inner
                .sessions
                .get(&key)
                .map(|slot| slot.session.task_id.clone())
            else {
                continue;
            };
            if retire_idle_locked(&mut inner, &key, reason, &mut pending) {
                retired.push(task_id);
            }
        }
        // The tick that found a batch of finished sessions owes a host close for
        // each, and this sweep runs every 250 ms: on the lock, one slow backend
        // would spend the whole window and every adapter frame queued behind it.
        drop(inner);
        close_retired(pending);
        retired
    }

    /// Retire the sessions whose plugin connection dropped and never came back, or
    /// whose connection stayed up while they went quiet, and answer which ones left
    /// and why.
    ///
    /// A connection that ends without a `detach` frame leaves its session tracked
    /// so an agent that restarts inside `[client] reconnect_grace_secs` finds the
    /// resource it was using. That promise has to expire: a process that is
    /// simply gone would otherwise hold a slot, a projected `idle` row, and a live
    /// host resource forever, and on a role with `max_sessions = 1` it stops every
    /// later delivery. The window answers for the agent itself, so a session still
    /// bound to a task goes with it: the plugin connection that would have
    /// reported the ending is the one that dropped. The agent-gone feed is what
    /// says the process left — the session's own tuple reaches `Exited` through
    /// `AgentPhase::Gone` rather than through a task result — and the reason the
    /// backend is handed is the one `grace_close_reason` reads off what the slot
    /// still owes.
    ///
    /// What the slot owed is settled too: the task a bound session was serving
    /// ends `failed` here, because the agent that would have reported its ending
    /// is the one that left. A task with no verdict stays open for the server to
    /// re-offer and for `open_tasks` to keep reading, and no later caller exists
    /// to write one.
    ///
    /// A slot this client holds read-only is not this sweep's to end, agent gone
    /// or not: the session id it would feed is the task id, so the ghost's death
    /// would take the live session's mirror and its delivery row down with it.
    /// That retirement belongs to `retire_revived`, which runs when the
    /// completion that answers the held connection merges.
    ///
    /// The window has a second way to open, and it is the one a socket cannot
    /// report: a plugin whose event loop is blocked keeps its connection and
    /// stops beating, so no socket ends and no clock this sweep could read
    /// before moved. What such a session leaves behind is a stamp going stale
    /// while its task stays bound and unsettled, and that is the reading this
    /// sweep takes now. It is the same window and the same verdict — one clock,
    /// one retirement, no second threshold beside `[client]
    /// reconnect_grace_secs` and no fault row of the kind `stall_report_secs`
    /// records and leaves behind.
    ///
    /// The sessions' own ids come back rather than a count, because a retirement
    /// still owes the server the session's own ending: it is the only writer left
    /// for that task, and a mirror nobody tells keeps that session's last reading —
    /// `working`, for one that had beaten — until the server's own observer records
    /// a fault about it. The publish is
    /// [`sync_session`](crate::session::dispatch::sync_session)'s, which is the
    /// report an ordinary ending travels on, and it cannot run under this lock —
    /// so the caller is handed what to publish instead of a second writer being
    /// invented here.
    pub fn retire_dropped_ghosts(&self, now: Instant, grace_secs: u64) -> Vec<Retired> {
        if grace_secs == 0 {
            return Vec::new();
        }
        let window = Duration::from_secs(grace_secs);
        let mut inner = self.inner.lock();
        // The arm is decided from the pair of readings here, while both still
        // describe the slot: which fact closed the window is what the operator
        // has to be able to tell apart once the retirement itself is a line in
        // the log.
        let due: Vec<(String, RetirementArm)> = inner
            .sessions
            .iter()
            .filter_map(|(key, slot)| {
                let arm = if dropped_past_window(&inner, key, slot, now, window) {
                    RetirementArm::Dropped
                } else if silent_past_window(&inner, key, slot, now) {
                    RetirementArm::Silent
                } else {
                    return None;
                };
                Some((key.clone(), arm))
            })
            .collect();
        let mut retired: Vec<Retired> = Vec::new();
        let mut pending: Vec<PendingClose> = Vec::new();
        for (key, arm) in due {
            let Some(slot) = inner.sessions.get(&key).cloned() else {
                continue;
            };
            // A slot this client holds read-only is not this sweep's to end, for
            // the reason the doc above gives. The demotion alone does not decide
            // it: the held connection that owns the slot can go without the task
            // ever completing, and a slot nothing owns any more is what this
            // window is for.
            if held_read_only(&inner, &key, &slot) {
                continue;
            }
            // A session judged dead on its silence is the one case where the
            // connection is still there: the socket has not ended and will not
            // while the agent behind it is blocked, so the client's own verdict
            // has to take the binding the way the death of the socket would have.
            // The retirement below refuses a slot an attached transport still
            // serves, and that refusal is what this unbinding answers: the
            // verdict has already been reached here, so the binding goes rather
            // than the death of the socket that would normally take it.
            if slot.dropped_at.is_none() {
                unbind_transports(&mut inner, &key, &slot);
            }
            let task_id = slot.session.task_id.clone();
            // Both the feed and the reason name the session's own task, so the id
            // that travels is the one the row and the resource are keyed by.
            let reason = grace_close_reason(&inner, &task_id);
            // The work this session still owed ends here, and this sweep is the
            // only writer left to say so: the plugin connection that would have
            // reported the ending is the one that dropped. A task nobody answers
            // stays `settled_at IS NULL` forever, so `open_tasks` keeps reading
            // it and the server keeps re-offering a delivery no client can take.
            // The binding is what the slot owed — a slot past its window with no
            // task bound owes nothing — and `failed` is the verdict the close
            // reason above already carries for it. A row an earlier verdict
            // settled keeps that one: `settle_task` updates only where
            // `settled_at IS NULL` and answers `false`.
            //
            // The write runs before the agent-gone feed and before the binding
            // hand-back, both of which end this slot's turn through the sweep:
            // the id is captured here, and the verdict is on disk before the
            // only handle on it goes away.
            if let Some(owed) = slot.task_id.clone() {
                // A verdict written here answers the task for good, so a
                // `control` command that is still waiting for its plugin's report
                // has nothing left to authorise: the note goes with the verdict
                // that outranks it.
                inner.control_settles.retain(|noted| noted.task_id != owed);
                if let Err(error) = inner.store.settle_task(&owed, TaskState::Failed) {
                    tracing::warn!(
                        task = %owed,
                        error = %error,
                        "the task of a retired ghost was not settled"
                    );
                }
                // The delivery handle this session was holding is spent as a
                // refusal that names the death, and it is the ledger half of the
                // verdict above. A row left `in_flight` is handed to a pull no
                // longer — `pull` passes by a row whose ticket is armed, and a
                // role-level pull's ticket carries no session id for the release
                // path to match — so nothing would answer for this task until the
                // link dropped, and an operator reading `onlyne ledger` would see
                // a session that has been buried as one still holding its
                // delivery. The reason is this client's own word for a death
                // (`SESSION_DEAD`), and a refusal is terminal: the work comes back
                // through `repair retry`, not by itself.
                let handle = inner
                    .sessions
                    .get_mut(&key)
                    .and_then(|slot| slot.msg_id.take());
                if let Some(msg_id) = handle {
                    store_ack(
                        &inner,
                        AckArgs {
                            msg_id,
                            op_id: None,
                            accepted: false,
                            reason: Some(SESSION_DEAD.to_string()),
                        },
                    );
                }
            }
            if let Err(error) = feed_agent_gone(&inner.bridge, &inner.store, &task_id) {
                tracing::warn!(
                    task = %task_id,
                    error = %error,
                    "agent-gone projection failed for a retired ghost"
                );
            }
            // The agent left, so the session owes no task any more: the binding
            // goes back before the idle retirement takes the slot.
            if let Some(slot) = inner.sessions.get_mut(&key) {
                slot.task_id = None;
            }
            if retire_idle_locked(&mut inner, &key, reason, &mut pending) {
                // The id that travels is the session's own, the one whose row was
                // just fed agent-gone and resource-closed: that row is what the
                // server mirrors, and its ending is what the caller publishes. The
                // ages are read off the slot as it stood before the retirement.
                retired.push(Retired {
                    session_id: task_id,
                    arm,
                    quiet_secs: quiet_secs(&slot, now),
                    away_secs: away_secs(&slot, now),
                });
            }
        }
        // Every session this sweep took is off the slot map and fully written down
        // by now; what remains is the host's own work, and the lock that answers an
        // agent's next frame is not to be held through it.
        drop(inner);
        close_retired(pending);
        retired
    }
}