liminal-server 0.25.4

use beamr::native::native_process::{NativeContext, NativeHandler, NativeOutcome};
use beamr::process::ExitReason;
use beamr::scheduler::Interest;
use beamr::term::Term;

use crate::ServerError;
use crate::server::mount::MountKind;
use crate::server::participant::ConnectionFateClass;

use liminal_protocol::wire::ConnectionIncarnation;

use std::io::Write;
use std::net::SocketAddr;
use std::sync::Arc;

#[path = "process/buffer.rs"]
mod buffer;
#[path = "process/close.rs"]
mod close;
#[path = "process/control.rs"]
mod control;
#[path = "process/force_close.rs"]
mod force_close;
#[path = "process/participant_pushes.rs"]
mod participant_pushes;
#[path = "process/slice_readiness.rs"]
mod slice_readiness;
#[path = "process/tcp_transport.rs"]
mod tcp_transport;

use buffer::process_buffer;
use participant_pushes::participant_publication_error;
use tcp_transport::TcpTransport;

use super::delivery::service_subscriptions;
use super::outbound::{DrainOutcome, OutboundWriter};
use super::participant_delivery::has_held_participant_head;
use super::state::{ConnectionProcessState, ProcessStatus};
use super::supervisor::runtime::{ConnectionControl, ConnectionRuntime};

#[cfg(test)]
#[path = "process_teardown_tests.rs"]
mod teardown_tests;

#[cfg(test)]
#[path = "process_fate_tests.rs"]
mod fate_tests;

#[cfg(test)]
#[path = "process_terminal_tests.rs"]
mod terminal_tests;

#[cfg(test)]
#[path = "process_pre_admission_tests.rs"]
mod pre_admission_tests;

#[cfg(test)]
#[path = "process_wake_tests.rs"]
mod wake_tests;

#[cfg(test)]
#[path = "process_readiness_stop_tests.rs"]
pub(super) mod readiness_stop_tests;

/// Per-slice inbound read size, and the per-slice outbound drain budget.
///
/// Shared with the sibling transports so one connection cannot monopolize a
/// scheduler thread on one door and not another.
pub(super) const READ_BUFFER_BYTES: usize = 8192;
/// Application stream id used for server-initiated push frames. Push is an
/// application-stream frame (non-zero stream id), like publish and conversation.
const PUSH_STREAM_ID: u32 = 1;

/// One supervised connection, generic over the transport half it owns.
///
/// EVERYTHING above the read/write halves is here exactly once: the slice
/// ordering, the `apply_frame` seam, the pending-reply table, the delivery and
/// participant-publication pumps, the control vocabulary, the close/fate paths,
/// and the teardown backstop. A sibling transport supplies a
/// [`ConnectionTransport`] and inherits all of it — no second slice loop, no
/// second probe, no second close path to keep in step. That is the design's
/// "third sibling, not a new seam" answer, and it is why the loopback needs no
/// parallel process type.
#[derive(Debug)]
pub(super) struct TransportConnectionProcess<T: ConnectionTransport> {
    runtime: Arc<ConnectionRuntime>,
    peer_addr: Option<SocketAddr>,
    /// The read/write halves and their transport-specific readiness discipline.
    transport: T,
    buffer: Vec<u8>,
    state: ConnectionProcessState,
    /// Per-connection outbound byte buffer. EVERY server-originated frame (acks,
    /// errors, `Push`, `Disconnect`, `Pong`, `Deliver`, ...) is enqueued here and
    /// drained cooperatively on the slice loop with partial-write tracking, so a
    /// frame larger than the socket send buffer streams out across slices instead
    /// of failing a `write_all` on the non-blocking socket (ledger G4).
    outbound: OutboundWriter,
    /// Set once the transport's wake has been installed, so the install runs on
    /// the first serviced slice and never again.
    wake_installed: bool,
    agent_connection: Option<u64>,
    /// Fails the next participant publication pump with a non-capacity fault.
    /// Test-only: the publication teardown, forced.
    #[cfg(test)]
    fail_next_pump: bool,
}

/// The supervised TCP connection process: the original, and the transport every
/// sibling is measured against.
pub(super) type ConnectionProcess = TransportConnectionProcess<TcpTransport>;

/// Whether a connection slice's socket/inbound servicing leaves the connection
/// running or has already resolved its lifecycle.
enum SliceStep {
    Continue,
    Stop(ExitReason),
}

/// A connection's read and write halves, plus the transport-specific parts of
/// parking: how readiness is armed, how the reader is told, and what a drop of
/// the transport owes the runtime.
///
/// Implementors supply ONLY what genuinely differs between doors. Every
/// transport-neutral decision — when to park, when to requeue, what a close
/// costs, which fate a hangup folds — stays in
/// [`TransportConnectionProcess`], so a new mount cannot accidentally acquire
/// different lifecycle semantics by copying a slice loop slightly wrong.
pub(super) trait ConnectionTransport: Send + 'static {
    /// The door this transport IS (design §10).
    ///
    /// An associated const rather than a constructor argument: the mount fact
    /// is then a property of the transport type itself, so no spawn path can
    /// stamp a mount that disagrees with the halves it actually handed over,
    /// and no runtime value — least of all an inbound byte — can move it.
    const MOUNT: MountKind;

    /// Whether the transport still holds its halves. A transport that has been
    /// released (or never received them) can neither read nor write.
    fn is_connected(&self) -> bool;

    /// Reads whatever is available into `buffer` without blocking.
    ///
    /// # Errors
    /// Returns [`ServerError::ListenerAccept`] on an unrecoverable read
    /// failure; `WouldBlock`/`Interrupted` are reported as
    /// [`ReadStatus::WouldBlock`], not errors.
    fn read_available(&mut self, buffer: &mut Vec<u8>) -> Result<ReadStatus, ServerError>;

    /// The outbound sink, or `None` when the transport has no write half left.
    fn sink(&mut self) -> Option<&mut dyn Write>;

    /// Whether inbound bytes are already waiting, without consuming them.
    ///
    /// A transport with no read half answers `true`: quiescence was never
    /// established, so parking on it would be a claim the probe cannot make.
    ///
    /// # Errors
    /// Returns [`ServerError::ListenerAccept`] when the probe itself fails.
    fn probe_inbound(&self) -> Result<bool, ServerError>;

    /// Arms this transport's readiness interest before the connection parks.
    ///
    /// # Errors
    /// Returns [`ServerError::ListenerAccept`] when the readiness contract
    /// cannot be satisfied.
    fn arm_readiness(
        &mut self,
        pid: u64,
        ctx: &NativeContext<'_>,
        interest: Interest,
        runtime: &ConnectionRuntime,
    ) -> Result<(), ServerError>;

    /// Installs this transport's wake on the connection's first serviced slice,
    /// once the host record exists (so the connection's `READY` waker can be
    /// built). Transports woken by the readiness facility need nothing here.
    ///
    /// # Errors
    /// Returns [`ServerError::ListenerAccept`] when the wake cannot be
    /// installed — for a transport that has no other way to be told, a failed
    /// install is fatal rather than a silently deaf connection.
    fn install_wake(&mut self, pid: u64, runtime: &ConnectionRuntime) -> Result<(), ServerError>;

    /// Drops the transport's halves at an orderly server-forced close, after the
    /// host record has been removed.
    fn release(&mut self);

    /// Observes the transport's halves actually dropping, for the descriptor
    /// allocator boundary tests. Production keeps the ordinary field-drop path.
    #[cfg(test)]
    fn note_process_drop(&mut self, runtime: &ConnectionRuntime);
}

impl<T: ConnectionTransport> TransportConnectionProcess<T> {
    /// Builds a connection process over `transport`.
    ///
    /// The ONE construction: pending-reply table sized from the runtime's
    /// configured §5 caps, publication inbox iff a durable incarnation and a
    /// participant service both exist, and the mount stamped from the
    /// transport's own [`ConnectionTransport::MOUNT`]. Every sibling transport
    /// enters here, so none of those decisions can drift per door.
    pub(super) fn over_transport(
        runtime: Arc<ConnectionRuntime>,
        peer_addr: Option<SocketAddr>,
        transport: T,
        connection_incarnation: Option<ConnectionIncarnation>,
    ) -> Self {
        // Build the pending-reply table from the runtime's configured §5 caps
        // (R1(vi), §1.2(3b)). The default table carries the signed defaults; this
        // uses the connection's actual limits.
        let limits = runtime.limits();
        let pending_replies = super::pending_reply::PendingReplyTable::new(
            limits.max_pending_replies_per_conversation,
            limits.max_pending_conversation_replies_per_connection,
            std::time::Duration::from_millis(limits.reply_timeout_ms),
        );
        let participant_publication = connection_incarnation.and_then(|_| {
            runtime
                .participant_service()
                .map(crate::server::participant::InstalledParticipantService::new_publication_inbox)
        });
        let state = ConnectionProcessState {
            connection_incarnation,
            participant_publication,
            pending_replies,
            // Design §10: stamped from the transport TYPE, so the door that
            // handed over the halves is the door on the record.
            mount: T::MOUNT,
            ..ConnectionProcessState::default()
        };
        // The standing bound is the operator's; the test seam overrides it for ONE
        // connection at a time (a fixture needs a small buffer on one peer while
        // its siblings stay normal), which a server-wide setting cannot express.
        #[cfg(test)]
        let outbound = OutboundWriter::with_capacity(
            runtime
                .take_next_outbound_capacity()
                .unwrap_or(limits.max_connection_outbound_bytes),
        );
        #[cfg(not(test))]
        let outbound = OutboundWriter::with_capacity(limits.max_connection_outbound_bytes);
        Self {
            runtime,
            peer_addr,
            transport,
            buffer: Vec::new(),
            state,
            outbound,
            wake_installed: false,
            agent_connection: None,
            #[cfg(test)]
            fail_next_pump: false,
        }
    }

    /// The mount this process's transport stamped onto its connection state.
    ///
    /// The test-side read of the design §10 fact: it proves the door's stamp
    /// arrived on the state `apply_frame` builds the handler context from,
    /// rather than trusting that the constructor passed it along.
    #[cfg(test)]
    pub(super) const fn mount(&self) -> MountKind {
        self.state.mount
    }

    /// Runs one connection scheduler slice: service inbound socket/control work,
    /// then service subscriptions into the outbound buffer, then drain the
    /// outbound buffer to the socket.
    ///
    /// The ordering is load-bearing (subscriptions are pumped AFTER socket/control
    /// work), and the whole slice preserves the no-sleep `Continue` discipline: a
    /// slice never parks a scheduler thread, it re-queues the process to poll again.
    fn handle_slice(&mut self, pid: u64, ctx: &mut NativeContext<'_>) -> NativeOutcome {
        self.agent_connection = Some(pid);
        // R7 (§1.2(6)): count this serviced slice. The park-flip's permanent
        // rule-1 quiescence assertion — a parked connection's counter must not
        // advance without an event — reads this; under the busy loop the counter
        // advances every slice, and the instrument proves it counts.
        #[cfg(test)]
        self.runtime.record_slice(pid);
        // `spawn_native` may schedule the first slice before the spawn thread has
        // inserted the host record. Do not mint an unpublishable token; yield once
        // and register after the record exists.
        if !self.runtime.is_registered(pid) {
            return NativeOutcome::Continue;
        }
        if let Err(error) = self.ensure_transport_wake_installed(pid) {
            return self.fail_slice(pid, &error);
        }
        if let Err(error) = self.ensure_participant_publication_registered(pid) {
            return self.fail_slice(pid, &error);
        }
        // ACCESS-005 R6: a revocation is acted on at the connection's own
        // wake even when its control could not be published.
        if self.state.revocation.is_some()
            && let Some(outcome) = self.revocation_sweep(pid)
        {
            return outcome;
        }
        if let Some(outcome) = self.force_close_if_pending(pid) {
            return outcome;
        }
        match self.service_socket(pid) {
            SliceStep::Stop(reason) => return NativeOutcome::Stop(reason),
            SliceStep::Continue => {}
        }
        if let Err(error) = super::supervisor::relay_pump::pump(&self.state, &mut self.outbound) {
            return self.tear_down_outbound(pid, &error, "relay pump failed; tearing down");
        }
        // R1(vi) (§1.2(3b)): service the pending-reply table each slice — expire
        // due deadlines (writing timeout frames) and drain/correlate any replies
        // the participants produced. A fatal outbound condition here tears the
        // connection down, matching the delivery path.
        if let Err(error) = self.service_pending_replies() {
            return self.tear_down_outbound(
                pid,
                &error,
                "conversation replies could not be written or a far call could not be settled; tearing down",
            );
        }
        if let Some(outcome) = self.service_participant_pushes(pid) {
            return outcome;
        }
        // Pump subscriptions into the outbound buffer. An overflow (or an encode
        // fault) is fatal: a dropped or truncated delivery would desync the stream,
        // so the connection is torn down rather than allowed to continue desynced.
        let slice_budget = self.runtime.limits().delivery_slice_budget;
        match service_subscriptions(&mut self.state, &mut self.outbound, slice_budget) {
            Ok(shed) => self.shed_subscriptions(shed),
            Err(error) => {
                return self.tear_down_outbound(
                    pid,
                    &error,
                    "outbound overflow while delivering; tearing down the connection",
                );
            }
        }
        // Drain queued outbound bytes with partial-write tracking. A hard write
        // error (or overflow surfaced here) tears the connection down.
        let drain = match self.drain_outbound() {
            Ok(drain) => drain,
            Err(error) => {
                if let Err(fate_error) =
                    self.complete_connection_fate(ConnectionFateClass::ConnectionLost)
                {
                    tracing::error!(connection_pid = pid, %fate_error, "connection-loss fate fold failed");
                }
                return self.tear_down_outbound(
                    pid,
                    &error,
                    "outbound drain failed; tearing down the connection",
                );
            }
        };
        #[cfg(test)]
        self.report_frames_written(pid, drain);
        self.finish_slice_readiness(pid, ctx, drain)
    }

    /// Installs the transport's wake exactly once, on the first slice that finds
    /// the host record in place.
    ///
    /// The ordering is what closes the lost-wake window for a transport with no
    /// descriptor. The install happens at the TOP of the slice, so it strictly
    /// precedes this slice's [`Self::final_probe`] — and the probe is the only
    /// way to `Wait`. A write that lands before the probe is seen BY the probe
    /// (which requeues); a write that lands after it fires the now-installed
    /// wake. There is no third case, so a parked connection is never a deaf one.
    /// It cannot run earlier than this: the wake names the connection's pid and
    /// its record, and neither exists until the spawn thread has registered.
    fn ensure_transport_wake_installed(&mut self, pid: u64) -> Result<(), ServerError> {
        if self.wake_installed {
            return Ok(());
        }
        self.transport.install_wake(pid, &self.runtime)?;
        self.wake_installed = true;
        Ok(())
    }

    /// Services the inbound half of a slice: reads available bytes and applies any
    /// complete frames, enqueuing responses into the outbound buffer.
    fn service_socket(&mut self, pid: u64) -> SliceStep {
        if let Some(outcome) = self.force_close_if_pending(pid) {
            return match outcome {
                NativeOutcome::Stop(reason) => SliceStep::Stop(reason),
                _ => SliceStep::Continue,
            };
        }
        if !self.transport.is_connected() {
            self.release_conversations();
            self.runtime
                .mark_crashed(pid, ExitReason::Error, self.peer_addr);
            return SliceStep::Stop(ExitReason::Error);
        }
        let read = self.transport.read_available(&mut self.buffer);
        if let Some(outcome) = self.force_close_if_pending(pid) {
            return match outcome {
                NativeOutcome::Stop(reason) => SliceStep::Stop(reason),
                _ => SliceStep::Continue,
            };
        }
        match read {
            Ok(ReadStatus::Closed) => {
                crate::agents::runtime::connection_closed(
                    self.runtime.services(),
                    pid,
                    "transport_eof",
                );
                // Best-effort flush of anything still queued (e.g. WouldBlock residue
                // from earlier slices to a half-closed peer) before we let the stream
                // drop, mirroring the ForceClose drain. The buffered-writer refactor
                // removed this on the EOF path; without it, queued responses are lost.
                //
                // The drain error is only debug-logged, and what is lost with it is the
                // residue this call could not write — the peer never sees those
                // frames. That is accepted because the peer has already closed
                // its write half, so the residue had nowhere to go regardless,
                // and this path is past the close decision: retrying or sleeping
                // to shrink the window is forbidden here.
                if let Err(error) = self.drain_outbound() {
                    tracing::debug!(connection_pid = pid, %error, "the close path's best-effort drain did not finish; the peer's half-close ends it");
                }
                if let Err(error) =
                    self.complete_connection_fate(ConnectionFateClass::ConnectionLost)
                {
                    return self.fail_fate(pid, &error);
                }
                self.release_conversations();
                self.runtime.finish(pid);
                return SliceStep::Stop(ExitReason::Normal);
            }
            Ok(ReadStatus::WouldBlock) => {
                // No bytes ready on this non-blocking socket right now. Do NOT
                // sleep: that would block a beamr scheduler worker thread (the
                // supervisor runs `CONNECTION_SCHEDULER_THREADS`) on every idle
                // poll and starve every other connection process sharing it. The
                // caller returns `NativeOutcome::Continue` (mapping to
                // `SliceOutcome::Requeue`), which re-queues this pid behind every
                // other runnable process (cooperative round-robin) and reschedules
                // us to poll — and, crucially, to drain any queued outbound bytes —
                // without parking. `Wait` is wrong here: it parks until a *message*
                // arrives, but socket readiness does not enqueue a message, so the
                // connection would hang forever.
                return SliceStep::Continue;
            }
            Ok(ReadStatus::Read) => {}
            Err(error) => {
                tracing::warn!(connection_pid = pid, %error, "connection read failed");
                if let Err(fate_error) =
                    self.complete_connection_fate(ConnectionFateClass::ConnectionLost)
                {
                    return self.fail_fate(pid, &fate_error);
                }
                self.release_conversations();
                self.runtime
                    .mark_crashed(pid, ExitReason::Error, self.peer_addr);
                return SliceStep::Stop(ExitReason::Error);
            }
        }
        match process_buffer(
            pid,
            &self.runtime,
            &mut self.state,
            &mut self.buffer,
            &mut self.outbound,
        ) {
            Ok(ProcessStatus::Continue) => SliceStep::Continue,
            Ok(ProcessStatus::Close) => self.finish_normal_close(pid),
            Ok(ProcessStatus::CloseWithFate(ConnectionFateClass::ServerShutdown)) => {
                self.runtime.take_force_close(pid);
                match self.handle_control(pid, ConnectionControl::ForceClose) {
                    Some(NativeOutcome::Stop(reason)) => SliceStep::Stop(reason),
                    _ => SliceStep::Continue,
                }
            }
            Ok(ProcessStatus::CloseWithFate(class)) => self.finish_fate_close(pid, class),
            Err(error) => {
                tracing::warn!(connection_pid = pid, %error, "connection process failed");
                self.release_conversations();
                self.runtime
                    .mark_crashed(pid, ExitReason::Error, self.peer_addr);
                SliceStep::Stop(ExitReason::Error)
            }
        }
    }

    /// R1(vi) (§1.2(3b)): services the pending-reply table for this slice.
    ///
    /// 1. DEADLINE-CHECK SEAM: expire every pending entry whose deadline passed,
    ///    tombstoning it and enqueuing its timeout error frame. Under the busy loop
    ///    this runs every slice; PARK-FLIP adds a timer-driven `READY` wake at each
    ///    entry's deadline (contract R1(vi) as amended) so a parked connection with
    ///    zero other traffic still wakes to write the timeout — that wake feeds this
    ///    same seam.
    /// 2. Drain and correlate: for each conversation still awaiting a reply, pull
    ///    buffered participant replies non-blocking and match them FIFO, enqueuing
    ///    each correlated reply frame. A conversation gone from the map (closed) is
    ///    swept from the table so its entries do not linger.
    ///
    /// # Errors
    /// Returns [`OutboundError`](super::outbound::OutboundError) when a reply or
    /// timeout frame cannot be enqueued (a fatal outbound condition).
    fn service_pending_replies(&mut self) -> Result<(), super::outbound::OutboundError> {
        let frames = super::reply_service::settled_frames(&mut self.state, &self.runtime)
            .map_err(super::outbound::OutboundError::LinkCalls)?;
        for frame in frames {
            self.outbound.enqueue_frame(&frame)?;
        }
        Ok(())
    }

    /// Removes and releases subscriptions the delivery pump shed on inbox overflow
    /// (§5). The pump has already enqueued each subscription's typed `SubscribeError`
    /// frame; here the subscription is dropped from connection state (its
    /// delivery-sequence counter and any held frame with it, so a re-subscribe that
    /// reuses the id starts clean) and released through the services adapter, the
    /// same teardown path an explicit `Unsubscribe` uses. A slow consumer thus sheds
    /// its own subscription without growing server memory or tearing down the
    /// connection's other streams.
    fn shed_subscriptions(&mut self, shed: Vec<u64>) {
        for subscription_id in shed {
            self.state.delivery_seqs.remove(&subscription_id);
            self.state.held_deliveries.remove(&subscription_id);
            if let Some(subscription) = self.state.subscriptions.remove(&subscription_id) {
                if let Err(error) = self.runtime.services().unsubscribe(subscription) {
                    tracing::warn!(
                        subscription_id,
                        %error,
                        "releasing a shed (inbox-overflowed) subscription failed"
                    );
                }
            }
        }
    }

    fn drain_outbound(
        &mut self,
    ) -> Result<super::outbound::DrainOutcome, super::outbound::OutboundError> {
        let Some(sink) = self.transport.sink() else {
            return Ok(super::outbound::DrainOutcome::Drained);
        };
        self.outbound.drain(sink, Some(READ_BUFFER_BYTES))
    }

    /// Makes one unbudgeted, nonblocking attempt to flush terminal output.
    ///
    /// The outbound queue is bounded, so removing the normal slice budget cannot
    /// create unbounded work. A full socket can still return
    /// [`DrainOutcome::WouldBlockWithResidue`]; terminal delivery is best-effort
    /// at that transport boundary and this method never waits or retries.
    fn drain_outbound_for_close(
        &mut self,
    ) -> Result<super::outbound::DrainOutcome, super::outbound::OutboundError> {
        let Some(sink) = self.transport.sink() else {
            return Ok(super::outbound::DrainOutcome::Drained);
        };
        self.outbound.drain(sink, None)
    }

    fn arm_readiness(
        &mut self,
        pid: u64,
        ctx: &NativeContext<'_>,
        interest: Interest,
    ) -> Result<(), ServerError> {
        self.transport
            .arm_readiness(pid, ctx, interest, &self.runtime)
    }

    fn sync_deadline_timers(
        &mut self,
        pid: u64,
        ctx: &mut NativeContext<'_>,
    ) -> Result<(), ServerError> {
        for timer in self.state.pending_replies.take_retired_timers() {
            ctx.cancel_timer(timer);
        }
        for (op_id, delay) in self
            .state
            .pending_replies
            .timers_to_arm(std::time::Instant::now())
        {
            let timer = ctx
                .send_after(delay, pid, Term::atom(self.runtime.ready_atom()))
                .ok_or_else(|| ServerError::ListenerAccept {
                    message: "connection scheduler has no timer facility for reply deadlines"
                        .to_owned(),
                })?;
            if !self.state.pending_replies.install_timer(op_id, timer) {
                ctx.cancel_timer(timer);
            }
        }
        Ok(())
    }

    /// Post-arm C1/C4 barrier. Each query is nonblocking and non-consuming,
    /// including the native mailbox so READY enqueued by this slice cannot be
    /// mistaken for quiescence after its inbox work was already consumed.
    fn final_probe(&self, pid: u64, ctx: &NativeContext<'_>) -> Result<bool, ServerError> {
        let socket_ready = self.transport.probe_inbound()?;
        let subscription_ready = !self.state.held_deliveries.is_empty()
            || self
                .state
                .subscriptions
                .values()
                .any(|subscription| subscription.is_overflowed() || subscription.has_pending());
        let participant_ready = if self.state.held_pushes.capacity_refused()
            && has_held_participant_head(&self.state)
        {
            false
        } else {
            match self.state.participant_publication.as_ref() {
                Some(inbox) => inbox.has_pending().map_err(participant_publication_error)?,
                None => false,
            }
        };
        let reply_ready = self
            .state
            .pending_replies
            .has_due(std::time::Instant::now())
            || self
                .state
                .pending_replies
                .conversations_with_entries()
                .into_iter()
                .filter_map(|id| self.state.conversations.get(&id))
                .any(super::conversation::ConnectionConversation::has_pending_reply);
        Ok(socket_ready
            || subscription_ready
            || participant_ready
            || reply_ready
            || super::supervisor::relay_pump::has_ready(&self.state, &self.outbound)?
            || self.runtime.has_control(pid)
            || self.runtime.ready_pending(pid)
            || ctx.has_messages())
    }

    /// Drains queued outbound bytes to the socket, if the stream is still present.
    ///
    /// Returns the [`DrainOutcome`](super::outbound::DrainOutcome) tri-state. Under
    /// the busy loop every caller ignores the distinction and re-services on the
    /// next slice; the value is reported so the seam is honest.
    ///
    /// PARK-FLIP SEAM (R2, §1.2(1)): when this returns
    /// `Ok(DrainOutcome::WouldBlockWithResidue)` the park-flip commit arms writable
    /// readiness interest for this connection's socket (and only then), so the
    /// connection re-wakes when the kernel send buffer drains instead of parking
    /// with unflushed residue. `Drained`/`Progress` arm nothing. The current live
    /// slice uses the fixed read-buffer budget; terminal close uses
    /// [`Self::drain_outbound_for_close`] instead.
    /// R3b F8, ruling 2a: once the buffer is empty, every frame enqueued
    /// since the last report has been written to the transport. Test-only.
    #[cfg(test)]
    fn report_frames_written(&mut self, pid: u64, drain: DrainOutcome) {
        if drain == DrainOutcome::Drained && self.outbound.queued_len() == 0 {
            self.runtime
                .record_frames_written(pid, self.outbound.take_enqueued());
        }
    }
}

impl<T: ConnectionTransport> NativeHandler for TransportConnectionProcess<T> {
    fn handle(&mut self, ctx: &mut NativeContext<'_>) -> NativeOutcome {
        let pid = ctx.self_pid();
        // FIX A-ii: the moment this connection is scheduled it is executing, so it
        // is no longer delivery-quiescent. Marked BEFORE the mailbox drain and
        // `acknowledge_ready` below, so the shutdown flush barrier can never sample
        // this connection as parked-with-no-pending-READY in the window between a
        // fan-out wake clearing `ready_pending` and its delivering slice starting.
        self.runtime.mark_running(pid);
        if let Some(outcome) = self.force_close_if_pending(pid) {
            return outcome;
        }
        // Registration is owned solely by the spawn thread (`SupervisorInner::
        // spawn_connection` calls `runtime.register` before returning the
        // handle), so the handler never writes the registry — it only reads its
        // record via `mark_crashed`/`finish`. This removes the previous
        // double-write (spawn-thread `insert` racing a handler `or_insert`) and
        // its lost-update/duplicate-record hazard.
        while let Some(message) = ctx.recv() {
            if let Some(outcome) = self.handle_message(pid, message) {
                return outcome;
            }
        }
        self.runtime.acknowledge_ready(pid);
        self.handle_slice(pid, ctx)
    }
}

impl<T: ConnectionTransport> Drop for TransportConnectionProcess<T> {
    fn drop(&mut self) {
        if let Some(pid) = self.agent_connection {
            crate::agents::runtime::connection_closed(
                self.runtime.services(),
                pid,
                "transport_drop",
            );
        }
        // Backstop for termination paths that never run another handler slice:
        // external termination (killed via the scheduler) and scheduler shutdown
        // drop the handler directly. This drop does NOT reclaim the host record;
        // W4 leg 1 delivers that reclamation TOLD, the instant beamr publishes
        // the process's exit event, through the supervisor's exit-event reactor
        // into the ordinary `remove()` funnel (replacing the retired per-accept
        // `reap_crashed` scan for this class). On the explicit in-handler teardown
        // paths the conversation map is already drained, so this is a no-op there.
        self.release_conversations();
        let timers = self.state.pending_replies.take_retired_timers();
        self.runtime.cancel_deadline_timers(timers);
        #[cfg(test)]
        {
            let runtime = Arc::clone(&self.runtime);
            self.transport.note_process_drop(&runtime);
        }
    }
}

/// A connection's read half, asked whether inbound bytes are already waiting.
///
/// [`ConnectionProcess::final_probe`] answers "is there inbound work pending"
/// before the connection parks, and that one term of the probe is a transport
/// fact rather than a connection fact: a socket answers with a non-consuming
/// `peek`, a transport with no socket answers from its own queued bytes. The
/// remaining terms (subscriptions, participant publications, reply deadlines,
/// control, mailbox) are shared across transports unchanged, so a sibling
/// process type replaces only this answer.
pub(super) trait InboundPending {
    /// Whether readable bytes are already waiting, without consuming them.
    ///
    /// `Interrupted` reports pending (the probe never ran, so quiescence is not
    /// established); `WouldBlock` is the honest "nothing waiting".
    ///
    /// # Errors
    /// Returns [`ServerError::ListenerAccept`] when the read half fails the
    /// probe with anything other than `WouldBlock` or `Interrupted`.
    fn inbound_pending(&self) -> Result<bool, ServerError>;
}

/// What one non-blocking inbound read found.
///
/// Shared across transports so every door's hangup, idle, and progress answers
/// select the SAME lifecycle branch in [`TransportConnectionProcess::
/// service_socket`] — a sibling cannot map its own end of file onto a different
/// fate by classifying it differently.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(super) enum ReadStatus {
    /// Bytes were appended to the buffer.
    Read,
    /// Nothing is available right now; the peer is still there.
    WouldBlock,
    /// End of file: the peer hung up and the read half is drained.
    Closed,
}