ts_runtime/

control_runner.rs

use core::{
    net::{Ipv4Addr, Ipv6Addr},
    time::Duration,
};
use std::{collections::HashMap, sync::Arc, time::Instant};

use futures::StreamExt;
use kameo::{
    actor::{ActorRef, Spawn},
    message::{Context, StreamMessage},
    prelude::Message,
};
use tokio::sync::watch;
use ts_control::{
    AsyncControlClient, Endpoint, EndpointType, Error as ControlError, IdTokenError, LogoutError,
    Node, SetDnsError, SshPolicy, StateUpdate, TkaStatus, TkaSyncError, tka_disable,
    tka_init_begin, tka_init_finish, tka_submit_signature,
};
use ts_magicsock::SelfEndpointType;

use crate::{
    derp_latency::{DerpLatencyMeasurement, DerpLatencyMeasurer},
    direct::EndpointAdvertisement,
};

/// Actor responsible for maintaining the connection to control.
///
/// This actor is responsible for proxying the map response stream onto the message bus.
pub struct ControlRunner {
    client: AsyncControlClient,
    params: Params,

    self_node: watch::Sender<Option<Node>>,
    /// Latest Tailscale SSH policy pushed by control, or `None` until control sends one. The SSH
    /// server reads this to authorize incoming connections; absent policy means deny-all.
    ssh_policy: watch::Sender<Option<SshPolicy>>,
    /// Latest Tailnet Lock status pushed by control, or `None` until control sends one.
    tka: watch::Sender<Option<TkaStatus>>,
    /// The locally-synced Tailnet-Lock state (verified `Authority` + AUM store), or `None` until a
    /// successful bootstrap+sync. Held here because `ControlRunner` owns the netmap stream that
    /// triggers resync. Mutated only on the actor thread (the netmap handler spawns the sync RPC and
    /// the result returns via the [`TkaSynced`] self-message).
    tka_synced: Option<crate::tka_sync::SyncedTka>,
    /// The Tailnet-Lock chain persisted beside the cached netmap for the cold-start replay, and the
    /// single ordered path that writes or removes it. See [`PersistedTkaChain`].
    tka_chain: PersistedTkaChain,
    /// The verified TKA [`Authority`](ts_tka::Authority) the peer tracker **enforces** (Go
    /// `tkaFilterNetmapLocked`). `None` until the first successful sync, and reset to `None` when the
    /// lock is disabled. This is the SOLE delivery channel to the peer tracker (which holds the
    /// matching `Receiver` and reads it on every peer upsert): a `watch` cell, not a bus message, so
    /// the latest value is always readable, never dropped under load, and writes are strictly ordered
    /// by this actor — a disable (`None`) can never be reordered behind or dropped before a stale
    /// `Some`. Written only from [`apply_tka_synced`] (enable) and [`maybe_sync_tka`] (disable), both
    /// on the actor thread. The published `Authority` has always passed `VerifiedAumChain::verify`.
    tka_authority: watch::Sender<Option<Arc<ts_tka::Authority>>>,
    /// In-flight guard: `true` while a sync RPC task is running, so a burst of netmap updates does
    /// not spawn overlapping syncs (Go serializes sync under `b.mu`).
    tka_syncing: bool,
    /// Monotonic generation stamped when a disable (or a fresh sync) supersedes any in-flight sync.
    /// `maybe_sync_tka` bumps this on a disable transition and captures it into each spawned sync;
    /// [`apply_tka_synced`] discards a sync result whose captured generation is stale, so a lock
    /// disabled *while a sync was in flight* is never re-enabled by that sync's late `Ok(Some)`
    /// (the in-flight window the `tka_synced.is_some()` disable guard alone does not cover).
    tka_generation: u64,
    /// Latest cert-domain list from control's netmap DNS config (Go `nm.DNS.CertDomains`), or empty
    /// until control sends a DNS config carrying one. The facade reads this for `Device::cert_domains`.
    cert_domains: watch::Sender<Vec<String>>,
    /// Latest full DNS config from control's netmap (Go `netmap.NetworkMap.DNS`), or `None` until
    /// control sends one. The facade reads this for `Device::dns_config` (the daemon's
    /// `tnet dns status`). A superset of [`cert_domains`](Self::cert_domains), which is kept as its
    /// own cell for the narrower TLS-cert use.
    dns_config: watch::Sender<Option<ts_control::DnsConfig>>,
    /// Latest interactive-login / consent URL control asked this node to open
    /// (`MapResponse.PopBrowserURL`), or `None` until control sends one. The facade reads this for
    /// `Device::pop_browser_url` (a daemon driving a non-authkey login surfaces it to the user), and
    /// [`Runtime::watch_ipn_bus`](crate::Runtime::watch_ipn_bus) subscribes to it for the bus's
    /// `browse_to_url` running-node events.
    ///
    /// **Sticky, not per-update** (Go `controlclient` `sess.lastPopBrowserURL`): control sends
    /// `MapResponse.PopBrowserURL` empty on nearly every netmap tick, so this cell is updated ONLY on
    /// a non-empty URL that differs from its current value (`sticky_update_pop_browser_url`, via
    /// `send_if_modified` — the cell's own value is the "last URL seen", so no separate mirror is
    /// needed). It is never reset to `None` by an empty update — matching Go's `direct.go` guard
    /// `u != "" && u != sess.lastPopBrowserURL`. Updating on every tick would thrash the cell to
    /// `None` and coalesce the URL away for a `watch` subscriber.
    pop_browser_url: watch::Sender<Option<url::Url>>,
    /// Latest network-conditions report (preferred DERP region + per-region latencies), updated each
    /// time the DERP-latency measurer reports in. The facade reads this for `Device::netcheck` (the
    /// daemon's `tnet netcheck`). Empty until the first measurement.
    netcheck: watch::Sender<crate::status::NetcheckReport>,
    /// The DERP home region currently selected, with the latency measured for it at selection time.
    /// `None` until the first home region is chosen. Used to apply selection **hysteresis** (Go
    /// `netcheck.addReportHistoryAndSetPreferredDERP`): the home region is only switched when a new
    /// region is *meaningfully* lower-latency than the current one, so jitter between near-equal
    /// regions does not flap the home relay (which would cause repeated reconnects + brief loss).
    home_region: Option<(ts_derp::RegionId, core::time::Duration)>,
    /// Rolling history of per-cycle DERP-latency reports within the last [`DERP_HISTORY_MAX_AGE`]
    /// (Go `netcheck` `maxAge`), each stamped with its arrival `Instant`. This is Go's one
    /// `netcheck.Client.prev` history and, like it, has two readers:
    ///
    /// 1. the `bestRecent` smoothing (Go `addReportHistoryAndSetPreferredDERP`) — the new home
    ///    candidate is chosen by each region's **minimum** latency over this window, not its raw
    ///    current sample, so a best region whose latency oscillates across the switch boundary does
    ///    not flap the home relay; and
    /// 2. [`ControlRunner::recent_region_latency`] (Go `RecentRegionLatency`) — the per-region
    ///    latencies the exit-node suggestion ranks candidates on, which need the union of several
    ///    *partial* measurements to cover every region at all.
    ///
    /// Aged entries are evicted on each measurement; the buffer is therefore bounded by the netcheck
    /// cadence × the window.
    derp_report_history: Vec<(Instant, Arc<Vec<ts_netcheck::RegionResult>>)>,
    /// Consecutive automatic-reauth attempts that have NOT yet recovered the node to a good (non-
    /// expired) self-node. The circuit breaker for [`expiry_action`]'s `Reauthenticate` path: a
    /// one-shot / already-consumed auth key cannot re-register, so without a bound an expired node
    /// would sit in [`DeviceState::Reauthenticating`] indefinitely (the rotated re-register keeps
    /// failing or control keeps returning a still-expired self-node).
    ///
    /// Incremented once per expired self-node seen while *already* `Reauthenticating` (each such
    /// arrival is evidence the prior reauth did not recover); reset to `0` whenever a good
    /// (non-expired) self-node arrives (the node recovered) — see the [`StreamMessage::Next`] handler.
    /// At [`MAX_REAUTH_ATTEMPTS`] the runner stops re-arming reauth and flips the cell to the terminal
    /// [`DeviceState::Expired`], giving the cell a stable terminal state (a genuinely good self-node
    /// can still recover it later). The one-shot `Command::Reauth` is fired ONLY on the transition
    /// INTO `Reauthenticating`, never re-fired while already reauthenticating, so the node key is
    /// rotated at most once per episode (a second rotation would lose the original `OldNodeKey`
    /// anchor control needs to link the rotation).
    reauth_attempts: u32,
    /// Background task that bridges the control client's mid-session re-auth URL cell onto
    /// [`Self::params`]'s device-state cell (sets [`DeviceState::NeedsLogin`] when control returns
    /// `MachineNotAuthorized` on a live re-register — see [`bridge_reauth_url_to_state`]). Aborted on
    /// [`Drop`] so it cannot outlive the actor (the [`DataplaneActor`](crate::dataplane) pattern).
    reauth_bridge: tokio::task::JoinHandle<()>,
}

/// The number of consecutive failed automatic-reauth attempts after which the runner gives up
/// re-arming reauth and flips the device to the terminal [`DeviceState::Expired`] (the circuit
/// breaker from the design's deferred-question Q1). A one-shot auth key was consumed by the first
/// registration, so an auto re-register with it cannot succeed; this bounds that case to today's
/// terminal behavior instead of an indefinite `Reauthenticating` spell. The first arrival fires the
/// reauth; each subsequent expired self-node while still reauthenticating counts toward this cap.
const MAX_REAUTH_ATTEMPTS: u32 = 3;

impl Drop for ControlRunner {
    fn drop(&mut self) {
        // Stop the re-auth bridge so it does not outlive the actor (mirrors `DataplaneActor`).
        self.reauth_bridge.abort();
    }
}

/// Control runner args.
pub struct Params {
    /// Control config.
    pub(crate) config: ts_control::Config,

    /// Auth key (if needed).
    pub(crate) auth_key: Option<String>,

    /// The [`crate::Env`] for this actor.
    pub(crate) env: crate::Env,

    /// Sender for the device connection-state cell. Created in [`Runtime::spawn`](crate::Runtime)
    /// so it outlives the actor's `on_start` (which may publish [`DeviceState::Failed`] and then
    /// return `Err`, before `Self` exists). The runtime keeps the matching `Receiver` for
    /// [`watch_state`](crate::Runtime::watch_state) / [`wait_until_running`](crate::Runtime::wait_until_running).
    pub(crate) state_tx: watch::Sender<crate::DeviceState>,

    /// Sender for the TKA enforcement-authority cell the peer tracker reads (Go
    /// `tkaFilterNetmapLocked`). Created in [`Runtime::spawn`](crate::Runtime) and threaded into BOTH
    /// the peer tracker (the `Receiver`) and this runner (the `Sender`), so the runner is the sole
    /// writer and the tracker reads the latest verified `Authority` on demand. `None` = no lock /
    /// disabled (admit all).
    pub(crate) tka_authority: watch::Sender<Option<Arc<ts_tka::Authority>>>,

    /// Sender for the selected DERP home region (the **smoothed** `bestRecent` + hysteresis choice,
    /// Go `report.PreferredDERP`). Created in [`Runtime::spawn`](crate::Runtime); the runner is the
    /// sole writer and [`Multiderp`](crate::multiderp) holds the `Receiver`, so the local DERP relay
    /// follows the SAME home the runner advertises to control — not the raw per-cycle latency
    /// minimum (which would flap on jitter and disagree with the advertised home). `None` until the
    /// first home is chosen.
    pub(crate) home_region: watch::Sender<Option<ts_derp::RegionId>>,
}

#[doc(hidden)]
#[derive(Debug, thiserror::Error)]
pub enum ControlRunnerError {
    #[error(transparent)]
    Control(#[from] ControlError),

    #[error(transparent)]
    Crate(#[from] crate::Error),
}

impl kameo::Actor for ControlRunner {
    type Args = Params;
    type Error = ControlRunnerError;

    async fn on_start(params: Params, slf: ActorRef<Self>) -> Result<Self, Self::Error> {
        // Cold start: replay the netmap this node cached the last time control granted it
        // `cache-network-maps` (Go `nodecap.CacheNetworkMaps`, capability version 135 — upstream
        // does the same from `ipnlocal.Start`, feeding the cached map through
        // `setControlClientStatusLocked` before it builds its control client). Publishing it here,
        // ahead of registration, is the whole point: the peer tracker, dataplane and packet filter
        // get the last-known peers, DERP map and rules while this node is still doing its
        // register/poll round trips, so peer connectivity can start coming up before control has
        // said a word. Control's own first netmap lands on the same bus moments later and replaces
        // it, and a netmap that no longer carries the attribute deletes the cache (see
        // `NetmapCache::observe`).
        //
        // Nothing here can fail the start-up: no cache configured, nothing cached, or an
        // undecodable cache all mean "carry on without one".
        replay_cached_netmap(&params).await;

        // The interactive AuthURL, captured on the first unauthorized reply and reused as the
        // `followup` on every subsequent poll so control long-polls ONE stable URL (rather than
        // minting a fresh, racing URL each retry) until the user visits it.
        let mut login_url: Option<url::Url> = None;
        loop {
            match AsyncControlClient::check_auth(
                &params.config,
                &params.env.keys,
                params.auth_key.as_deref(),
                login_url.as_ref(),
            )
            .await
            {
                Ok(()) => break,
                Err(ControlError::MachineNotAuthorized(u)) => {
                    // Capture the FIRST url and keep showing/following-up on it, so the link the
                    // user opened stays valid instead of being superseded by the next poll.
                    let url = login_url.get_or_insert(u).clone();
                    tracing::info!(auth_url = %url, "please authorize this machine or pass an auth key");
                    // Publish `NeedsLogin(url)` only when it actually changes the cell. With `followup`
                    // set, the SAME URL is re-affirmed on every long-poll timeout; a bare `send_replace`
                    // re-notifies `state_rx` each cycle, so the bus re-emits `browse_to_url` for a link
                    // the user already has — reopening it repeatedly. `send_if_modified` dedups the
                    // no-op, mirroring `bridge_reauth_url_to_state` (the mid-session re-auth path).
                    let next = crate::DeviceState::NeedsLogin(url);
                    params.state_tx.send_if_modified(|current| {
                        if *current == next {
                            false
                        } else {
                            *current = next.clone();
                            true
                        }
                    });
                    // With followup set, check_auth long-polls until the URL is visited; this short
                    // sleep only applies if control times the poll out, and we re-followup the SAME url.
                    tokio::time::sleep(Duration::from_secs(2)).await;
                }
                Err(ControlError::NeedsMachineAuth) => {
                    // The node is registered with a valid key but awaiting ADMIN APPROVAL on an
                    // approval-gated tailnet, and control offered NO interactive URL. This is
                    // TRANSIENT (Go's `NeedsMachineAuth`): poll registration until an admin approves,
                    // then `check_auth` returns `Ok(())` → the loop breaks and the node comes up with
                    // no re-registration. Publishing the (no-URL) `NeedsMachineAuth` state — NOT a
                    // terminal `Failed` and NOT `NeedsLogin` (there is no URL to open) — lets a
                    // watcher / `wait_until_running` see "awaiting approval" instead of an opaque
                    // timeout. Same 5s poll cadence as the `MachineNotAuthorized(url)` arm.
                    tracing::info!(
                        "machine awaiting admin approval to join the tailnet; polling until approved"
                    );
                    params
                        .state_tx
                        .send_replace(crate::DeviceState::NeedsMachineAuth);
                    tokio::time::sleep(Duration::from_secs(5)).await;
                }
                Err(ControlError::RateLimited(retry_after)) => {
                    // Control asked us to slow down (HTTP 429). Wait exactly the server-requested
                    // cooldown and retry — this is transient, NOT a terminal `Failed`, so we must
                    // not stop the runner (mirrors Go's `authRoutine` sleeping `rle.retryAfter`).
                    tracing::warn!(
                        ?retry_after,
                        "control rate-limited registration; waiting before retry"
                    );
                    tokio::time::sleep(retry_after).await;
                }
                Err(e) => {
                    // Followup auth path gone: while long-polling a STABLE AuthURL, control returns
                    // an HTTP registration error (410 "auth path not found") once the path is either
                    // VISITED + approved (consumed) or expired. This is NOT terminal — drop the
                    // followup and re-register ONCE: an approved node key comes back `MachineAuthorized`
                    // (`Ok(())` → login completes), an expired one comes back with a fresh AuthURL.
                    // Without this, the user's own approval (which consumes the path → 410) kills the
                    // runner. Bounded: on re-register `login_url` is None, so a repeat HTTP error falls
                    // through to the terminal handling below (no infinite loop).
                    if login_url.is_some()
                        && matches!(
                            e,
                            ControlError::Internal(
                                ts_control::InternalErrorKind::Http,
                                ts_control::Operation::Registration,
                            )
                        )
                    {
                        tracing::info!(
                            "followup auth path gone (approved or expired); re-registering"
                        );
                        login_url = None;
                        tokio::time::sleep(Duration::from_secs(1)).await;
                        continue;
                    }
                    // A hard registration failure (bad/expired/unknown auth key, etc.). Log the
                    // specific reason control gave AND publish it as a typed `Failed` state so
                    // `Device::wait_until_running` returns the actionable reason (tsr-kqj) instead
                    // of the opaque `Internal(Actor)` the caller would otherwise see once the
                    // stopped actor is next asked. Publishing before `return Err` is why the state
                    // sender lives on `Runtime`, not on `Self` (which never gets constructed here).
                    let reason = crate::RegistrationError::from(&e);
                    tracing::error!(error = %e, "registration failed; control runner stopping");
                    params
                        .state_tx
                        .send_replace(crate::DeviceState::Failed(reason));
                    return Err(e.into());
                }
            }
        }
        // check_auth succeeded, but the node is not "up" until the netmap stream is actually
        // attached below. Publish `Running` only AFTER `attach_stream` so `wait_until_running` never
        // resolves `Ok` for a device whose stream connect failed (which would leave a stopped actor
        // behind). If the connect/subscribe steps fail, publish a transient `Failed` first so the
        // waiter sees an actionable reason instead of the opaque post-mortem `Internal(Actor)`.
        // The control client's live map-poll loop publishes a mid-session re-auth URL here (set when
        // a re-register returns `MachineNotAuthorized` because the node key expired/was revoked). The
        // runtime owns the receiver; `connect` takes the sender. Created before `connect` so the
        // sender is in place for the very first poll, and so the receiver outlives `bring_up`.
        let (auth_url_tx, auth_url_rx) = watch::channel::<Option<url::Url>>(None);

        // `connect` issues its own `machine/register` POST (a second one after `check_auth`'s), so
        // it too can hit a 429. Wrap it in the same honor-`Retry-After` retry as the `check_auth`
        // loop above: a rate-limit is transient — sleeping the server-requested cooldown and
        // retrying must NOT stop the runtime (a 429 here previously fell into the `Err(e)` arm →
        // `Failed` → actor stop). `connect` consumes a `watch::Sender` by value (and drops it on a
        // failed register, before it would be moved into the live-poll task), so we keep the original
        // `auth_url_tx` alive here across all attempts and hand `connect` a CLONE each time. That
        // keeps the runtime's `auth_url_rx` (the `reauth_bridge` receiver) paired with a live sender:
        // recreating a fresh sender per attempt instead would orphan the bridge the moment the
        // original dropped, silently killing mid-session re-auth-URL delivery.
        let client = loop {
            let bring_up = async {
                let (client, stream) = AsyncControlClient::connect(
                    &params.config,
                    &params.env.keys,
                    params.auth_key.as_deref(),
                    auth_url_tx.clone(),
                )
                .await?;

                DerpLatencyMeasurer::spawn_link(&slf, params.env.clone()).await;

                params.env.subscribe::<DerpLatencyMeasurement>(&slf).await?;
                params.env.subscribe::<EndpointAdvertisement>(&slf).await?;
                slf.attach_stream(stream.boxed(), (), ());
                Ok::<_, ControlRunnerError>(client)
            };

            match bring_up.await {
                Ok(client) => break client,
                Err(ControlRunnerError::Control(ControlError::RateLimited(retry_after))) => {
                    tracing::warn!(
                        ?retry_after,
                        "control rate-limited the session bring-up; waiting before retry"
                    );
                    tokio::time::sleep(retry_after).await;
                }
                Err(ControlRunnerError::Control(ControlError::NeedsMachineAuth)) => {
                    // `connect` issues its OWN `machine/register` POST (a second one after
                    // `check_auth`'s), so it too can come back "awaiting admin approval, no URL". In
                    // the normal flow the `check_auth` loop above already gated this (it only breaks
                    // once registration returns `Ok`, and approval is monotonic), so this arm is the
                    // defensive twin for a node de-authorized in the race window between the two POSTs:
                    // treat it as TRANSIENT exactly like the `check_auth` arm — publish the (no-URL)
                    // `NeedsMachineAuth` state, poll the same 5s, and retry the bring-up — rather than
                    // collapsing into the terminal `Failed` arm below (which would permanently stop
                    // the runner on a recoverable await-approval). Mirrors Go's `NeedsMachineAuth`.
                    tracing::info!(
                        "machine awaiting admin approval during session bring-up; polling until \
                         approved"
                    );
                    params
                        .state_tx
                        .send_replace(crate::DeviceState::NeedsMachineAuth);
                    tokio::time::sleep(Duration::from_secs(5)).await;
                }
                Err(e) => {
                    tracing::error!(error = %e, "bringing up the control session failed");
                    // The control session never came up; surface it as a transient registration
                    // failure (a retry / fresh `Device::new` may succeed) rather than leaving the
                    // state stuck at `Connecting`.
                    params.state_tx.send_replace(crate::DeviceState::Failed(
                        crate::RegistrationError::NetworkUnreachable,
                    ));
                    return Err(e);
                }
            }
        };

        // The netmap stream is attached: the node is up. The stream `Next` handler keeps this
        // current (and flips to `Expired` if the self-node's key lapses).
        params.state_tx.send_replace(crate::DeviceState::Running);

        // Bridge the control client's mid-session re-auth URL cell onto the device-state cell: a
        // `Some(url)` (control returned `MachineNotAuthorized` on a live re-register) becomes
        // `DeviceState::NeedsLogin(url)` so the IPN bus surfaces `browse_to_url` and the embedder can
        // prompt the user — the live-session analogue of the initial `check_auth` loop above. The
        // recovery to `Running` is the netmap self-node handler's job (next good self-node), so this
        // bridge only forwards `Some`. The task ends when the sender drops (the client's `run` task
        // ended) and is aborted on actor `Drop`, so it cannot leak past the actor.
        let reauth_bridge = {
            let state_tx = params.state_tx.clone();
            let mut auth_url_rx = auth_url_rx;
            tokio::spawn(async move {
                while auth_url_rx.changed().await.is_ok() {
                    let url = auth_url_rx.borrow_and_update().clone();
                    bridge_reauth_url_to_state(&state_tx, url.as_ref());
                }
            })
        };

        // Clone the TKA authority publisher before `params` moves into `Self` below. The matching
        // `Receiver` lives on the peer tracker; this sender is the sole writer (enforce on sync,
        // clear on disable).
        let tka_authority = params.tka_authority.clone();

        // Same reason: the owner of the persisted Tailnet-Lock chain is built from the config before
        // `params` moves. It starts out assuming a chain may already be on disk — an earlier process
        // of this node can have left one there, and this one has synced nothing yet.
        let tka_chain = PersistedTkaChain::new(params.config.netmap_cache_dir.as_ref());

        Ok(Self {
            client,
            params,
            self_node: Default::default(),
            ssh_policy: Default::default(),
            tka: Default::default(),
            tka_synced: None,
            tka_chain,
            tka_authority,
            tka_syncing: false,
            tka_generation: 0,
            cert_domains: Default::default(),
            dns_config: Default::default(),
            pop_browser_url: Default::default(),
            netcheck: Default::default(),
            home_region: None,
            derp_report_history: Vec::new(),
            reauth_attempts: 0,
            reauth_bridge,
        })
    }
}

impl ControlRunner {
    /// Decide whether the latest netmap's Tailnet-Lock status warrants a (re)sync and, if so, spawn
    /// the bootstrap+sync RPC off the actor thread (so the netmap stream never blocks on a control
    /// round-trip). The result returns via the [`TkaSynced`] self-message.
    ///
    /// Triggers when control reports TKA enabled (`is_enabled`) AND we are not already syncing AND
    /// either we hold no `Authority` yet (→ bootstrap) or control's head differs from ours (→ catch
    /// up). When TKA is disabled, clears any synced state (the lock was turned off). Mirrors Go's
    /// `tkaSyncIfNeeded`: a no-op when our head already matches.
    async fn maybe_sync_tka(&mut self, tka: &TkaStatus, self_ref: ActorRef<Self>) {
        if !tka.is_enabled() {
            // Lock disabled (or never enabled): clear enforcement by writing `None` to the authority
            // cell the peer tracker reads — synchronously, so it can never be reordered behind or
            // dropped before a stale `Some` (the failure a best-effort broadcast had). Always bump the
            // generation so ANY sync currently in flight is invalidated: without this, a disable that
            // races an in-flight sync (whose `take()` already cleared `tka_synced`) would be a no-op
            // here, and the sync's late `Ok(Some)` would silently re-enable a lock control just turned
            // off (the in-flight window the `tka_synced.is_some()` guard alone misses). Cheap and
            // idempotent: clearing an already-`None` cell and bumping the generation are harmless.
            self.tka_generation = self.tka_generation.wrapping_add(1);
            if self.tka_synced.is_some() {
                tracing::info!("TKA lock disabled; clearing enforcement (admitting all peers)");
                self.tka_synced = None;
            }
            // Outside that guard, deliberately: the chain on disk is not this process's to have
            // synced. A cold start holds `tka_synced == None` while an earlier process's chain is
            // still in the cache directory, and that chain is exactly the one the *next* cold start
            // would find. So a lock control reports off clears it whether or not we ever enforced it.
            // Dropping it is not what keeps the next cold start honest (the replay only vouches when
            // the cached frame itself says the lock was on, and only with a chain at that frame's
            // head) — it is so a disabled lock leaves nothing of itself on disk.
            self.tka_chain.discard().await;
            self.tka_authority.send_replace(None);
            return;
        }
        if self.tka_syncing {
            return; // a sync is already in flight; the next netmap will re-trigger if still stale
        }
        // Up-to-date check: if we already have an Authority whose head matches control's, nothing to
        // do. A malformed control head is treated as "different" (we'll attempt a sync, which
        // fail-closes harmlessly).
        if let Some(synced) = &self.tka_synced
            && let Some(control_head) = ts_tka::AumHash::from_base32(&tka.head)
            && synced.authority.head_matches(&control_head)
        {
            return;
        }

        // Spawn the sync. Move the current synced state out (the driver takes it by value and returns
        // the advanced state); `tka_synced` stays `None` until the result lands, guarded by
        // `tka_syncing` so we don't spawn a second concurrent sync. Capture the current generation so
        // `apply_tka_synced` can discard this result if a disable bumped the generation while the sync
        // was in flight (H1: don't re-enable a lock that was disabled mid-sync).
        self.tka_syncing = true;
        let generation = self.tka_generation;
        let current = self.tka_synced.take();
        let config = self.params.config.clone();
        let keys = self.params.env.keys.clone();
        tokio::spawn(async move {
            let result = crate::tka_sync::sync_tka(&config, &keys, current).await;
            // Hand the outcome back to the actor thread to apply (mutating actor state off-thread is
            // not allowed). A send failure just means the actor is gone — nothing to do.
            if let Err(e) = self_ref.tell(TkaSynced { result, generation }).await {
                tracing::debug!(error = ?e, "TKA sync result not delivered (actor gone)");
            }
        });
    }

    /// Apply the outcome of a spawned [`maybe_sync_tka`] task on the actor thread: store the advanced
    /// state + publish the `Authority` to the peer tracker's enforcement cell (or, on inert/failed
    /// sync, leave peers unaffected). Always clears the in-flight guard.
    ///
    /// `generation` is the value captured when the sync was spawned. If it no longer matches
    /// `self.tka_generation`, the lock was disabled (or re-synced) while this sync was in flight, so
    /// the result is discarded — never re-enabling an authority control has since turned off.
    async fn apply_tka_synced(
        &mut self,
        result: Result<Option<crate::tka_sync::SyncedTka>, crate::tka_sync::TkaSyncDriverError>,
        generation: u64,
    ) {
        self.tka_syncing = false;

        // H1 guard: a disable (or a superseding sync) bumped the generation while this sync ran. Drop
        // the stale result — `maybe_sync_tka`'s disable branch already cleared enforcement to `None`,
        // and re-applying this `Some` would re-enforce a lock that is no longer active.
        if generation != self.tka_generation {
            tracing::info!(
                "TKA sync result superseded (lock disabled or re-synced mid-flight); discarding"
            );
            return;
        }

        match result {
            Ok(Some(synced)) => {
                tracing::info!(
                    head = %synced.authority.head().to_base32(),
                    "TKA sync succeeded; enforcing verified Authority (Go tkaFilterNetmapLocked)"
                );
                // Deliver the verified Authority to the peer tracker's enforcement cell. The tracker
                // reads it on every peer upsert and drops unauthorized peers. `Some(..)` = enforce; a
                // `None` is written on disable. `watch` is the sole channel (last-write-wins, never
                // dropped, ordered by this actor) — no bus, no re-publish-for-replay needed.
                self.tka_authority
                    .send_replace(Some(synced.authority.clone()));

                // Observability (Go `tkaFilterNetmapLocked`'s self check → `LockedOut` health
                // warning): verify SELF's own node-key signature against the freshly-synced
                // Authority and warn if self is NOT authorized. We never FILTER self (self never
                // enters the peer db, so enforcement can't lock us out of our own netmap), but Go
                // raises an operator-facing warning here because a self that the lock does not
                // authorize means this node's key-signature is missing/invalid for the current lock
                // — it will be unable to prove itself to locked peers. This fork has no health
                // subsystem, so the signal is a `tracing::warn!` (its observability channel).
                //
                // `self_node` is a sticky cell set on every netmap carrying a self-node; if a sync
                // somehow lands before the first self-node ever arrived it is `None`, so we skip the
                // advisory this cycle and re-evaluate on the next sync — fine for observability-only.
                // The `borrow()` ref is scoped to this `if let` and dropped before the `&mut self`
                // write below.
                if let Some(self_node) = self.self_node.borrow().as_ref() {
                    log_self_lockout(self_node, &synced.authority);
                }

                // Persist the verified chain beside the cached netmap so the *next* cold start can
                // run this same filter over the cached peers before control answers — Go's cold
                // start filters its cached map against the authority it re-opens from disk, and this
                // is the disk it re-opens (see `replay_cached_netmap`).
                self.persist_tka_chain(&synced).await;

                self.tka_synced = Some(synced);
            }
            Ok(None) => {
                // Control has no lock for us (no genesis / disabled). Clear any authority we were
                // previously enforcing — symmetric with the disable path — so a transition to
                // "no lock" stops dropping peers. Not an error.
                if self.tka_synced.is_some() {
                    tracing::info!("TKA sync: control reports no lock; clearing enforcement");
                    self.tka_synced = None;
                }
                // Outside that guard for the same reason as the disable path: a chain an earlier
                // process persisted outlives the synced state this one holds.
                self.tka_chain.discard().await;
                self.tka_authority.send_replace(None);
            }
            Err(e) => {
                // Transport or verify failure: log and leave the prior authority in place (a failed
                // sync must not drop enforcement — that would fail OPEN). NEVER errors the netmap.
                // The next netmap update re-triggers a sync attempt.
                tracing::warn!(error = %e, "TKA sync failed; keeping prior enforcement state");
            }
        }
    }

    /// Persist the freshly-synced AUM chain next to the cached netmap, so a cold start can vouch for
    /// the peers in that netmap ([`load_cached_netmap`]).
    ///
    /// Written only where the netmap itself is written: with no
    /// [`netmap_cache_dir`](ts_control::Config::netmap_cache_dir) there is nothing to vouch for, and
    /// with the `cache-network-maps` grant absent or overridden nothing is cached either, so the
    /// chain is removed instead of written — the same withdrawal rule `NetmapCache::observe` applies
    /// to the netmap. The grant is read from the last self node control sent, which is the same
    /// source `observe` reads it from: nothing is ever cached before one arrives, so "no self node
    /// yet" is also "nothing to vouch for". A write failure is logged inside the cache and changes
    /// nothing: the next cold start simply withholds the cached peers.
    async fn persist_tka_chain(&mut self, synced: &crate::tka_sync::SyncedTka) {
        // Scoped so the `watch` borrow is dropped before the awaits below (it is not `Send`).
        let caching_granted = {
            let self_node = self.self_node.borrow();
            self_node
                .as_ref()
                .is_some_and(|node| ts_control::netmap_caching_enabled(&node.cap_map))
        };
        if !caching_granted {
            self.tka_chain.discard().await;
            return;
        }

        match crate::tka_sync::encode_chain(&synced.store, synced.oldest) {
            Some(blob) => self.tka_chain.store(&blob).await,
            None => {
                // The store cannot be walked genesis→head, so there is no chain to persist. Drop any
                // older one rather than leaving a chain that no longer matches what we enforce.
                tracing::warn!("TKA: synced chain is not walkable; not persisting it");
                self.tka_chain.discard().await;
            }
        }
    }

    fn with_self_node<F, R>(&self, f: F) -> impl Future<Output = Option<R>> + use<F, R>
    where
        F: FnOnce(&Node) -> R,
    {
        let mut sub = self.self_node.subscribe();
        let mut shutdown = self.params.env.shutdown.clone();

        async move {
            tokio::select! {
                _ = shutdown.wait_for(|x| *x) => {
                    None
                },
                node = sub.wait_for(Option::is_some) => {
                    Some(f(node.ok()?.as_ref()?))
                },
            }
        }
    }
}

/// Apply Go's sticky `PopBrowserURL` semantics to the consent-URL `watch` cell.
///
/// Control sends `MapResponse.PopBrowserURL` empty on nearly every netmap update, so the cell is
/// updated ONLY when `incoming` is a non-empty URL that differs from the cell's current value —
/// Go's `direct.go` guard `u != "" && u != sess.lastPopBrowserURL`. The cell is **never reset to
/// `None`** by an empty/absent update — the running-node consent URL is sticky for the session.
/// Updating unconditionally would thrash the cell to `None` on every tick and coalesce the URL away
/// for a `watch`/bus subscriber.
///
/// The dedupe is in-place via [`watch::Sender::send_if_modified`] — the cell's own value is the
/// "last URL sent" (this sticky path is its only writer), so no separate mirror field is needed and
/// the watch is woken only on a genuine change (Go's `sess.lastPopBrowserURL` role, for free). This
/// matches the [`send_if_modified`](watch::Sender::send_if_modified) idiom already used for the
/// device-state cell in this handler.
///
/// Factored out of the netmap-update handler so the (easy-to-regress) sticky logic is unit-testable
/// against a plain `watch` channel without standing up the actor.
fn sticky_update_pop_browser_url(
    cell: &watch::Sender<Option<url::Url>>,
    incoming: Option<&url::Url>,
) {
    if let Some(url) = incoming {
        cell.send_if_modified(|current| {
            if current.as_ref() == Some(url) {
                false
            } else {
                *current = Some(url.clone());
                true
            }
        });
    }
}

/// Map a mid-session re-auth URL surfaced by the control client onto the device-state cell.
///
/// The control client's live map-poll loop publishes an `Option<url::Url>` into a `watch` cell when
/// a re-register hits `MachineNotAuthorized` (the node key expired/was revoked mid-session — see
/// [`ts_control::AsyncControlClient::connect`]'s `auth_url_tx`). `ts_control` cannot name
/// [`DeviceState`] (it must not depend on this crate), so this bridge fn does the translation:
/// a `Some(url)` sets [`DeviceState::NeedsLogin`]`(url)` so the IPN bus derives `browse_to_url` and
/// the embedder can prompt the user, exactly like the initial-registration `check_auth` path.
///
/// **Only `Some` drives a transition; `None` is ignored here.** The clear back to
/// [`DeviceState::Running`] is owned by the netmap self-node handler (the next good self-node flips
/// it — see the `StreamMessage::Next` arm), which is the authoritative "we are up again" signal; an
/// independent `None`-clear in this bridge could race that and is unnecessary. The
/// [`send_if_modified`](watch::Sender::send_if_modified) guard fires the watch only on a genuine
/// state change (it is a no-op when the cell already holds `NeedsLogin(url)` for the same URL), so a
/// re-auth URL re-surfaced across retries does not thrash the cell — mirroring the device-state
/// dedupe in the netmap handler.
///
/// Factored out so the (regress-prone) map-and-guard is unit-testable against a plain `watch`
/// channel without standing up the actor (mirrors [`sticky_update_pop_browser_url`]).
/// **Auto-reauth ownership.** While an automatic re-auth is in flight (the cell holds
/// [`DeviceState::Reauthenticating`]), this bridge must NOT downgrade it to `NeedsLogin`. An
/// auto-reauth's rotated re-register surfaces an auth URL through the SAME `auth_url_tx` cell this
/// bridge watches, but on a headless / auth-key node there is no human to visit it — flipping to
/// `NeedsLogin` would surface a misleading `browse_to_url` AND, by moving the cell off
/// `Reauthenticating`, let the next expired self-node re-fire reauth (a second node-key rotation that
/// loses the original `OldNodeKey` anchor). The auto-reauth path owns the cell until it recovers to
/// `Running` (the netmap self-node handler) or its circuit breaker trips it to `Expired`; this bridge
/// stands down for the duration.
pub(crate) fn bridge_reauth_url_to_state(
    state_tx: &watch::Sender<crate::DeviceState>,
    incoming: Option<&url::Url>,
) {
    if let Some(url) = incoming {
        let next = crate::DeviceState::NeedsLogin(url.clone());
        state_tx.send_if_modified(|current| {
            // Do not clobber an in-flight automatic re-auth (see the ownership note above): leave
            // `Reauthenticating` untouched so it can recover to `Running` or trip to `Expired` on its
            // own terms.
            if *current == crate::DeviceState::Reauthenticating || *current == next {
                false
            } else {
                *current = next.clone();
                true
            }
        });
    }
}

/// What to do when control delivers a self-node whose node-key expiry has passed — the decision
/// behind the [`StreamMessage::Next`] handler's expiry branch, factored into a pure function so the
/// full input matrix is unit-testable (mirrors [`bridge_reauth_url_to_state`] being pure).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum ExpiryAction {
    /// The key is not expired — the node is up. (`→ DeviceState::Running`.)
    Running,
    /// The key expired and auto-reauth is permitted (auth key retained, reauth enabled, TKA NOT
    /// enforcing): rotate the node key + re-register (`→ DeviceState::Reauthenticating` +
    /// `Command::Reauth`).
    Reauthenticate,
    /// The key expired and auto-reauth is NOT permitted (no auth key, reauth disabled, or TKA
    /// enforcing): fall back to today's terminal behavior (`→ DeviceState::Expired`).
    Expired,
}

/// Decide the action for an expired-or-not self node (pure; the live handler at
/// [`StreamMessage::Next`] applies it). Go's `ipnlocal` runs `doLogin` (rotate the node key +
/// re-register with the stored auth key) when an auth-key node's key expires; this fork does the
/// same, gated by three safety conditions:
///
/// - `key_expired` — control reported the self-node's key expiry is in the past.
/// - `has_auth_key` — a usable auth key is retained for a non-interactive re-register (without one
///   there is nothing to re-register with → fall back to `Expired`, today's behavior).
/// - `reauth_enabled` — the `reauth_on_expiry` config opt-out is on (default true).
/// - `tka_active` — Tailnet Lock enforcement is currently active. **Hard safety gate:** a node-key
///   rotation on a locked tailnet would install an UNSIGNED key, locking the node out of locked
///   peers (the TKA re-sign is a separate follow-up — see `keystate.rs` `rotate_node_key`). So when
///   the lock is enforcing, never rotate — fall back to `Expired`.
///
/// Truth table: not expired → `Running`; expired AND auth-key AND reauth-enabled AND NOT TKA →
/// `Reauthenticate`; otherwise → `Expired`.
pub(crate) fn expiry_action(
    key_expired: bool,
    has_auth_key: bool,
    reauth_enabled: bool,
    tka_active: bool,
) -> ExpiryAction {
    if !key_expired {
        return ExpiryAction::Running;
    }
    if has_auth_key && reauth_enabled && !tka_active {
        ExpiryAction::Reauthenticate
    } else {
        ExpiryAction::Expired
    }
}

/// The outcome of one step of the bounded reauth sub-state-machine: the device state to publish, the
/// updated consecutive-attempt counter, and whether to fire the one-shot `Command::Reauth` this step.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) struct ReauthStep {
    /// The device state to publish for this self-node.
    pub next: ReauthState,
    /// The consecutive-failed-reauth counter after this step (the caller stores it back).
    pub attempts: u32,
    /// Whether to fire the one-shot `Command::Reauth` (rotate + re-register) this step. Set ONLY on
    /// the transition INTO reauthenticating, so the node key rotates at most once per episode.
    pub fire_reauth: bool,
}

/// The device state a [`ReauthStep`] resolves to — the subset of [`DeviceState`](crate::DeviceState)
/// the circuit breaker can produce. Kept as its own enum so the breaker stays pure (no dependency on
/// the URL-carrying `DeviceState` variants) and the truth table is exhaustively testable.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum ReauthState {
    /// The node is up — `→ DeviceState::Running`.
    Running,
    /// An automatic re-auth is in flight — `→ DeviceState::Reauthenticating`.
    Reauthenticating,
    /// Terminal expiry — `→ DeviceState::Expired` (either the non-reauth path or the breaker tripped).
    Expired,
}

/// One step of the bounded reauth sub-state-machine (pure; the [`StreamMessage::Next`] handler stores
/// the result back into [`ControlRunner::reauth_attempts`] and publishes [`ReauthStep::next`]). This
/// is the circuit breaker for the design's deferred-question Q1: a one-shot / already-consumed auth
/// key cannot re-register, so the `Reauthenticate` path must settle rather than loop forever.
///
/// Inputs:
/// - `action` — the [`expiry_action`] verdict for this self-node.
/// - `already_reauthing` — whether the published state is currently
///   [`DeviceState::Reauthenticating`](crate::DeviceState::Reauthenticating) (the prior reauth has
///   not yet recovered the node).
/// - `attempts` — the current consecutive-failed-reauth counter.
///
/// Rules:
/// - `Running` → reset the counter to `0` (the node recovered) and report `Running`.
/// - `Reauthenticate` while NOT already reauthing → ENTER reauthenticating: counter `= 1`, fire the
///   one-shot reauth.
/// - `Reauthenticate` while ALREADY reauthing → the prior reauth did not recover; COUNT it (counter
///   `+= 1`) and do NOT re-fire (a second rotation would lose the original `OldNodeKey` anchor). At
///   [`MAX_REAUTH_ATTEMPTS`] the breaker trips to terminal `Expired`; below the cap, stay reauthing.
/// - `Expired` → terminal; the counter is irrelevant (this path never armed the reauth machine), so
///   it is reset to `0`.
pub(crate) fn reauth_circuit_step(
    action: ExpiryAction,
    already_reauthing: bool,
    attempts: u32,
) -> ReauthStep {
    match action {
        ExpiryAction::Running => ReauthStep {
            next: ReauthState::Running,
            attempts: 0,
            fire_reauth: false,
        },
        ExpiryAction::Reauthenticate if !already_reauthing => ReauthStep {
            next: ReauthState::Reauthenticating,
            attempts: 1,
            fire_reauth: true,
        },
        ExpiryAction::Reauthenticate => {
            let attempts = attempts.saturating_add(1);
            if attempts >= MAX_REAUTH_ATTEMPTS {
                ReauthStep {
                    next: ReauthState::Expired,
                    attempts,
                    fire_reauth: false,
                }
            } else {
                ReauthStep {
                    next: ReauthState::Reauthenticating,
                    attempts,
                    fire_reauth: false,
                }
            }
        }
        ExpiryAction::Expired => ReauthStep {
            next: ReauthState::Expired,
            attempts: 0,
            fire_reauth: false,
        },
    }
}

/// The classification of SELF against the active network lock — the observability analog of Go
/// `tkaFilterNetmapLocked`'s self check (which raises a `LockedOut` health warning).
#[derive(Debug, Clone, PartialEq, Eq)]
enum SelfLockVerdict {
    /// Self carries no key-signature at all (empty). The common "not signed yet" case: the node
    /// simply has not been signed for this lock — not locked out, just unsigned.
    Unsigned,
    /// Self's key-signature is authorized by the active lock; nothing to warn about.
    Authorized,
    /// Self has a key-signature but the lock does NOT authorize it (the message is the verify
    /// error). The operator-facing `LockedOut` condition: locked peers will reject this node.
    LockedOut(String),
}

/// Classify a node key + its key-signature against `authority` (pure: verify-and-classify, no
/// logging, no I/O). Takes only the two fields it needs — not the whole `Node` — so the decision is
/// unit-testable without constructing a full `Node` or standing up the actor.
fn self_lock_verdict(
    node_key: &ts_keys::NodePublicKey,
    key_signature: &[u8],
    authority: &ts_tka::Authority,
) -> SelfLockVerdict {
    // Mirror the peer path (`peer_tracker` `tka_snapshot_admits`): treat an empty signature as
    // "unsigned" rather than the `LockedOut` bucket Go's `NodeKeyAuthorized` would put a nil sig in
    // (it errors at decode). This is a deliberate, narrow divergence from a literal Go port: it
    // avoids `warn`-spam on a lock that simply has not signed this node yet, and keeps self and peer
    // classification consistent.
    if key_signature.is_empty() {
        return SelfLockVerdict::Unsigned;
    }
    match authority.node_key_authorized(&node_key.to_bytes(), key_signature) {
        Ok(()) => SelfLockVerdict::Authorized,
        Err(e) => SelfLockVerdict::LockedOut(e.to_string()),
    }
}

/// Emit the self-locked-out observability signal (Go `tkaFilterNetmapLocked`'s self check → a
/// `LockedOut` health warning): classify SELF against the freshly-synced `authority` and log.
///
/// This is **observability, not enforcement** — self never enters the peer db, so the lock can never
/// filter our own node out of the netmap. But a self the lock does not authorize means this node's
/// key-signature is absent or invalid for the active lock, so it cannot prove itself to locked peers
/// (they will drop it); surfacing that lets an operator notice and re-sign. A never-signed node
/// (empty signature) logs at `info`, distinct from a present-but-invalid signature (`warn`), so the
/// common unsigned case does not spam a warning. This fork has no health subsystem, so the operator
/// signal is a `tracing` event (its observability channel).
fn log_self_lockout(self_node: &Node, authority: &ts_tka::Authority) {
    match self_lock_verdict(&self_node.node_key, &self_node.key_signature, authority) {
        SelfLockVerdict::Unsigned => tracing::info!(
            "TKA: this node has no key-signature for the active lock; it cannot prove itself to \
             locked peers until control signs it (not locked out, just unsigned)"
        ),
        SelfLockVerdict::Authorized => {
            tracing::debug!("TKA: self node-key is authorized by the active lock")
        }
        SelfLockVerdict::LockedOut(error) => tracing::warn!(
            %error,
            "TKA self locked out: this node's key-signature is not authorized by the active \
             network lock; locked peers will reject it until control re-signs this node \
             (Go LockedOut)"
        ),
    }
}

// The `#[kameo::messages]` macro generates message structs whose fields mirror the method params;
// those generated fields carry no doc and can't take attributes, so wrap in a module where
// missing-docs is allowed (same pattern as PeerTracker's `msg_impl`). The generated message structs
// are re-exported so callers keep referencing them at `control_runner::<Name>`.
pub use msg_impl::*;

#[allow(missing_docs)]
mod msg_impl {
    use kameo::{message::Context, reply::DelegatedReply};

    use super::*;

    #[kameo::messages]
    impl ControlRunner {
        /// Fetch the IPv4 address for this tailscale device.
        #[message(ctx)]
        pub fn ipv4(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Option<Ipv4Addr>>>,
        ) -> DelegatedReply<Option<Ipv4Addr>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let fut = self.with_self_node(|node| node.tailnet_address.ipv4.addr());

                tokio::spawn(async move {
                    let ip = fut.await;
                    replier.send(ip);
                });
            }

            deleg
        }

        /// Fetch the IPv6 address for this tailscale device.
        #[message(ctx)]
        pub fn ipv6(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Option<Ipv6Addr>>>,
        ) -> DelegatedReply<Option<Ipv6Addr>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let fut = self.with_self_node(|node| node.tailnet_address.ipv6.addr());

                tokio::spawn(async move {
                    let ip = fut.await;
                    replier.send(ip);
                });
            }

            deleg
        }

        /// Fetch the self node for this tailscale device.
        #[message(ctx)]
        pub fn self_node(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Option<Node>>>,
        ) -> DelegatedReply<Option<Node>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let node = self.with_self_node(|node| node.clone());

                tokio::spawn(async move {
                    let node = node.await;
                    replier.send(node)
                });
            }

            deleg
        }

        /// Fetch the current Tailscale SSH policy, if control has pushed one.
        ///
        /// Returns `None` when control has not sent an SSH policy (the SSH server treats this as
        /// deny-all — fail-closed). Unlike `self_node` this does not block waiting
        /// for a value: an absent policy is a legitimate, immediate answer.
        #[message]
        pub fn current_ssh_policy(&self) -> Option<SshPolicy> {
            self.ssh_policy.borrow().clone()
        }

        /// Fetch the current Tailnet Lock status, if control has pushed one.
        ///
        /// Returns `None` when control has sent no `TKAInfo` (tailnet lock not in use / no change seen).
        #[message]
        pub fn current_tka_status(&self) -> Option<TkaStatus> {
            self.tka.borrow().clone()
        }

        /// Read up to `limit` entries of the Tailnet-Lock update-chain log, **head-first** (newest →
        /// oldest), from the locally-synced AUM chain (Go `NetworkLockLog`).
        ///
        /// A **pure local read** — no crypto, no mutation, no control round-trip: it walks the
        /// already-verified `SyncedTka` store this actor owns. Returns an empty `Vec` when no lock is
        /// synced (Go's `b.tka == nil`). Synchronous (no spawn), like `current_tka_status` — the
        /// chain is in memory.
        #[message]
        pub fn tka_log(&self, limit: usize) -> Vec<crate::tka_sync::TkaLogEntry> {
            let Some(synced) = self.tka_synced.as_ref() else {
                return Vec::new();
            };
            crate::tka_sync::tka_log_entries(&synced.store, synced.oldest, limit)
        }

        /// Sign `node_key` directly with this node's network-lock key and submit the signature to
        /// control (Go `tka.sign` for the Direct case → `tkaSubmitSignature`).
        ///
        /// Builds a `Direct` [`NodeKeySignature`](ts_tka::NodeKeySignature) via
        /// [`sign_direct`](ts_tka::NodeKeySignature::sign_direct) over this node's inner ed25519
        /// network-lock signing key, serializes it (raw CBOR), and POSTs it to `/machine/tka/sign`.
        /// Mirrors `set_dns`/`get_certificate`: clones the control config + node keys into a spawned
        /// task (delegated reply, so the round-trip doesn't block the mailbox) over a fresh Noise
        /// channel.
        ///
        /// **Posture: this only *submits* a signature to control — it does NOT mutate the local
        /// [`Authority`](ts_tka::Authority).** The local trusted-key state advances solely through the
        /// existing verified-sync path (`sync_tka` → `VerifiedAumChain::verify`); a `tka_sign` success
        /// is acknowledged to the caller, and the resulting AUM is picked up on the next netmap-driven
        /// sync. Verify-and-log is unchanged.
        #[message(ctx)]
        pub fn tka_sign(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<(), TkaSyncError>>>,
            node_key: [u8; 32],
        ) -> DelegatedReply<Result<(), TkaSyncError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    // Sign the node key with our network-lock key, then submit the raw-CBOR NKS.
                    let nks = ts_tka::NodeKeySignature::sign_direct(
                        &node_key,
                        &keys.network_lock_keys.private.signing_key(),
                    );
                    let req = ts_control::TkaSubmitSignatureRequest {
                        // node_key + version are stamped by the RPC client from `keys`.
                        version: Default::default(),
                        node_key: keys.node_keys.public,
                        signature: nks.serialize(),
                    };
                    let result = tka_submit_signature(
                        &config.server_url,
                        &keys,
                        req,
                        config.allow_http_key_fetch,
                    )
                    .await
                    .map(|_response| ());
                    replier.send(result);
                });
            }

            deleg
        }

        /// Disable Tailnet Lock by presenting the disablement secret to control (Go
        /// `tka.disable` → `/machine/tka/disable`).
        ///
        /// Targets the **current** authority head (read from the cached [`TkaStatus`]); the caller
        /// supplies the `disablement_secret` out of band (it is the operator-held capability that
        /// authorizes turning the lock off). Mirrors `tka_sign`: clones config + keys into a spawned
        /// task (delegated reply). Returns [`TkaSyncError::Unsupported`] when there is no known TKA
        /// head (lock not in use / control hasn't pushed a status), since there is nothing to disable.
        ///
        /// **Submit-only, like `tka_sign`:** this POSTs the disablement to control and does NOT mutate
        /// the local [`Authority`](ts_tka::Authority). Control acts on the disablement; this node
        /// observes the result through the existing verified-sync path. Verify-and-log unchanged.
        #[message(ctx)]
        pub fn tka_disable(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<(), TkaSyncError>>>,
            disablement_secret: Vec<u8>,
        ) -> DelegatedReply<Result<(), TkaSyncError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                // Read the current head from the cached status BEFORE the spawn (can't borrow &self
                // across the await). No head ⇒ no lock to disable ⇒ Unsupported.
                let head = self.tka.borrow().as_ref().map(|s| s.head.clone());
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = match head {
                        Some(head) => {
                            let req = ts_control::TkaDisableRequest {
                                // node_key + version are stamped by the RPC client from `keys`.
                                version: Default::default(),
                                node_key: keys.node_keys.public,
                                head,
                                disablement_secret,
                            };
                            tka_disable(&config.server_url, &keys, req, config.allow_http_key_fetch)
                                .await
                                .map(|_response| ())
                        }
                        None => Err(TkaSyncError::Unsupported),
                    };
                    replier.send(result);
                });
            }

            deleg
        }

        /// Initialize Tailnet Lock with this node as the sole initial trusted key, gated by
        /// `disablement_secret` (Go `LocalClient.NetworkLockInit` — the "lock yourself in" case).
        ///
        /// Builds + signs a genesis Checkpoint AUM whose only trusted key is this node's network-lock
        /// public key (votes 1) and whose single DisablementValue is `disablement_value(secret)`, then
        /// drives the two-phase init: `tka/init/begin` (submit the genesis) → if control needs no
        /// further node signatures (`NeedSignatures` empty, the case when this node is the only key) →
        /// `tka/init/finish` carrying the raw `disablement_secret` as `SupportDisablement`. Mirrors
        /// `tka_sign`/`tka_disable`: cloned config + keys into a spawned task (delegated reply).
        ///
        /// If control returns a non-empty `NeedSignatures` (other nodes must be re-signed under the new
        /// lock — a multi-node tailnet), this returns [`TkaSyncError::Unsupported`]: re-signing each
        /// listed node (incl. the Rotation-key case) is a larger flow deferred to a fuller
        /// `tka_init(keys, secrets)` — the single-node lock-init is the shipped subset.
        ///
        /// **Submit-only**, like `tka_sign`/`tka_disable`: this creates the lock at control and does
        /// NOT seed the local [`Authority`](ts_tka::Authority) — the node picks up the new lock through
        /// the existing verified netmap-sync (control pushes a `TKAInfo`, `maybe_sync_tka` bootstraps
        /// the genesis through `VerifiedAumChain::verify`). Verify-and-log posture unchanged.
        #[message(ctx)]
        pub fn tka_init(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<(), TkaSyncError>>>,
            disablement_secret: Vec<u8>,
        ) -> DelegatedReply<Result<(), TkaSyncError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = tka_init_run(&config, &keys, disablement_secret).await;
                    replier.send(result);
                });
            }

            deleg
        }

        /// The cert-eligible DNS names from control's netmap DNS config (Go `nm.DNS.CertDomains`).
        ///
        /// Returns an empty `Vec` when control has sent no DNS config, or one carrying no cert
        /// domains (an empty list is a legitimate, immediate answer — like `current_ssh_policy`, this
        /// does not block waiting for a value).
        #[message]
        pub fn cert_domains(&self) -> Vec<String> {
            self.cert_domains.borrow().clone()
        }

        /// The full DNS config from control's netmap (Go `netmap.NetworkMap.DNS`), or `None` when
        /// control has sent no DNS config yet. An immediate answer (does not block); the facade
        /// surfaces this for `Device::dns_config` (the daemon's `tnet dns status`).
        #[message]
        pub fn dns_config(&self) -> Option<ts_control::DnsConfig> {
            self.dns_config.borrow().clone()
        }

        /// The interactive-login / consent URL control last asked this node to open
        /// (`MapResponse.PopBrowserURL`), or `None` when control has sent none. An immediate answer
        /// (does not block); the facade surfaces this for `Device::pop_browser_url`.
        #[message]
        pub fn pop_browser_url(&self) -> Option<url::Url> {
            self.pop_browser_url.borrow().clone()
        }

        /// Subscribe to the interactive-login / consent URL cell (`MapResponse.PopBrowserURL`).
        ///
        /// Returns a [`watch::Receiver`] whose value is the latest running-node consent URL, used by
        /// [`Runtime::watch_ipn_bus`](crate::Runtime::watch_ipn_bus) to surface `browse_to_url`
        /// events mid-session. The cell is sticky (updated only on a new non-empty URL, never reset
        /// to `None` by an empty update — see the field docs), so a subscriber is not thrashed and a
        /// late subscriber sees the current URL. The initial value is `None` until control sends one.
        #[message(derive(Clone))]
        pub fn watch_browser_url(&self) -> watch::Receiver<Option<url::Url>> {
            self.pop_browser_url.subscribe()
        }

        /// The latest network-conditions report (preferred DERP region + per-region latencies). An
        /// immediate answer (does not block); empty before the first DERP-latency measurement. The
        /// facade surfaces this for `Device::netcheck` (the daemon's `tnet netcheck`).
        #[message]
        pub fn netcheck(&self) -> crate::status::NetcheckReport {
            self.netcheck.borrow().clone()
        }

        /// The lowest latency seen for each DERP region across the retained measurement history —
        /// the Rust analog of Go `netcheck.Client.RecentRegionLatency()`, and what
        /// [`Runtime::suggest_exit_node`](crate::Runtime::suggest_exit_node) ranks exit-node
        /// candidates on.
        ///
        /// Ranking on this rather than on the newest report is the whole point: every measurement
        /// this fork makes is *partial* (`ts_netcheck::Config::complete_threshold` ends it once
        /// three regions have answered), so the newest report names a handful of regions and a
        /// candidate homed anywhere else would be ranked against nothing — collapsing the
        /// suggestion to a uniform random pick. The union over the retained window gives every
        /// region a recent latency to be ranked on.
        ///
        /// The window (`DERP_HISTORY_MAX_AGE`) is re-applied *here*, at read time, not just when
        /// a measurement lands: the history is only pruned on arrival, so a node that has gone
        /// quiet must not serve latencies that have since aged out. An immediate answer (does not
        /// block); empty before the first DERP-latency measurement.
        #[message]
        pub fn recent_region_latency(&self) -> HashMap<ts_derp::RegionId, Duration> {
            best_recent(
                &self.derp_report_history,
                Instant::now(),
                DERP_HISTORY_MAX_AGE,
            )
        }

        /// Request an OIDC ID token from control scoped to `audience` (workload-identity federation).
        ///
        /// Opens a fresh Noise channel and POSTs `/machine/id-token`; returns the signed JWT or an
        /// [`IdTokenError`]. Runs on a spawned task (delegated reply) so the actor mailbox isn't blocked
        /// for the round-trip.
        #[message(ctx)]
        pub fn fetch_id_token(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<String, IdTokenError>>>,
            audience: String,
        ) -> DelegatedReply<Result<String, IdTokenError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = ts_control::fetch_id_token(&config, &keys, &audience).await;
                    replier.send(result);
                });
            }

            deleg
        }

        /// Log this node out of the tailnet: deregister it by expiring its current node key.
        ///
        /// Mirrors `fetch_id_token`: clones the control config + node keys
        /// into a spawned task (delegated reply, so the round-trip doesn't block the mailbox) and
        /// re-POSTs `/machine/register` with a past expiry over a fresh Noise channel. This is a
        /// control-plane state change only — it does NOT stop this actor or tear down the datapath
        /// (the caller follows up with the normal runtime shutdown), and it does not touch the
        /// on-disk node key, so re-registering with the same key is the re-login path.
        #[message(ctx)]
        pub fn logout(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<(), LogoutError>>>,
        ) -> DelegatedReply<Result<(), LogoutError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = ts_control::logout(&config, &keys).await;
                    replier.send(result);
                });
            }

            deleg
        }

        /// Publish a DNS record for this node via control's `/machine/set-dns` (Go
        /// `LocalClient.SetDNS`).
        ///
        /// Mirrors `fetch_id_token`: clones the control config + node keys
        /// into a spawned task (delegated reply, so the round-trip doesn't block the mailbox) and
        /// POSTs the record over a fresh Noise channel. Go's `SetDNS` is `TXT`-only (its sole use is
        /// the ACME DNS-01 `_acme-challenge` record); the record type is fixed to `"TXT"` here to
        /// match, so the surfaced API takes only `name` + `value`.
        #[message(ctx)]
        pub fn set_dns(
            &self,
            ctx: &mut Context<Self, DelegatedReply<Result<(), SetDnsError>>>,
            name: String,
            value: String,
        ) -> DelegatedReply<Result<(), SetDnsError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = ts_control::set_dns(&config, &keys, &name, "TXT", &value).await;
                    replier.send(result);
                });
            }

            deleg
        }
    }

    /// The reply type of the [`get_cert_pair`](ControlRunner::get_cert_pair) message: the issued
    /// `(cert_chain_pem, key_pem)` PEM pair (the `tnet cert` surface) or a [`ts_control::CertError`].
    /// Aliased so the message's `Context` type stays under clippy's `type_complexity` bar (the
    /// nested `Result<(String, String), _>` trips it inline).
    #[cfg(feature = "acme")]
    pub type CertPairReply = Result<(String, String), ts_control::CertError>;

    // The `acme`-gated cert-issuance message lives in its own `#[kameo::messages]` impl block so the
    // proc-macro never sees it in a non-`acme` build (a `#[cfg]` *inside* a single messages-impl
    // block is not honored by the macro's generated dispatch — it would emit a `GetCertificate`
    // handler calling a `get_certificate` method that the same `#[cfg]` strips). A separate gated
    // block keeps the default build clean.
    #[cfg(feature = "acme")]
    #[kameo::messages]
    impl ControlRunner {
        /// Issue a real Let's Encrypt certificate for this node's MagicDNS `name` via the
        /// client-side ACME DNS-01 engine (`acme` feature).
        ///
        /// Mirrors `fetch_id_token`: clones the control config + node keys
        /// into a spawned task (delegated reply, so the round-trip doesn't block the mailbox), loads
        /// or generates the ACME account key, and runs issuance against Let's Encrypt production,
        /// publishing the DNS-01 challenge TXT through the node's `POST /machine/set-dns` RPC.
        ///
        /// The account key is loaded from [`ts_keys::NodeState::acme_account_key`] (PKCS#8 DER) when
        /// present, so the same ACME account persists across renewals; otherwise an ephemeral key is
        /// generated for this call only (a fresh ACME account each issuance — acceptable for v1; LE
        /// allows it). Persisting a generated key back into the key file is the embedder's job (no
        /// write-back path here). SaaS-only: against a self-hosted control plane the set-dns
        /// publish 501s.
        #[message(ctx)]
        pub fn get_certificate(
            &self,
            ctx: &mut Context<
                Self,
                DelegatedReply<Result<ts_control::tls::CertifiedKey, ts_control::CertError>>,
            >,
            name: String,
        ) -> DelegatedReply<Result<ts_control::tls::CertifiedKey, ts_control::CertError>> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = issue_certificate(&config, &keys, &name).await;
                    replier.send(result);
                });
            }

            deleg
        }

        /// Issue a real Let's Encrypt certificate for this node's MagicDNS `name` and return the
        /// **PEM pair** — `(cert_chain_pem, key_pem)` — for writing the on-disk `.crt` + `.key`
        /// (the daemon's `tnet cert`, Go's `LocalClient.CertPair`). `acme` feature.
        ///
        /// Identical issuance to [`get_certificate`](Self::get_certificate) (same client-side ACME
        /// DNS-01 flow, same set-dns publish, same account-key handling), only the *shape* of the
        /// result differs: this surfaces the raw chain + leaf-key PEMs instead of the opaque
        /// [`CertifiedKey`](ts_control::tls::CertifiedKey). The leaf **private key** PEM is the
        /// second tuple element and is NEVER logged — the spawned task sends it straight back to the
        /// replier. SaaS-only: against a self-hosted control plane the set-dns publish 501s.
        #[message(ctx)]
        pub fn get_cert_pair(
            &self,
            ctx: &mut Context<Self, DelegatedReply<CertPairReply>>,
            name: String,
        ) -> DelegatedReply<CertPairReply> {
            let (deleg, replier) = ctx.reply_sender();

            if let Some(replier) = replier {
                let config = self.params.config.clone();
                let keys = self.params.env.keys.clone();
                tokio::spawn(async move {
                    let result = issue_cert_pair(&config, &keys, &name).await;
                    replier.send(result);
                });
            }

            deleg
        }
    }
}

/// The `tka_init` body (the genesis-build + two-phase init/begin→init/finish choreography),
/// factored out of the actor handler so it runs in the spawned task. See [`ControlRunner::tka_init`].
///
/// "Lock yourself in": the genesis trusts only this node's network-lock key (votes 1) and stores one
/// DisablementValue = `disablement_value(secret)`. On a non-empty `NeedSignatures` (multi-node
/// tailnet needing re-signs) it returns [`TkaSyncError::Unsupported`] — the single-node subset.
async fn tka_init_run(
    config: &ts_control::Config,
    keys: &ts_keys::NodeState,
    disablement_secret: Vec<u8>,
) -> Result<(), TkaSyncError> {
    // Build the genesis: this node's NL public key as the sole trusted key, one disablement value.
    let nl_public = keys.network_lock_keys.public.to_bytes().to_vec();
    let genesis_key = ts_tka::AumKey {
        kind: ts_tka::KeyKind::Ed25519,
        votes: 1,
        public: nl_public,
        meta: Vec::new(),
    };
    let dvalue = ts_tka::disablement_value(&disablement_secret).to_vec();
    let mut genesis = ts_tka::Aum::new_genesis_checkpoint(vec![genesis_key], vec![dvalue])
        // A malformed genesis is a local construction bug, not a transient RPC failure — surface it as a
        // coarse internal error rather than NetworkError (which would invite a pointless retry).
        .map_err(|_| TkaSyncError::Internal(ts_control::TkaSyncInternalErrorKind::SerDe))?;
    genesis.sign(&keys.network_lock_keys.private.signing_key());

    // Phase 1: submit the genesis. node_key + version are stamped by the RPC client from `keys`.
    let begin_req = ts_control::TkaInitBeginRequest {
        version: Default::default(),
        node_key: keys.node_keys.public,
        genesis_aum: genesis.serialize(),
    };
    let begin_resp = tka_init_begin(
        &config.server_url,
        keys,
        begin_req,
        config.allow_http_key_fetch,
    )
    .await?;

    // Single-node case only: control must need no further node signatures. A non-empty
    // NeedSignatures means other nodes must be re-signed under the new lock — deferred.
    if !begin_resp.need_signatures.is_empty() {
        tracing::warn!(
            need = begin_resp.need_signatures.len(),
            "tka_init: control requires re-signing other nodes; the multi-node init is not yet \
             implemented (single-node lock-init only)"
        );
        return Err(TkaSyncError::Unsupported);
    }

    // Phase 2: finish, carrying the raw disablement secret as SupportDisablement (Go sends the raw
    // secret here; only the genesis stores its Argon2i hash).
    let finish_req = ts_control::TkaInitFinishRequest {
        version: Default::default(),
        node_key: keys.node_keys.public,
        signatures: std::collections::BTreeMap::new(),
        support_disablement: disablement_secret,
    };
    tka_init_finish(
        &config.server_url,
        keys,
        finish_req,
        config.allow_http_key_fetch,
    )
    .await
    .map(|_response| ())
}

/// Load or generate the ACME account key, then issue a cert for `name` via set-dns DNS-01,
/// returning just the ready-to-serve [`CertifiedKey`](ts_control::tls::CertifiedKey) (the
/// `get_certificate` / `ListenTLS` path).
///
/// Thin wrapper over [`issue_cert_pair`] that drops the PEMs — one issuance, this caller just
/// doesn't need the on-disk pair. See [`issue_cert_pair`] for the account-key handling.
#[cfg(feature = "acme")]
async fn issue_certificate(
    config: &ts_control::Config,
    keys: &ts_keys::NodeState,
    name: &str,
) -> Result<ts_control::tls::CertifiedKey, ts_control::CertError> {
    issue_cert_pair_inner(config, keys, name)
        .await
        .map(|issued| issued.certified)
}

/// Load or generate the ACME account key, then issue a cert for `name` via set-dns DNS-01,
/// returning the **PEM pair** `(cert_chain_pem, key_pem)` for the daemon's on-disk `.crt`/`.key`
/// (`tnet cert`, Go `LocalClient.CertPair`).
///
/// Same single issuance as [`issue_certificate`]; only the result shape differs. The leaf
/// **private key** PEM is the second element and is NEVER logged here.
#[cfg(feature = "acme")]
async fn issue_cert_pair(
    config: &ts_control::Config,
    keys: &ts_keys::NodeState,
    name: &str,
) -> Result<(String, String), ts_control::CertError> {
    issue_cert_pair_inner(config, keys, name)
        .await
        .map(|issued| (issued.cert_chain_pem, issued.key_pem))
}

/// Shared issuance core for [`issue_certificate`] and [`issue_cert_pair`]: load (or generate) the
/// ACME account key, target Let's Encrypt production, and run one DNS-01 issuance, returning the
/// full [`IssuedCert`](ts_control::acme::IssuedCert) so each caller projects out what it needs (one
/// ACME order, two consumers).
///
/// Reuses the persisted [`ts_keys::NodeState::acme_account_key`] (PKCS#8 DER) when present so the
/// same Let's Encrypt account survives renewals; otherwise generates an ephemeral per-call key
/// (logged at debug — a new ACME account each issuance, with no write-back). Always targets Let's
/// Encrypt production ([`ts_control::acme::LETS_ENCRYPT_PRODUCTION_DIRECTORY`]). Never logs the leaf
/// private key.
#[cfg(feature = "acme")]
async fn issue_cert_pair_inner(
    config: &ts_control::Config,
    keys: &ts_keys::NodeState,
    name: &str,
) -> Result<ts_control::acme::IssuedCert, ts_control::CertError> {
    let account_key = match keys.acme_account_key.as_deref() {
        Some(der) => ts_control::acme::AcmeAccountKey::from_pkcs8(der)?,
        None => {
            tracing::debug!(
                "no persisted ACME account key in key state; generating an ephemeral per-call key \
                 (a new ACME account this issuance — not persisted back)"
            );
            ts_control::acme::AcmeAccountKey::generate()?.0
        }
    };
    let directory = ts_control::acme::LETS_ENCRYPT_PRODUCTION_DIRECTORY
        .parse()
        .map_err(|e| {
            ts_control::CertError::Acme(format!("parsing Let's Encrypt directory URL: {e}"))
        })?;
    ts_control::issue_cert_pair_via_setdns(config, keys, name, &account_key, &directory).await
}

/// Publish the cached netmap, if this node has one, onto the netmap bus.
///
/// The cold-start half of the netmap cache (Go `nodecap.CacheNetworkMaps`). Reads
/// [`Config::netmap_cache_dir`](ts_control::Config::netmap_cache_dir) — `None` means the embedder
/// configured no storage, so there is nothing to replay — and decodes whatever is there with the
/// same decoder the live map poll uses.
///
/// The read is deliberately **not** gated on the node attributes. Nothing is ever written without
/// the grant, so the presence of a cache is itself the record that control asked for one (Go makes
/// the same argument: at this point in start-up the client has not spoken to control yet, so the
/// grant is not knowable). If the grant has since been withdrawn, the first netmap of this session
/// says so and the cache is discarded then.
///
/// **Peers cached under Tailnet Lock are filtered, not withheld.** Go replays its cached map through
/// `tkaFilterNetmapLocked`, which it can do at cold start because its TKA authority is persisted on
/// disk; the peers that hold a valid signature survive and are dialed. This port persists the same
/// authority next to the cached netmap (see [`load_cached_netmap`]) and runs the same filter over the
/// cached peers. With no persisted authority to run it with — a cache written before this node ever
/// completed a TKA sync — the peers are withheld and control's first netmap brings them back moments
/// later, behind a synced authority.
///
/// The bus has no replay, so this reaches only subscribers already registered. Every netmap
/// subscriber is spawned by `Runtime::spawn` before the control runner and registers from its own
/// `on_start` with no I/O in the way, while this path awaits a file read first — so in practice the
/// subscribers are there. A subscriber that is not simply misses the head start and is brought
/// current by control's first netmap, which is exactly the behaviour of a node with no cache.
async fn replay_cached_netmap(params: &Params) {
    let Some(dir) = params.config.netmap_cache_dir.as_ref() else {
        return;
    };

    let Some(update) = load_cached_netmap(&ts_control::NetmapCache::new(dir)).await else {
        return;
    };

    let peers = match update.peer_update.as_ref() {
        Some(ts_control::PeerUpdate::Full(peers)) => peers.len(),
        _ => 0,
    };
    tracing::info!(peers, "replaying cached netmap on cold start");

    if let Err(e) = params.env.publish(Arc::new(update)).await {
        tracing::warn!(error = %e, "publishing the cached netmap");
    }
}

/// Read the cached netmap, vouching for its peers with the Tailnet-Lock authority persisted beside
/// it — the whole of the cold-start decision, with no actor state in it so it can be tested directly.
///
/// A netmap cached while the lock was **off** replays whole; there is nothing to enforce (Go's
/// `tkaFilterNetmapLocked` returns early on `b.tka == nil`). A netmap cached while the lock was
/// **on** replays exactly the peers
/// [`PeerTracker::tka_keep_verdicts`](crate::peer_tracker::PeerTracker::tka_keep_verdicts) admits —
/// the same pass the live netmap path runs, so a peer that is dialed from the cache is one the
/// authority authorized, and an unsigned peer, a peer whose signature fails, or a peer a newer
/// rotation obsoletes is dropped.
///
/// Three things make the persisted authority safe to enforce with:
///
/// * it is re-verified from genesis on load ([`crate::tka_sync::authority_from_encoded_chain`]), so a
///   blob that is not a signed chain yields no authority;
/// * it is read back out of the same directory the netmap is, vetted as private to this user, so a
///   local attacker cannot choose which chain we start from any more than they can choose the netmap;
/// * its head must equal the head the **cached frame** recorded (`MapResponse.TKAInfo.Head`) — see
///   [`vouch_cached_peers`]. Go does not need that check because it has no second file to reconcile,
///   only the chonk.
pub(crate) async fn load_cached_netmap(cache: &ts_control::NetmapCache) -> Option<StateUpdate> {
    let authority = match cache.load_tka_chain().await {
        None => None,
        Some(blob) => match crate::tka_sync::authority_from_encoded_chain(&blob) {
            Ok(authority) => Some(authority),
            Err(e) => {
                tracing::warn!(
                    error = %e,
                    "persisted tailnet-lock chain did not verify; replaying the cached netmap \
                     without its peers"
                );
                None
            }
        },
    };

    cache
        .load_state_update_vouched(|tka, peers| vouch_cached_peers(authority.as_ref(), tka, peers))
        .await
}

/// The peers of a netmap cached under an **active** Tailnet Lock that may be replayed on this cold
/// start: those the persisted `authority` authorizes, by exactly the pass a live netmap's peers go
/// through (Go's `tkaFilterNetmapLocked`, via
/// [`PeerTracker::tka_keep_verdicts`](crate::peer_tracker::PeerTracker::tka_keep_verdicts) — the
/// per-peer signature verdict plus the cross-peer rotation filter).
///
/// Nothing is replayed unless the authority can be held to the netmap:
///
/// * **no persisted authority** (this node has never completed a TKA sync, or the chain did not
///   verify) ⇒ no peers. There is nothing to check a `key_signature` against, and admitting them
///   unchecked would dial a peer the lock may have revoked while this node was off.
/// * **the chain is not at the head the cached frame recorded** (`MapResponse.TKAInfo.Head`) ⇒ no
///   peers. The netmap and the chain are written on separate events, so they can disagree — a crash
///   between them, or a lock disabled and re-enabled under a fresh genesis — and a chain that
///   describes a different lock says nothing about these peers.
///
/// Separated from the I/O in [`load_cached_netmap`] so the decision itself is directly testable with
/// a real chain-derived [`Authority`](ts_tka::Authority) and real signatures.
pub(crate) fn vouch_cached_peers(
    authority: Option<&ts_tka::Authority>,
    tka: &TkaStatus,
    peers: Vec<Node>,
) -> Vec<Node> {
    let Some(authority) = authority else {
        tracing::info!(
            "no persisted tailnet-lock authority to verify the cached peers against; withholding \
             them until control's first netmap"
        );
        return Vec::new();
    };

    if !ts_tka::AumHash::from_base32(&tka.head).is_some_and(|head| authority.head_matches(&head)) {
        tracing::warn!(
            head = %tka.head,
            "persisted tailnet-lock chain is not at the head the cached netmap recorded; \
             withholding its peers"
        );
        return Vec::new();
    }

    let keep = {
        let refs: Vec<&Node> = peers.iter().collect();
        crate::peer_tracker::PeerTracker::tka_keep_verdicts(Some(authority), &refs)
    };
    peers
        .into_iter()
        .zip(keep)
        .filter_map(|(peer, keep)| keep.then_some(peer))
        .collect()
}

/// The Tailnet-Lock chain persisted beside the cached netmap ([`load_cached_netmap`]), and the only
/// path that writes or removes it.
///
/// It is one owner because the two operations have to be ordered **against each other**. They used
/// not to be: the write was awaited on the actor thread while the removal was `tokio::spawn`ed, so a
/// removal scheduled by a lock *disable* could still be sitting in the runtime's queue when a sync
/// that followed it persisted a fresh chain, and then delete that fresh chain. Nothing tells the node
/// afterwards — the loss shows up one restart later, as a cold start that withholds peers it was
/// entitled to dial. Every operation here is awaited by its caller instead, and both callers
/// ([`ControlRunner::persist_tka_chain`] and the two lock-off paths) run on the actor thread, so the
/// actor's mailbox is the order.
///
/// This owns no policy about *which* chain may be replayed: the cold-start replay judges whatever it
/// finds on disk against the head the cached netmap frame itself recorded, so nothing here can admit
/// a peer, and a failed removal is a chain the replay refuses rather than one it trusts.
struct PersistedTkaChain {
    /// The cache directory the chain lives in, or `None` when the embedder configured no
    /// [`netmap_cache_dir`](ts_control::Config::netmap_cache_dir) — then no netmap is cached either,
    /// so there is nothing to vouch for and nothing to remove.
    cache: Option<ts_control::NetmapCache>,
    /// Whether a chain may be on disk right now.
    ///
    /// **Starts `true`.** A fresh process has synced nothing, but an earlier process of this node can
    /// have left a chain in the cache directory, and that chain is precisely what the next cold start
    /// would load. So "this process never synced a lock" is not evidence that there is nothing to
    /// remove, and the lock-off paths must not treat it as such.
    ///
    /// Its only job is to keep the repeated removal free: a tailnet with no lock reports the lock off
    /// on *every* netmap, and one `remove_file` per netmap tick is a syscall for nothing. It is never
    /// trusted in the other direction — a stale `false` cannot admit a peer, because admitting is the
    /// replay's decision and the replay reads the disk.
    maybe_on_disk: bool,
}

impl PersistedTkaChain {
    /// The chain owner for a node configured with `dir` as its netmap cache directory.
    fn new(dir: Option<&std::path::PathBuf>) -> Self {
        Self {
            cache: dir.map(ts_control::NetmapCache::new),
            maybe_on_disk: true,
        }
    }

    /// Persist `blob` as the chain the cached netmap's peers were admitted under. A write failure is
    /// logged inside the cache and swallowed; the flag stays set, so the next lock-off still tries to
    /// remove whatever did land.
    async fn store(&mut self, blob: &[u8]) {
        let Some(cache) = self.cache.as_ref() else {
            return;
        };
        cache.store_tka_chain(blob).await;
        self.maybe_on_disk = true;
    }

    /// Remove the persisted chain. Awaited, so it has happened by the time this returns and cannot
    /// overtake a later [`store`](Self::store).
    async fn discard(&mut self) {
        if !self.maybe_on_disk {
            return;
        }
        let Some(cache) = self.cache.as_ref() else {
            // No directory was ever configured, so nothing can be on disk and nothing ever will be.
            // Settle the flag rather than re-deciding this on every netmap.
            self.maybe_on_disk = false;
            return;
        };
        cache.discard_tka_chain().await;
        self.maybe_on_disk = false;
    }
}

impl Message<StreamMessage<Arc<StateUpdate>, (), ()>> for ControlRunner {
    type Reply = ();

    async fn handle(
        &mut self,
        msg: StreamMessage<Arc<StateUpdate>, (), ()>,
        ctx: &mut Context<Self, Self::Reply>,
    ) {
        match msg {
            StreamMessage::Started(_) => {
                tracing::trace!("started listening to state updates");
            }

            StreamMessage::Next(msg) => {
                if let Some(node) = msg.node.as_ref() {
                    // Reflect node-key expiry into the device state. Control delivering a self-node
                    // whose key is in the past means the node must re-authenticate; the arrival of a
                    // fresh (non-expired) self-node confirms we are Running (recovering the state if a
                    // prior update had flipped it to Expired/Reauthenticating). On expiry, decide
                    // between an automatic re-auth (Go `doLogin`: rotate key + re-register with the
                    // stored auth key) and the terminal Expired state via the pure `expiry_action`:
                    //   - auth key retained, reauth enabled, and TKA NOT enforcing → Reauthenticate.
                    //   - otherwise → Expired (no auth key / reauth disabled / TKA-locked).
                    // The TKA gate is a hard safety constraint: rotating on a locked tailnet would
                    // install an unsigned key and lock this node out of locked peers (the TKA re-sign
                    // is a separate follow-up). Recovery from Reauthenticating is automatic — the next
                    // good self-node flips back to Running at this same handler.
                    let now_unix = std::time::SystemTime::now()
                        .duration_since(std::time::UNIX_EPOCH)
                        .map(|d| d.as_secs() as i64)
                        .unwrap_or(0);
                    let action = expiry_action(
                        node.key_expired_at_unix(now_unix),
                        self.params.auth_key.is_some(),
                        self.params.config.reauth_on_expiry,
                        self.tka_authority.borrow().is_some(),
                    );

                    // Bounded reauth sub-state-machine (circuit breaker), evaluated by the pure
                    // `reauth_circuit_step`. The `Reauthenticate` path must settle — a one-shot /
                    // already-consumed auth key cannot re-register, so an unbounded reauth would sit in
                    // `Reauthenticating` forever. The step takes the `expiry_action` verdict, whether we
                    // are ALREADY reauthenticating (i.e. the prior reauth has not recovered), and the
                    // current attempt counter, and returns the target state, the new counter, and
                    // whether to fire the one-shot reauth (fired ONLY on entry, so the node key rotates
                    // at most once per episode — a second rotation would lose the original `OldNodeKey`
                    // anchor). At `MAX_REAUTH_ATTEMPTS` it trips to terminal `Expired`.
                    //
                    // NOTE: we may act (count, and eventually flip to `Expired`) even when the published
                    // state does NOT change — a repeated `Reauthenticating` self-node is exactly the
                    // "prior reauth didn't recover" signal we must tally, so the counting reads the
                    // pre-step state directly here and `send_if_modified`'s `changed` only gates the
                    // log/firing below, never the counting.
                    let already_reauthing = matches!(
                        &*self.params.state_tx.borrow(),
                        crate::DeviceState::Reauthenticating
                    );
                    let step = reauth_circuit_step(action, already_reauthing, self.reauth_attempts);
                    self.reauth_attempts = step.attempts;
                    if step.next == ReauthState::Expired
                        && action == ExpiryAction::Reauthenticate
                        && already_reauthing
                    {
                        tracing::warn!(
                            attempts = step.attempts,
                            "automatic re-auth did not recover the node after {MAX_REAUTH_ATTEMPTS} \
                             attempts (auth key likely one-shot / consumed); falling back to terminal \
                             Expired"
                        );
                    }
                    let next = match step.next {
                        ReauthState::Running => crate::DeviceState::Running,
                        ReauthState::Reauthenticating => crate::DeviceState::Reauthenticating,
                        ReauthState::Expired => crate::DeviceState::Expired,
                    };

                    // `send_if_modified` avoids waking watchers when the state is unchanged (a fresh
                    // self-node arrives on every netmap update). Returns whether the state changed.
                    let changed = self.params.state_tx.send_if_modified(|s| {
                        if *s != next {
                            *s = next.clone();
                            true
                        } else {
                            false
                        }
                    });

                    if changed && step.fire_reauth {
                        tracing::info!(
                            "self node-key expired; starting automatic re-auth (rotate node key + \
                             re-register with stored auth key)"
                        );
                        self.client.reauth().await;
                    }

                    self.self_node.send_replace(Some(node.clone()));
                }

                if let Some(policy) = msg.ssh_policy.as_ref() {
                    self.ssh_policy.send_replace(Some(policy.clone()));
                }

                if let Some(tka) = msg.tka.as_ref() {
                    self.tka.send_replace(Some(tka.clone()));
                    self.maybe_sync_tka(tka, ctx.actor_ref().clone()).await;
                }

                // Track the cert-domain list from the netmap DNS config (Go `nm.DNS.CertDomains`).
                // An update with no DNS config, or one carrying no cert domains, means "none" — Go
                // reads an empty slice off an absent config too, so mirror that as an empty `Vec`.
                let cert_domains = msg
                    .dns_config
                    .as_ref()
                    .map(|d| d.cert_domains.clone())
                    .unwrap_or_default();
                self.cert_domains.send_replace(cert_domains);

                // Track the full DNS config for `Device::dns_config` (the daemon's `tnet dns status`).
                // `None` when control sent no DNS config on this update — distinct from a present but
                // empty config (Go `netmap.NetworkMap.DNS`).
                self.dns_config.send_replace(msg.dns_config.clone());

                // Track the interactive-login URL for `Device::pop_browser_url` /
                // `Runtime::watch_ipn_bus`. See `sticky_update_pop_browser_url` for the Go-faithful
                // sticky semantics (update only on a new non-empty URL; never reset to `None`).
                sticky_update_pop_browser_url(&self.pop_browser_url, msg.pop_browser_url.as_ref());

                if let Err(e) = self.params.env.publish(msg).await {
                    tracing::error!(error = %e, "publishing netmap update");
                }
            }

            StreamMessage::Finished(_) => {
                tracing::error!("state update stream terminated")
            }
        }
    }
}

/// The outcome of a spawned TKA bootstrap+sync task, delivered back to the actor thread so the
/// result can be applied to actor state (which a spawned task cannot touch directly). Sent by
/// [`ControlRunner::maybe_sync_tka`]; handled by applying via
/// [`ControlRunner::apply_tka_synced`](ControlRunner).
#[doc(hidden)]
pub struct TkaSynced {
    pub(crate) result:
        Result<Option<crate::tka_sync::SyncedTka>, crate::tka_sync::TkaSyncDriverError>,
    /// The [`ControlRunner::tka_generation`] captured when this sync was spawned; the handler
    /// discards the result if it no longer matches (the lock was disabled/re-synced mid-flight).
    pub(crate) generation: u64,
}

impl Message<TkaSynced> for ControlRunner {
    type Reply = ();

    async fn handle(&mut self, msg: TkaSynced, _ctx: &mut Context<Self, Self::Reply>) {
        self.apply_tka_synced(msg.result, msg.generation).await;
    }
}

impl Message<DerpLatencyMeasurement> for ControlRunner {
    type Reply = ();

    async fn handle(&mut self, msg: DerpLatencyMeasurement, _ctx: &mut Context<Self, Self::Reply>) {
        let measurements = msg.measurement.as_ref().clone();

        // Publish the net-report snapshot for `Device::netcheck` (the daemon's `tnet netcheck`) from
        // the same measurements, before the home-region short-circuit below — an empty set still
        // yields a (default/empty) report rather than a stale one.
        self.netcheck
            .send_replace(crate::status::NetcheckReport::from_region_results(
                &measurements,
            ));

        if measurements.is_empty() {
            tracing::debug!("derp latency measurements empty");
            return;
        };

        // Record this cycle into the rolling history and evict reports older than the smoothing
        // window, then compute each region's `bestRecent` (5-min min). `Instant::now()` is the
        // arrival stamp; `best_recent` takes it as a param so the decision stays unit-testable.
        let now = Instant::now();
        self.derp_report_history
            .push((now, msg.measurement.clone()));
        self.derp_report_history
            .retain(|(stamp, _)| now.saturating_duration_since(*stamp) <= DERP_HISTORY_MAX_AGE);
        let best_recent = best_recent(&self.derp_report_history, now, DERP_HISTORY_MAX_AGE);

        // Apply selection hysteresis (the pure decision lives in `select_home_region` for testability)
        // so jitter between near-equal regions does not flap the home relay. Go's asymmetric
        // smoothed-best vs raw-old comparison lives in `select_home_region`; here we just resolve the
        // chosen id back to its current-cycle latency for the home-region record + control update.
        let selected_id = select_home_region(
            self.home_region.map(|(id, _)| id),
            &measurements,
            &best_recent,
        )
        .expect("non-empty measurements always yield a selection");
        // `select_home_region` only ever returns an id drawn from `measurements`, so this lookup
        // always succeeds (same invariant the prior impl relied on when it returned the result by
        // reference). We record the current-cycle (raw) latency for the chosen region.
        let selected_latency = measurements
            .iter()
            .find(|m| m.id == selected_id)
            .expect("the selected region id is always one of the measurements")
            .latency;

        let iter = measurements.iter().map(|result| {
            (
                result.latency_map_key.as_str(),
                result.latency.as_secs_f64(),
            )
        });

        if self.home_region.map(|(id, _)| id) != Some(selected_id) {
            tracing::debug!(selected_region_id = ?selected_id, "updating home region");
        }
        self.home_region = Some((selected_id, selected_latency));
        // Advertise the smoothed home to control AND drive the local DERP relay to the same region
        // (Go `report.PreferredDERP` feeds both). `send_replace` wakes the watch on every send (it
        // does NOT coalesce same-value writes), so Multiderp's bridge sees a `SetHomeRegion` each
        // cycle; the de-dup is one layer down — `home_transition` returns `Unchanged` for a
        // re-selection of the current home, so the relay only churns on an actual home change.
        self.client.set_home_region(selected_id, iter).await;
        self.params.home_region.send_replace(Some(selected_id));
    }
}

/// The cadence on which this node re-measures the derp map — the fork's analog of Go netcheck's
/// `fullReportInterval` (also 5 minutes). This fork has no netcheck timer of its own: a re-measure
/// is triggered by control pushing a derp map and by a link change
/// ([`MeasureNow`](crate::derp_latency::MeasureNow)), so this is the *assumed* cycle length that
/// [`DERP_HISTORY_MAX_AGE`] is sized from, not a period anything here sleeps on.
const FULL_REPORT_INTERVAL: Duration = Duration::from_secs(5 * 60);

/// The window over which the rolling DERP-latency history is retained: it feeds both `best_recent`
/// (the home-region smoothing) and [`ControlRunner::recent_region_latency`] (the exit-node
/// suggestion's per-region ranking), exactly as Go's one `netcheck.Client.prev` history feeds both
/// `bestRecent` and `RecentRegionLatency`.
///
/// Sized as Go sizes it — `maxAge = fullReportInterval + ReportTimeout` — so the window covers a
/// whole re-measure cycle *plus* the time one measurement is allowed to take, and a report cannot
/// age out before its successor exists. Go's stronger guarantee ("the window always holds one
/// *full* report") does not carry over: `ts_netcheck`'s `complete_threshold` means **every**
/// measurement here is partial, so the window holds a union of partial reports rather than one
/// complete one — which is precisely why the union is what the suggestion ranks on.
const DERP_HISTORY_MAX_AGE: Duration =
    FULL_REPORT_INTERVAL.saturating_add(ts_netcheck::REPORT_TIMEOUT);

/// Compute each region's `bestRecent` — its **minimum** latency over the reports within
/// `max_age` of `now` (Go `addReportHistoryAndSetPreferredDERP`'s `bestRecent` map). Reports older
/// than the window are ignored. `now` and `max_age` are parameters (not clock-read) so this is
/// deterministically unit-testable. A region absent from every in-window report is absent from the
/// result.
///
/// This is also the body of Go's `netcheck.Client.RecentRegionLatency()` — the same "lowest latency
/// seen per region across the retained history" map — which is why
/// [`ControlRunner::recent_region_latency`] serves the exit-node suggestion straight out of it
/// rather than keeping a second history.
fn best_recent(
    history: &[(Instant, Arc<Vec<ts_netcheck::RegionResult>>)],
    now: Instant,
    max_age: Duration,
) -> HashMap<ts_derp::RegionId, Duration> {
    let mut best: HashMap<ts_derp::RegionId, Duration> = HashMap::new();
    for (stamp, report) in history {
        // Skip reports outside the window. `saturating_duration_since` guards a `stamp` that is
        // somehow after `now` (clock skew): age 0, always in-window.
        if now.saturating_duration_since(*stamp) > max_age {
            continue;
        }
        for r in report.iter() {
            best.entry(r.id)
                .and_modify(|d| {
                    if r.latency < *d {
                        *d = r.latency;
                    }
                })
                .or_insert(r.latency);
        }
    }
    best
}

/// Choose the DERP home region id, applying Go's selection hysteresis
/// (`netcheck.addReportHistoryAndSetPreferredDERP`). Pure so the decision is unit-testable.
///
/// `measurements` is the current cycle sorted by latency ascending (so `measurements[0]` is the
/// raw-current best). `best_recent` is each region's smoothed (5-min-min) latency. Matching Go's
/// **asymmetric** comparison exactly: the new best candidate is chosen by the *smoothed* `best_recent`
/// latency (`bestAny`), while the old/home region is compared using its *current-cycle* (raw)
/// latency (`oldRegionCurLatency`). Smoothing the best damps oscillation of the best region across
/// the switch boundary that the raw-vs-raw comparison (the prior impl) would still flap on.
///
/// Keeps the `current` home region unless the new best is *meaningfully* lower-latency — switching
/// only when BOTH the current region's raw latency exceeds the smoothed-best by at least
/// `PREFERRED_DERP_ABSOLUTE_DIFF` (10ms) AND the smoothed-best is at most two-thirds of the current
/// region's raw latency (a >~33% improvement). On the first selection (`current` is `None`), when the
/// smoothed-best already IS the current region, or when the current region dropped out of the
/// measurements, returns the best directly. `None` only if `measurements` is empty.
fn select_home_region(
    current: Option<ts_derp::RegionId>,
    measurements: &[ts_netcheck::RegionResult],
    best_recent: &HashMap<ts_derp::RegionId, Duration>,
) -> Option<ts_derp::RegionId> {
    /// Go `netcheck.preferredDERPAbsoluteDiff`.
    const PREFERRED_DERP_ABSOLUTE_DIFF: Duration = Duration::from_millis(10);

    // The smoothed latency for a region: its `best_recent` if present, else its current sample (a
    // region seen only this cycle has a 1-sample history, so its min == its current latency anyway).
    let smoothed = |m: &ts_netcheck::RegionResult| -> Duration {
        best_recent.get(&m.id).copied().unwrap_or(m.latency)
    };

    // Pick the best candidate by SMOOTHED latency (Go `bestAny = min over regions of bestRecent`).
    // `measurements` is sorted by raw latency, but smoothing can reorder, so scan for the smoothed
    // minimum explicitly rather than trusting `measurements[0]`.
    let best = measurements.iter().min_by_key(|m| smoothed(m))?;
    let best_any = smoothed(best);

    let Some(old_id) = current.filter(|id| *id != best.id) else {
        // First selection, or the smoothed-best already is the current home region.
        return Some(best.id);
    };

    // Compare against the old region's CURRENT (raw) latency this cycle, if it is still present —
    // Go's `oldRegionCurLatency`, deliberately unsmoothed (the asymmetry).
    match measurements.iter().find(|m| m.id == old_id) {
        Some(old) => {
            // Byte-faithful to Go: `oldRegionCurLatency - bestAny < 10ms || bestAny >
            // oldRegionCurLatency/3*2`. `saturating_sub` matches Go's signed subtraction for the
            // `< 10ms` test (when `old < best_any` Go is negative → `< 10ms` true; saturating_sub
            // floors to 0 → also true). The two-thirds rule uses INTEGER `Duration` division
            // `(old/3)*2` — NOT float `* 2.0/3.0`: Go computes the threshold in integer nanoseconds
            // (`oldNs/3` truncates), and float arithmetic diverges from it at the exact 2/3 boundary
            // with whole-millisecond inputs (e.g. old=36ms, best=24ms: Go's `24ms > 24ms` is false →
            // switch, but float `0.024 > 0.0239999997` is true → keep). `Duration / u32` truncates
            // nanos exactly like Go and `* u32` is exact, reproducing `oldRegionCurLatency/3*2`.
            let keep_old = old.latency.saturating_sub(best_any) < PREFERRED_DERP_ABSOLUTE_DIFF
                || best_any > (old.latency / 3) * 2;
            Some(if keep_old { old.id } else { best.id })
        }
        // The current region is no longer reachable this cycle: take the new best.
        None => Some(best.id),
    }
}

impl Message<EndpointAdvertisement> for ControlRunner {
    type Reply = ();

    async fn handle(&mut self, msg: EndpointAdvertisement, _ctx: &mut Context<Self, Self::Reply>) {
        let endpoints: Vec<Endpoint> = msg
            .endpoints
            .iter()
            .map(|ep| Endpoint {
                endpoint: ep.addr,
                ty: match ep.ty {
                    SelfEndpointType::Local => EndpointType::Local,
                    SelfEndpointType::Stun => EndpointType::Stun,
                    SelfEndpointType::Stun4LocalPort => EndpointType::Stun4LocalPort,
                },
            })
            .collect();

        tracing::debug!(
            n_endpoints = endpoints.len(),
            "advertising endpoints to control"
        );

        self.client.set_endpoints(endpoints).await;
    }
}

/// Re-advertise this node's routable IP prefixes (`Hostinfo.RoutableIPs`) to control — the wire
/// half of a runtime [`Runtime::set_advertise_routes`](crate::Runtime::set_advertise_routes). Sent
/// as a direct `ask` from the runtime (not over the bus), so the route change reaches the live
/// map-poll client. `routes` is the final advertised set the caller wants control to grant.
#[derive(Debug)]
pub struct SetAdvertiseRoutes {
    /// The prefixes to advertise to control (already filtered to the final set).
    pub routes: Vec<ipnet::IpNet>,
}

impl Message<SetAdvertiseRoutes> for ControlRunner {
    type Reply = ();

    async fn handle(&mut self, msg: SetAdvertiseRoutes, _ctx: &mut Context<Self, Self::Reply>) {
        tracing::debug!(n_routes = msg.routes.len(), "advertising routes to control");
        self.client.set_routable_ips(msg.routes).await;
    }
}

/// Update this node's `Hostinfo.Hostname` at control — the wire half of a runtime
/// [`Runtime::set_hostname`](crate::Runtime::set_hostname). A direct `ask` from the runtime, so the
/// change reaches the live map-poll client.
#[derive(Debug)]
pub struct SetHostname {
    /// The new hostname to report to control.
    pub hostname: String,
}

impl Message<SetHostname> for ControlRunner {
    type Reply = ();

    async fn handle(&mut self, msg: SetHostname, _ctx: &mut Context<Self, Self::Reply>) {
        tracing::debug!("updating hostname at control");
        self.client.set_hostname(msg.hostname).await;
    }
}

#[cfg(test)]
mod reauth_bridge_tests {
    use tokio::sync::watch;

    use super::bridge_reauth_url_to_state;
    use crate::DeviceState;

    fn url(s: &str) -> url::Url {
        s.parse().unwrap()
    }

    /// The bridge maps a surfaced re-auth URL onto `DeviceState::NeedsLogin(url)` — the fix's core:
    /// a mid-session `MachineNotAuthorized` (forwarded by the control client as `Some(url)`) becomes
    /// the "needs login" state the IPN bus turns into `browse_to_url`.
    #[test]
    fn bridge_maps_auth_url_to_needs_login() {
        let u = url("https://login.example/auth");
        let (tx, rx) = watch::channel(DeviceState::Running);

        bridge_reauth_url_to_state(&tx, Some(&u));

        assert_eq!(*rx.borrow(), DeviceState::NeedsLogin(u));
    }

    /// `None` never drives a transition — the recovery to `Running` is the netmap self-node
    /// handler's job, so the bridge ignores a `None` and leaves the state untouched.
    #[test]
    fn bridge_none_leaves_state_unchanged() {
        let (tx, rx) = watch::channel(DeviceState::Running);

        bridge_reauth_url_to_state(&tx, None);

        assert_eq!(*rx.borrow(), DeviceState::Running);
    }

    /// Re-surfacing the same URL across retries does not re-fire the watch (`send_if_modified`
    /// dedupe against the cell's current value), so a stuck re-auth does not thrash subscribers.
    #[test]
    fn bridge_same_url_does_not_refire() {
        let u = url("https://login.example/auth");
        let (tx, mut rx) = watch::channel(DeviceState::Running);

        bridge_reauth_url_to_state(&tx, Some(&u)); // first: fires
        assert!(rx.has_changed().unwrap(), "first NeedsLogin fires");
        rx.mark_unchanged();
        bridge_reauth_url_to_state(&tx, Some(&u)); // same URL: deduped
        assert!(
            !rx.has_changed().unwrap(),
            "the same re-auth URL must not re-fire the state watch"
        );
    }

    /// A genuinely different re-auth URL after a prior one fires again (the dedupe tracks changes,
    /// it does not pin the first URL forever).
    #[test]
    fn bridge_new_url_after_prior_fires() {
        let a = url("https://login.example/a");
        let b = url("https://login.example/b");
        let (tx, rx) = watch::channel(DeviceState::Running);

        bridge_reauth_url_to_state(&tx, Some(&a));
        bridge_reauth_url_to_state(&tx, Some(&b));

        assert_eq!(*rx.borrow(), DeviceState::NeedsLogin(b));
    }

    /// End-to-end of the *clear* contract: after the bridge sets `NeedsLogin`, the netmap self-node
    /// path (modeled here as a direct `send_replace(Running)`, the exact transition the
    /// `StreamMessage::Next` handler performs on the next good self-node) flips back to `Running`.
    /// This pins that the bridge does NOT need a `None`-clear arm — recovery is owned elsewhere.
    #[test]
    fn running_netmap_clears_needs_login() {
        let u = url("https://login.example/auth");
        let (tx, rx) = watch::channel(DeviceState::Running);

        bridge_reauth_url_to_state(&tx, Some(&u));
        assert_eq!(*rx.borrow(), DeviceState::NeedsLogin(u));

        // The self-node handler's recovery transition (next good netmap self-node → Running).
        tx.send_replace(DeviceState::Running);
        assert_eq!(*rx.borrow(), DeviceState::Running);
    }

    /// Fix 2 — the bridge must NOT clobber an in-flight automatic re-auth. While the cell holds
    /// `Reauthenticating`, an auth URL surfaced by the rotated re-register (over the same
    /// `auth_url_tx` cell) must leave the state UNTOUCHED: flipping to `NeedsLogin` would surface a
    /// misleading `browse_to_url` on a headless node and, by moving the cell off `Reauthenticating`,
    /// re-arm a second node-key rotation that loses the original `OldNodeKey` anchor. The auto-reauth
    /// path owns the cell until it recovers to `Running` or trips to `Expired`.
    #[test]
    fn bridge_does_not_clobber_reauthenticating() {
        let u = url("https://login.example/auth");
        let (tx, rx) = watch::channel(DeviceState::Reauthenticating);

        bridge_reauth_url_to_state(&tx, Some(&u));

        assert_eq!(
            *rx.borrow(),
            DeviceState::Reauthenticating,
            "a surfaced auth URL must not downgrade an in-flight auto-reauth to NeedsLogin"
        );
    }

    /// The no-clobber guard is scoped to `Reauthenticating` only — it does not change the bridge's
    /// behavior in any other state. From `Running` (and likewise `Connecting`/`NeedsLogin`) a
    /// surfaced URL still drives `NeedsLogin` as before, so an ordinary interactive re-auth is
    /// unaffected.
    #[test]
    fn bridge_still_sets_needs_login_from_non_reauthenticating() {
        let u = url("https://login.example/auth");
        for start in [
            DeviceState::Running,
            DeviceState::Connecting,
            DeviceState::Expired,
        ] {
            let (tx, rx) = watch::channel(start);
            bridge_reauth_url_to_state(&tx, Some(&u));
            assert_eq!(*rx.borrow(), DeviceState::NeedsLogin(u.clone()));
        }
    }
}

#[cfg(test)]
mod sticky_pop_browser_url_tests {
    use tokio::sync::watch;

    use super::sticky_update_pop_browser_url;

    fn url(s: &str) -> url::Url {
        s.parse().unwrap()
    }

    /// A non-empty URL publishes to the cell.
    #[test]
    fn non_empty_url_publishes() {
        let (tx, rx) = watch::channel(None);
        let u = url("https://login.example/consent");
        sticky_update_pop_browser_url(&tx, Some(&u));
        assert_eq!(*rx.borrow(), Some(u));
    }

    /// An absent (`None`) update — the common netmap tick — must NOT reset the cell. This is the
    /// regression guard for the thrash bug (a reset-every-tick would coalesce the URL away on the bus).
    #[test]
    fn absent_update_does_not_reset() {
        let u = url("https://login.example/consent");
        let (tx, rx) = watch::channel(Some(u.clone()));
        // Simulate many empty netmap updates.
        for _ in 0..5 {
            sticky_update_pop_browser_url(&tx, None);
        }
        assert_eq!(
            *rx.borrow(),
            Some(u),
            "empty updates must not clear the URL"
        );
    }

    /// The same URL repeated does not re-fire the watch (in-place dedupe via `send_if_modified`), so
    /// a subscriber isn't woken spuriously. Proven by the borrow not having been marked changed.
    #[test]
    fn repeated_same_url_does_not_refire() {
        let u = url("https://login.example/consent");
        let (tx, mut rx) = watch::channel(None);
        sticky_update_pop_browser_url(&tx, Some(&u)); // first: fires
        assert!(rx.has_changed().unwrap(), "first non-empty URL fires");
        rx.mark_unchanged();
        sticky_update_pop_browser_url(&tx, Some(&u)); // same: deduped
        assert!(
            !rx.has_changed().unwrap(),
            "repeating the same URL must not re-fire the watch"
        );
    }

    /// A genuinely new URL after a prior one fires again (sticky but tracks changes).
    #[test]
    fn new_url_after_prior_fires() {
        let a = url("https://login.example/a");
        let b = url("https://login.example/b");
        let (tx, rx) = watch::channel(None);
        sticky_update_pop_browser_url(&tx, Some(&a));
        sticky_update_pop_browser_url(&tx, Some(&b));
        assert_eq!(*rx.borrow(), Some(b));
    }

    /// The realistic session sequence: a URL stays sticky through a run of `None` ticks, and a
    /// *different* URL after that gap still fires. Chains the legs the other tests cover in isolation
    /// (the actual control cadence is "URL, then many empty updates, then maybe a new URL").
    #[test]
    fn sticky_through_none_gap_then_new_url_fires() {
        let a = url("https://login.example/a");
        let b = url("https://login.example/b");
        let (tx, rx) = watch::channel(None);
        sticky_update_pop_browser_url(&tx, Some(&a));
        for _ in 0..3 {
            sticky_update_pop_browser_url(&tx, None);
        }
        assert_eq!(*rx.borrow(), Some(a), "stayed sticky through the None gap");
        sticky_update_pop_browser_url(&tx, Some(&b));
        assert_eq!(
            *rx.borrow(),
            Some(b),
            "a new URL after a None gap still fires"
        );
    }

    /// Returning to a previously-seen URL (A → B → A) re-fires: the dedupe is against the cell's
    /// *current* value, not a full history, so A after B is a genuine change.
    #[test]
    fn returning_to_prior_url_refires() {
        let a = url("https://login.example/a");
        let b = url("https://login.example/b");
        let (tx, mut rx) = watch::channel(None);
        sticky_update_pop_browser_url(&tx, Some(&a));
        sticky_update_pop_browser_url(&tx, Some(&b));
        rx.mark_unchanged();
        sticky_update_pop_browser_url(&tx, Some(&a)); // back to A: differs from current (B) → fires
        assert!(
            rx.has_changed().unwrap(),
            "returning to a prior URL re-fires"
        );
        assert_eq!(*rx.borrow(), Some(a));
    }

    /// End-to-end de-thrash: feed a realistic netmap cadence (empty, empty, URL, empty, empty)
    /// through the producer into a cell, and count the changes a `run_bus`-style subscriber would
    /// observe via `changed()`. The whole point of the fix is that exactly ONE change survives the
    /// surrounding `None` thrash — the pre-fix code (`send_replace` every tick) would have woken the
    /// subscriber on every empty tick and coalesced the URL away. This exercises the producer + the
    /// watch-subscribe path together (the two halves the unit tests cover in isolation).
    #[tokio::test]
    async fn end_to_end_one_change_survives_none_thrash() {
        let u = url("https://login.example/consent");
        let (tx, mut rx) = watch::channel(None);
        // The cadence control actually sends: mostly-empty MapResponses with one carrying the URL.
        let cadence = [None, None, Some(&u), None, None];
        for incoming in cadence {
            sticky_update_pop_browser_url(&tx, incoming);
        }
        // A subscriber sees exactly one change, and it carries the URL (not a coalesced `None`).
        let mut changes = 0;
        while rx.has_changed().unwrap() {
            let v = rx.borrow_and_update().clone();
            changes += 1;
            assert_eq!(v, Some(u.clone()), "the surviving change carries the URL");
        }
        assert_eq!(changes, 1, "exactly one change survives the None thrash");
    }
}

#[cfg(test)]
mod home_region_hysteresis_tests {
    use core::time::Duration;
    use std::{collections::HashMap, sync::Arc, time::Instant};

    use ts_derp::RegionId;
    use ts_netcheck::RegionResult;

    use super::{DERP_HISTORY_MAX_AGE, best_recent, select_home_region};

    fn region(id: u32, latency_ms: u64) -> RegionResult {
        RegionResult {
            latency: Duration::from_millis(latency_ms),
            id: RegionId(core::num::NonZeroU32::new(id).unwrap()),
            latency_map_key: format!("region-{id}"),
            connected_remote: "127.0.0.1:0".parse().unwrap(),
        }
    }

    fn rid(id: u32) -> RegionId {
        RegionId(core::num::NonZeroU32::new(id).unwrap())
    }

    /// Call `select_home_region` with NO smoothing history — `best_recent` empty, so each region's
    /// smoothed latency falls back to its current sample, reproducing the original raw-vs-raw
    /// hysteresis these tests pin. (The smoothing-specific tests below pass a populated map.)
    fn sel(current: Option<RegionId>, m: &[RegionResult]) -> Option<RegionId> {
        select_home_region(current, m, &HashMap::new())
    }

    /// Empty measurements yield no selection.
    #[test]
    fn empty_measurements_select_none() {
        assert!(sel(Some(rid(1)), &[]).is_none());
        assert!(sel(None, &[]).is_none());
    }

    /// First selection (no current home region) takes the best (lowest-latency) region directly.
    #[test]
    fn first_selection_takes_best() {
        let m = [region(1, 20), region(2, 50)];
        assert_eq!(sel(None, &m).unwrap(), rid(1));
    }

    /// Jitter within the 10ms absolute-diff band keeps the current region (no flap). Current=region 2
    /// at 25ms; new best=region 1 at 20ms (only 5ms better) -> keep region 2.
    #[test]
    fn keeps_current_when_within_absolute_diff() {
        let m = [region(1, 20), region(2, 25)];
        assert_eq!(
            sel(Some(rid(2)), &m).unwrap(),
            rid(2),
            "a 5ms improvement (< 10ms) must not flap the home region"
        );
    }

    /// A meaningful improvement (>10ms AND best <= 2/3 of current) switches. Current=region 2 at
    /// 100ms; new best=region 1 at 20ms -> switch to region 1.
    #[test]
    fn switches_on_meaningful_improvement() {
        let m = [region(1, 20), region(2, 100)];
        assert_eq!(
            sel(Some(rid(2)), &m).unwrap(),
            rid(1),
            "a large improvement must switch the home region"
        );
    }

    /// The two-thirds rule: even past the 10ms absolute diff, an improvement that does not beat 2/3
    /// of the current latency keeps the current region. current=60ms, best=45ms: diff=15ms (>10ms,
    /// so the absolute test alone would switch), but 45 > 60*2/3=40, so keep.
    #[test]
    fn keeps_current_when_two_thirds_rule_not_met() {
        let m = [region(1, 45), region(2, 60)];
        assert_eq!(
            sel(Some(rid(2)), &m).unwrap(),
            rid(2),
            "best (45ms) is not <= 2/3 of current (40ms), so keep current despite >10ms diff"
        );
    }

    /// When the current home region is no longer present in the measurements, take the new best.
    #[test]
    fn switches_when_current_region_absent() {
        let m = [region(1, 20), region(3, 25)];
        assert_eq!(
            sel(Some(rid(2)), &m).unwrap(),
            rid(1),
            "a current region absent from the measurements falls through to the best"
        );
    }

    /// When the best already IS the current home region, it is kept (no spurious change).
    #[test]
    fn keeps_current_when_it_is_already_best() {
        let m = [region(2, 20), region(1, 50)];
        assert_eq!(sel(Some(rid(2)), &m).unwrap(), rid(2));
    }

    /// The exit-node suggestion ranks on the *union* of the retained partial measurements, not on
    /// the newest one — this fork's form of upstream's `TestRecentReportsRetainFullNetcheck`.
    ///
    /// Go retains reports for `fullReportInterval + ReportTimeout` so the history is guaranteed to
    /// hold one *full* netcheck. This fork never produces a full one: `complete_threshold` ends
    /// every measurement once three regions have answered, so the window instead has to hold enough
    /// *partial* reports for their union to cover every region. Both halves are asserted against
    /// the production code — `best_recent`, which `ControlRunner::recent_region_latency` serves the
    /// suggestion from, and `suggest_exit_node`, the ranking itself.
    #[test]
    fn partial_reports_union_keeps_every_region_rankable() {
        use ts_control::StableNodeId;

        use crate::exit_node_suggest::{ExitNodeCandidate, suggest_exit_node};

        let now = Instant::now();

        // Six measurement cycles, each naming only three regions — what a `complete_threshold: 3`
        // measurement produces against a real derp map. The nearby regions 1/2/3 answer nearly every
        // cycle; region 9 (far away, 300ms) got in exactly ONE cycle, four minutes ago; region 7 has
        // never been measured at all. The 10-minute-old cycle is outside the retention window and
        // its absurdly good 5ms sample for region 9 must not survive into the ranking.
        let history = vec![
            (now - Duration::from_secs(600), Arc::new(vec![region(9, 5)])),
            (
                now - Duration::from_secs(240),
                Arc::new(vec![region(1, 13), region(2, 20), region(9, 300)]),
            ),
            (
                now - Duration::from_secs(180),
                Arc::new(vec![region(1, 11), region(2, 21), region(3, 30)]),
            ),
            (
                now - Duration::from_secs(120),
                Arc::new(vec![region(1, 14), region(2, 19), region(3, 31)]),
            ),
            (
                now - Duration::from_secs(60),
                Arc::new(vec![region(1, 10), region(2, 23), region(3, 29)]),
            ),
            (
                now,
                Arc::new(vec![region(1, 12), region(2, 18), region(3, 28)]),
            ),
        ];

        // Half one: every region any in-window cycle reached still has a latency to rank on, at its
        // lowest in-window sample. Region 9's stale 5ms is evicted; its 300ms stands.
        let recent = best_recent(&history, now, DERP_HISTORY_MAX_AGE);
        for (id, want_ms) in [(1, 10), (2, 18), (3, 28), (9, 300)] {
            assert_eq!(
                recent.get(&rid(id)).copied(),
                Some(Duration::from_millis(want_ms)),
                "region {id} must still be rankable off the retained history"
            );
        }
        assert!(
            !recent.contains_key(&rid(7)),
            "a region no in-window cycle ever measured has no latency"
        );

        fn exit_node(id: u32, derp: u32) -> ExitNodeCandidate {
            ExitNodeCandidate {
                stable_id: StableNodeId(format!("stable{id}")),
                name: format!("peer{id}"),
                derp_region: Some(rid(derp)),
                online: Some(true),
                advertises_exit_route: true,
                has_suggest_cap: true,
            }
        }
        // One exit node in the far-but-measured region 9, one in the never-measured region 7.
        let candidates = [exit_node(9, 9), exit_node(7, 7)];
        let first_node = |nodes: &[ExitNodeCandidate], _: Option<&StableNodeId>| nodes[0].clone();

        // Half two: ranked on the history, region 9's four-minute-old 300ms beats region 7's nothing,
        // and the uniform region pick is never reached.
        let no_random = |_: &[RegionId]| panic!("select_region must not be reached");
        let got = suggest_exit_node(Some(1), &recent, &candidates, None, &no_random, &first_node)
            .expect("ok")
            .expect("some");
        assert_eq!(
            got.id,
            StableNodeId("stable9".to_owned()),
            "a region measured four minutes ago must still beat an unmeasured one"
        );

        // The regression this replaces: ranked on the NEWEST report alone — regions 1/2/3, the only
        // ones that cycle answered — neither candidate's region has a latency, so the suggestion
        // falls through to the uniform random pick and the unmeasured region 7 can win the coin flip.
        let newest: HashMap<RegionId, Duration> = history
            .last()
            .unwrap()
            .1
            .iter()
            .map(|r| (r.id, r.latency))
            .collect();
        let pick_seven = |regions: &[RegionId]| {
            assert!(regions.contains(&rid(7)) && regions.contains(&rid(9)));
            rid(7)
        };
        let got = suggest_exit_node(
            Some(1),
            &newest,
            &candidates,
            None,
            &pick_seven,
            &first_node,
        )
        .expect("ok")
        .expect("some");
        assert_eq!(
            got.id,
            StableNodeId("stable7".to_owned()),
            "ranking on one partial report leaves both candidates unmeasured and picks at random"
        );
    }

    /// `best_recent` is each region's MINIMUM latency over the in-window reports; a report older than
    /// `max_age` is excluded.
    #[test]
    fn best_recent_is_min_over_window_and_evicts_aged() {
        let now = Instant::now();
        // Two in-window reports for region 1 (50ms then 20ms) → min 20ms; region 2 once at 30ms.
        // One aged report (region 1 at 5ms) outside the window must be ignored.
        let history = vec![
            (
                now - Duration::from_secs(10 * 60), // aged out (> 5min)
                Arc::new(vec![region(1, 5)]),
            ),
            (
                now - Duration::from_secs(60),
                Arc::new(vec![region(1, 50), region(2, 30)]),
            ),
            (now, Arc::new(vec![region(1, 20)])),
        ];
        let br = best_recent(&history, now, DERP_HISTORY_MAX_AGE);
        assert_eq!(
            br.get(&rid(1)).copied(),
            Some(Duration::from_millis(20)),
            "region 1 min over the window is 20ms (the aged 5ms is excluded)"
        );
        assert_eq!(br.get(&rid(2)).copied(), Some(Duration::from_millis(30)));
    }

    /// The asymmetric comparison: the new best is chosen by its SMOOTHED (best_recent) latency while
    /// the old region is compared on its RAW current latency. A best region whose CURRENT sample
    /// looks much better but whose 5-min MIN is only marginally better must NOT flap the home region
    /// — exactly the oscillation the raw-vs-raw comparison would have switched on.
    #[test]
    fn smoothed_best_damps_oscillation_across_boundary() {
        // Current home = region 2, raw 60ms this cycle. Region 1's CURRENT sample is 20ms (a >2/3,
        // >10ms improvement → raw-vs-raw would SWITCH), but its 5-min MIN (best_recent) is 50ms
        // (it oscillates). Smoothed-best 50ms vs raw-old 60ms: diff 10ms is NOT < 10ms, but
        // 50 > 60*2/3=40 → keepOld. So we KEEP region 2, where the raw comparison would have flapped.
        let m = [region(1, 20), region(2, 60)];
        let mut br = HashMap::new();
        br.insert(rid(1), Duration::from_millis(50)); // smoothed best is worse than its raw sample
        br.insert(rid(2), Duration::from_millis(60));
        assert_eq!(
            select_home_region(Some(rid(2)), &m, &br).unwrap(),
            rid(2),
            "a best region whose 5-min min is only marginally better must not flap the home region"
        );

        // Sanity: with NO smoothing (raw 20ms best), the same inputs WOULD switch — proving the
        // smoothing is what holds it.
        assert_eq!(
            select_home_region(Some(rid(2)), &m, &HashMap::new()).unwrap(),
            rid(1),
            "raw-vs-raw (no smoothing) switches on the 20ms-vs-60ms current samples"
        );
    }

    /// Smoothing can reorder which region is "best": `measurements` is sorted by raw latency, but the
    /// smoothed minimum may favor a different region. `select_home_region` must pick by smoothed
    /// latency, not blindly trust `measurements[0]`.
    #[test]
    fn smoothed_best_may_differ_from_raw_first() {
        // Raw order: region 1 (10ms) is first. But region 2's 5-min min is 5ms while region 1's is
        // 40ms (region 1's 10ms was a lucky low sample). Smoothed-best is region 2. First selection.
        let m = [region(1, 10), region(2, 12)];
        let mut br = HashMap::new();
        br.insert(rid(1), Duration::from_millis(40));
        br.insert(rid(2), Duration::from_millis(5));
        assert_eq!(
            select_home_region(None, &m, &br).unwrap(),
            rid(2),
            "the smoothed-best region wins even when it is not the raw-latency first"
        );
    }

    /// Byte-faithful integer two-thirds boundary (the float-vs-integer divergence): at exactly
    /// `best == old * 2/3` (old=36ms, best=24ms), Go's integer `bestAny > old/3*2` = `24ms > 24ms`
    /// is FALSE, so it does NOT keep on the 2/3 arm; and `cond_a` `36-24=12ms < 10ms` is also false,
    /// so Go SWITCHES. A float `0.024 > 0.036*2.0/3.0 = 0.0239999997` would wrongly KEEP. This test
    /// pins the integer math: it must switch to the best.
    #[test]
    fn two_thirds_boundary_is_integer_not_float() {
        let m = [region(1, 24), region(2, 36)];
        // No smoothing (raw == smoothed): isolates the 2/3 arithmetic at the exact boundary.
        assert_eq!(
            sel(Some(rid(2)), &m).unwrap(),
            rid(1),
            "at best == old*2/3 the integer rule does NOT keep (Go switches); a float rule would keep"
        );
    }

    /// The `cond_a` (absolute-diff) arm via `saturating_sub`: when the old region's RAW current
    /// latency is FASTER than the smoothed-best (old=20ms raw, smoothed-best=50ms), `old - best_any`
    /// underflows. Go's signed subtraction is negative (`< 10ms` → keepOld); `saturating_sub` floors
    /// to 0 (`< 10ms` → keepOld) — same outcome. The old region is kept.
    #[test]
    fn old_faster_than_smoothed_best_keeps_via_absolute_diff() {
        // Current home = region 2, raw 20ms. Region 1 is the raw-best at 15ms but its smoothed min is
        // 50ms (it oscillates badly). smoothed-best candidate by min = region 2 (raw 20 == smoothed
        // 20, since br[2]=20) vs region 1 smoothed 50 → best is region 2 itself → already-best path.
        // To exercise the old<best_any underflow we need best != old: make region 1 the smoothed best
        // at 18ms but the OLD region's raw 20ms... use: old=region2 raw 20, best=region1 smoothed 18.
        let m = [region(1, 15), region(2, 20)];
        let mut br = HashMap::new();
        br.insert(rid(1), Duration::from_millis(18)); // smoothed-best = region 1 at 18ms
        br.insert(rid(2), Duration::from_millis(25)); // region 2 smoothed worse than its raw 20ms
        // best_any = 18ms (region 1). old (region 2) RAW = 20ms. 20 - 18 = 2ms < 10ms → keepOld.
        assert_eq!(
            select_home_region(Some(rid(2)), &m, &br).unwrap(),
            rid(2),
            "old raw (20ms) within 10ms of smoothed-best (18ms) keeps via the absolute-diff arm"
        );
    }
}

#[cfg(test)]
mod expiry_action_tests {
    use super::{ExpiryAction, expiry_action};

    /// Not expired → `Running`, regardless of the other inputs (the gate fields are only consulted
    /// once the key is expired).
    #[test]
    fn not_expired_is_always_running() {
        for has_auth_key in [false, true] {
            for reauth_enabled in [false, true] {
                for tka_active in [false, true] {
                    assert_eq!(
                        expiry_action(false, has_auth_key, reauth_enabled, tka_active),
                        ExpiryAction::Running,
                        "a non-expired key is Running for any gate combination"
                    );
                }
            }
        }
    }

    /// The ONLY input combination that auto-reauths: expired AND auth key retained AND reauth enabled
    /// AND TKA not enforcing.
    #[test]
    fn expired_with_authkey_reauth_enabled_and_no_tka_reauthenticates() {
        assert_eq!(
            expiry_action(true, true, true, false),
            ExpiryAction::Reauthenticate
        );
    }

    /// Every other expired combination falls back to the terminal `Expired` (today's behavior, no
    /// regression): no auth key, reauth disabled, or TKA enforcing each independently forces Expired.
    #[test]
    fn expired_falls_back_to_expired_for_every_other_combination() {
        // The full expired-input matrix minus the single Reauthenticate cell above.
        for has_auth_key in [false, true] {
            for reauth_enabled in [false, true] {
                for tka_active in [false, true] {
                    let action = expiry_action(true, has_auth_key, reauth_enabled, tka_active);
                    if has_auth_key && reauth_enabled && !tka_active {
                        // The one Reauthenticate cell, asserted above.
                        assert_eq!(action, ExpiryAction::Reauthenticate);
                    } else {
                        assert_eq!(
                            action,
                            ExpiryAction::Expired,
                            "expired with has_auth_key={has_auth_key}, \
                             reauth_enabled={reauth_enabled}, tka_active={tka_active} must be Expired"
                        );
                    }
                }
            }
        }
    }

    /// The TKA safety gate in isolation: even with an auth key and reauth enabled, an ACTIVE lock
    /// forces `Expired` (never rotate an unsigned key on a locked tailnet). This is the hard
    /// constraint from the design — pinned as its own test so a regression that drops the `!tka_active`
    /// term is caught explicitly.
    #[test]
    fn tka_active_forces_expired_even_when_reauth_would_otherwise_fire() {
        assert_eq!(
            expiry_action(true, true, true, true),
            ExpiryAction::Expired,
            "an enforcing Tailnet Lock must veto auto-reauth (unsigned-key lockout safety gate)"
        );
    }

    /// No auth key forces `Expired`: there is nothing to non-interactively re-register with, so even
    /// with reauth enabled and no lock the node goes terminal (unchanged from today).
    #[test]
    fn no_auth_key_forces_expired() {
        assert_eq!(
            expiry_action(true, false, true, false),
            ExpiryAction::Expired
        );
    }

    /// The config opt-out: `reauth_on_expiry=false` forces `Expired` even with an auth key and no
    /// lock (the conservative posture / historical behavior).
    #[test]
    fn reauth_disabled_forces_expired() {
        assert_eq!(
            expiry_action(true, true, false, false),
            ExpiryAction::Expired
        );
    }
}

#[cfg(test)]
mod reauth_circuit_tests {
    use super::{ExpiryAction, MAX_REAUTH_ATTEMPTS, ReauthState, reauth_circuit_step};

    /// Entering reauth (a `Reauthenticate` verdict while NOT already reauthenticating) fires the
    /// one-shot reauth and starts the counter at 1.
    #[test]
    fn enter_reauth_fires_once_and_starts_counter() {
        let step = reauth_circuit_step(ExpiryAction::Reauthenticate, false, 0);
        assert_eq!(step.next, ReauthState::Reauthenticating);
        assert_eq!(
            step.attempts, 1,
            "the entering attempt starts the counter at 1"
        );
        assert!(step.fire_reauth, "entry fires the one-shot Command::Reauth");
    }

    /// A subsequent expired self-node while ALREADY reauthenticating (prior reauth not recovered)
    /// counts up but does NOT re-fire — re-firing would rotate the node key a second time and lose the
    /// original `OldNodeKey` anchor.
    #[test]
    fn still_reauthing_counts_but_does_not_refire() {
        let step = reauth_circuit_step(ExpiryAction::Reauthenticate, true, 1);
        assert_eq!(step.next, ReauthState::Reauthenticating);
        assert_eq!(
            step.attempts, 2,
            "a non-recovering reauth increments the counter"
        );
        assert!(
            !step.fire_reauth,
            "must NOT re-fire reauth while already reauthenticating (no second rotation)"
        );
    }

    /// At `MAX_REAUTH_ATTEMPTS` consecutive non-recovering reauths, the breaker trips to the terminal
    /// `Expired` and stops re-arming reauth (the one-shot auth-key case settles instead of looping).
    #[test]
    fn trips_to_expired_at_cap() {
        // One step below the cap still stays reauthenticating.
        let below =
            reauth_circuit_step(ExpiryAction::Reauthenticate, true, MAX_REAUTH_ATTEMPTS - 2);
        assert_eq!(below.next, ReauthState::Reauthenticating);
        assert!(!below.fire_reauth);

        // The step that reaches the cap flips to terminal Expired and does not fire.
        let at_cap =
            reauth_circuit_step(ExpiryAction::Reauthenticate, true, MAX_REAUTH_ATTEMPTS - 1);
        assert_eq!(at_cap.attempts, MAX_REAUTH_ATTEMPTS);
        assert_eq!(
            at_cap.next,
            ReauthState::Expired,
            "at the cap the circuit breaker trips to terminal Expired"
        );
        assert!(
            !at_cap.fire_reauth,
            "a tripped breaker never fires another reauth"
        );
    }

    /// A good (non-expired) self-node resets the counter to 0 (the node recovered) — so a later,
    /// genuine expiry cycle gets a fresh full budget of attempts, never a stale leftover count.
    #[test]
    fn running_resets_counter() {
        let step = reauth_circuit_step(ExpiryAction::Running, true, MAX_REAUTH_ATTEMPTS);
        assert_eq!(step.next, ReauthState::Running);
        assert_eq!(
            step.attempts, 0,
            "recovery to Running resets the attempt counter"
        );
        assert!(!step.fire_reauth);
    }

    /// The non-reauth terminal path (`expiry_action` → `Expired`: no auth key / reauth disabled /
    /// TKA-locked) reports `Expired` and never fires reauth; the counter is irrelevant (this path
    /// never armed the machine), so it resets to 0.
    #[test]
    fn expired_action_is_terminal_without_firing() {
        let step = reauth_circuit_step(ExpiryAction::Expired, false, 0);
        assert_eq!(step.next, ReauthState::Expired);
        assert!(!step.fire_reauth);
        assert_eq!(step.attempts, 0);
    }

    /// The full episode as the handler drives it, feeding each step's `attempts` into the next: a
    /// one-shot auth key that never recovers fires reauth exactly ONCE, counts each repeated expired
    /// self-node, and settles on terminal `Expired` at the cap — never an indefinite `Reauthenticating`
    /// spell and never a second rotation.
    #[test]
    fn full_episode_one_shot_key_settles_at_expired_after_one_fire() {
        // Step 1: first expired self-node (state was Running → not already reauthing): enter + fire.
        let s1 = reauth_circuit_step(ExpiryAction::Reauthenticate, false, 0);
        assert_eq!(s1.next, ReauthState::Reauthenticating);
        assert!(s1.fire_reauth);
        let mut attempts = s1.attempts;
        let mut fires = 1; // counted the one fire

        // Steps 2..: still expired while reauthenticating — count only, never re-fire — until the cap.
        loop {
            let s = reauth_circuit_step(ExpiryAction::Reauthenticate, true, attempts);
            if s.fire_reauth {
                fires += 1;
            }
            attempts = s.attempts;
            if s.next == ReauthState::Expired {
                break;
            }
            assert_eq!(s.next, ReauthState::Reauthenticating);
            assert!(attempts < MAX_REAUTH_ATTEMPTS);
        }

        assert_eq!(attempts, MAX_REAUTH_ATTEMPTS, "settles exactly at the cap");
        assert_eq!(
            fires, 1,
            "reauth (and thus a node-key rotation) fires exactly once across the whole episode"
        );
    }

    /// Recovery then a fresh failing episode: `Reauthenticating → Running` (reset) → a later expiry
    /// re-enters and re-fires. Proves the reset gives the next episode a full attempt budget rather
    /// than carrying a stale count that would trip the breaker early.
    #[test]
    fn recovery_then_new_episode_re_fires() {
        // Episode 1 entry.
        let e1 = reauth_circuit_step(ExpiryAction::Reauthenticate, false, 0);
        assert!(e1.fire_reauth);
        // A non-recovering step, then recovery to Running resets.
        let mid = reauth_circuit_step(ExpiryAction::Reauthenticate, true, e1.attempts);
        assert_eq!(mid.attempts, 2);
        let recovered = reauth_circuit_step(ExpiryAction::Running, true, mid.attempts);
        assert_eq!(recovered.attempts, 0);

        // Episode 2: a fresh expiry (state is Running again → not already reauthing) re-enters + fires.
        let e2 = reauth_circuit_step(ExpiryAction::Reauthenticate, false, recovered.attempts);
        assert_eq!(e2.next, ReauthState::Reauthenticating);
        assert_eq!(e2.attempts, 1, "the new episode starts fresh at 1");
        assert!(
            e2.fire_reauth,
            "a new episode after recovery fires reauth again"
        );
    }
}

#[cfg(test)]
mod self_lockout_tests {
    use ts_tka::{AumHash, Authority, State};

    use super::{SelfLockVerdict, self_lock_verdict};

    fn node_key() -> ts_keys::NodePublicKey {
        ts_keys::NodePrivateKey::random().public_key()
    }

    /// An empty key-signature is the "not signed yet" case: `Unsigned`, never a lockout warning —
    /// so a tailnet that simply has not signed this node does not spam a `warn`.
    #[test]
    fn empty_signature_is_unsigned_not_locked_out() {
        let authority = Authority::from_state(AumHash([0; 32]), State::default());
        assert_eq!(
            self_lock_verdict(&node_key(), &[], &authority),
            SelfLockVerdict::Unsigned
        );
    }

    /// A non-empty key-signature that does not authorize self classifies as `LockedOut` — the
    /// operator-facing condition — and the verdict carries the verify error string for the log. Here
    /// the blob is non-empty (so we attempt verification rather than short-circuiting to `Unsigned`)
    /// but is not a valid NodeKeySignature CBOR (`0x01` decodes as a bare uint with trailing bytes),
    /// so `node_key_authorized` returns a `Decode` error → `LockedOut`. The cryptographic-rejection
    /// arms (`UntrustedKey` / `BadSignature` for a well-formed-but-untrusted NKS) are covered by
    /// `ts_tka`'s own `node_key_authorized` tests; this only needs to prove the runtime classifier
    /// routes a verify `Err` to `LockedOut`.
    #[test]
    fn unverifiable_signature_is_locked_out() {
        let authority = Authority::from_state(AumHash([0; 32]), State::default());
        let verdict = self_lock_verdict(&node_key(), &[0x01, 0x02, 0x03], &authority);
        assert!(
            matches!(verdict, SelfLockVerdict::LockedOut(_)),
            "a signature the lock cannot authorize must classify as LockedOut, got {verdict:?}"
        );
    }
}

#[cfg(test)]
mod cached_replay_tests {
    //! The cold-start replay of a netmap cached under Tailnet Lock: which of its peers this node
    //! dials before control has answered.
    //!
    //! Go replays its cached map through `setNetMapLocked`, so `tkaFilterNetmapLocked`
    //! (`ipn/ipnlocal/tailnet-lock.go`, upstream `9ea7cba44591e0cd840c6c94d23274dd222059bf`) runs
    //! over it against the authority it re-opened from disk, and the peers holding a valid signature
    //! survive. These cover the same outcome here, over a chain persisted beside the netmap.

    use ed25519_dalek::SigningKey;
    use ts_tka::{Aum, AumHash, AumKey, Authority, KeyKind, MemAumStore, NodeKeySignature};

    use super::{TkaStatus, vouch_cached_peers};
    use crate::peer_tracker::tka_tests::peer_node;

    /// The node key of the peer the lock authorizes in these tests.
    const SIGNED_PEER_KEY: [u8; 32] = [9u8; 32];
    /// The node key of the peer that presents nothing.
    const UNSIGNED_PEER_KEY: [u8; 32] = [10u8; 32];

    /// A locked tailnet: a genesis checkpoint trusting `signer` (signed by it, so the chain verifies),
    /// plus the store and the [`Authority`] a completed sync would hold.
    fn locked_tailnet(signer: &SigningKey) -> (MemAumStore, AumHash, Authority) {
        let key = AumKey {
            kind: KeyKind::Ed25519,
            votes: 1,
            public: signer.verifying_key().to_bytes().to_vec(),
            meta: Vec::new(),
        };
        let mut genesis = Aum::new_genesis_checkpoint(vec![key], vec![vec![0x11; 32]])
            .expect("a well-formed genesis checkpoint");
        genesis.sign(signer);
        let oldest = genesis.hash();
        let store = MemAumStore::from_aums([genesis]);
        let authority = crate::tka_sync::authority_from_encoded_chain(
            &crate::tka_sync::encode_chain(&store, oldest).expect("encode"),
        )
        .expect("the chain verifies");
        (store, oldest, authority)
    }

    /// The status control stamped on the netmap that was cached under `authority`.
    fn cached_lock(authority: &Authority) -> TkaStatus {
        TkaStatus {
            head: authority.head().to_base32(),
            disabled: false,
        }
    }

    /// The headline: with the authority those peers were cached under, the cold start replays the
    /// peer that authority authorizes and drops the one it does not.
    ///
    /// Without the persisted authority this replays **nothing** — which is what made a locked
    /// tailnet lose the cache entirely, while Go kept dialing its authorized peers.
    #[test]
    fn a_persisted_authority_replays_the_peers_it_authorizes() {
        let signer = SigningKey::from_bytes(&[42u8; 32]);
        let (_store, _oldest, authority) = locked_tailnet(&signer);

        let signed = peer_node(
            "signed-peer",
            SIGNED_PEER_KEY,
            NodeKeySignature::sign_direct(&SIGNED_PEER_KEY, &signer).serialize(),
        );
        let unsigned = peer_node("unsigned-peer", UNSIGNED_PEER_KEY, Vec::new());

        let replayed = vouch_cached_peers(
            Some(&authority),
            &cached_lock(&authority),
            vec![signed, unsigned],
        );

        assert_eq!(
            replayed
                .iter()
                .map(|p| p.stable_id.0.as_str())
                .collect::<Vec<_>>(),
            vec!["signed-peer"],
            "the peer the lock authorizes is dialed at cold start; the unsigned one is dropped"
        );
    }

    /// No persisted authority (this node has never completed a sync, or the chain did not verify) ⇒
    /// no peers. There is nothing to check a signature against, and a peer the lock revoked while
    /// this node was off must not be dialed on the strength of a cache entry.
    #[test]
    fn without_an_authority_no_cached_peer_is_replayed() {
        let signer = SigningKey::from_bytes(&[42u8; 32]);
        let (_store, _oldest, authority) = locked_tailnet(&signer);
        let signed = peer_node(
            "signed-peer",
            SIGNED_PEER_KEY,
            NodeKeySignature::sign_direct(&SIGNED_PEER_KEY, &signer).serialize(),
        );

        assert!(
            vouch_cached_peers(None, &cached_lock(&authority), vec![signed]).is_empty(),
            "a peer nothing can vouch for waits for control's first netmap"
        );
    }

    /// A netmap cached on disk (raw `MapResponse` JSON, exactly what `NetmapCache` persists), in a
    /// directory private to this user so the cache's own vetting accepts it.
    #[cfg(unix)]
    fn cache_dir_with(label: &str, tka_info: &str, peers: &str) -> std::path::PathBuf {
        use std::os::unix::fs::{OpenOptionsExt as _, PermissionsExt as _};

        let dir = std::env::temp_dir().join(format!(
            "ts-rs-cached-replay-{}-{label}",
            std::process::id()
        ));
        std::fs::remove_dir_all(&dir).ok();
        std::fs::create_dir_all(&dir).expect("scratch dir");
        std::fs::set_permissions(&dir, std::fs::Permissions::from_mode(0o700)).expect("chmod");

        let body = format!(
            r#"{{
                {tka_info}
                "Node": {{
                    "ID": 1,
                    "StableID": "self-1",
                    "Name": "self.example.ts.net.",
                    "Addresses": ["100.64.0.1/32"],
                    "CapMap": {{"cache-network-maps": null}}
                }},
                "Peers": [{peers}],
                "DERPMap": {{ "Regions": {{ "3": {{
                    "RegionID": 3, "RegionCode": "tst", "RegionName": "Test", "Nodes": []
                }} }} }}
            }}"#
        );
        let mut file = std::fs::OpenOptions::new()
            .write(true)
            .create_new(true)
            .mode(0o600)
            .open(dir.join(ts_control::NETMAP_CACHE_FILE))
            .expect("cache file");
        std::io::Write::write_all(&mut file, body.as_bytes()).expect("write cached netmap");
        dir
    }

    /// One cached peer, with no `KeySignature` — an unsigned peer, which a lock drops.
    #[cfg(unix)]
    const UNSIGNED_CACHED_PEER: &str = r#"{
        "ID": 2,
        "StableID": "peer-2",
        "Name": "peer.example.ts.net.",
        "Addresses": ["100.64.0.2/32"],
        "Endpoints": ["192.0.2.7:41641"],
        "HomeDERP": 3
    }"#;

    /// End to end over the files a cold start actually finds: the persisted chain is read from the
    /// cache directory, re-verified, and used to judge the cached peers — an unsigned one is dropped
    /// while the rest of the netmap replays.
    ///
    /// The *admit* side is proven on [`vouch_cached_peers`] directly rather than here: a cached
    /// frame's `KeySignature` decodes to the raw bytes of its JSON string, so a real signature (64
    /// arbitrary bytes inside a CBOR structure) cannot be written into a JSON fixture at all.
    #[cfg(unix)]
    #[tokio::test]
    async fn a_cold_start_judges_the_cached_peers_against_the_persisted_chain() {
        let signer = SigningKey::from_bytes(&[42u8; 32]);
        let (store, oldest, authority) = locked_tailnet(&signer);
        let dir = cache_dir_with(
            "with-chain",
            &format!(
                r#""TKAInfo": {{ "Head": "{}", "Disabled": false }},"#,
                authority.head().to_base32()
            ),
            UNSIGNED_CACHED_PEER,
        );
        let cache = ts_control::NetmapCache::new(&dir);
        cache
            .store_tka_chain(&crate::tka_sync::encode_chain(&store, oldest).expect("encode"))
            .await;

        let replayed = super::load_cached_netmap(&cache)
            .await
            .expect("the cached netmap replays");

        assert!(
            replayed.peer_update.is_none(),
            "an unsigned peer is dropped by the same filter the live netmap runs, got {:?}",
            replayed.peer_update
        );
        assert!(
            replayed.node.is_some() && replayed.derp.is_some(),
            "the rest of the netmap carries no peer identity and still replays"
        );

        std::fs::remove_dir_all(&dir).ok();
    }

    /// A netmap cached while the lock was **off** replays whole: there is no authority to consult and
    /// nothing to enforce (Go's filter returns early on `b.tka == nil`).
    #[cfg(unix)]
    #[tokio::test]
    async fn a_cold_start_without_a_lock_replays_every_cached_peer() {
        let dir = cache_dir_with("no-lock", "", UNSIGNED_CACHED_PEER);

        let replayed = super::load_cached_netmap(&ts_control::NetmapCache::new(&dir))
            .await
            .expect("the cached netmap replays");

        assert!(
            matches!(replayed.peer_update, Some(ts_control::PeerUpdate::Full(ref p)) if p.len() == 1),
            "an unlocked tailnet replays its cached peers untouched: {:?}",
            replayed.peer_update
        );

        std::fs::remove_dir_all(&dir).ok();
    }

    /// A chain that does not verify is no authority at all, so the cached peers are withheld — the
    /// same outcome as never having persisted one. The netmap around them still replays.
    #[cfg(unix)]
    #[tokio::test]
    async fn a_cold_start_withholds_peers_when_the_persisted_chain_does_not_verify() {
        let signer = SigningKey::from_bytes(&[42u8; 32]);
        let (_store, _oldest, authority) = locked_tailnet(&signer);
        let tka_info = format!(
            r#""TKAInfo": {{ "Head": "{}", "Disabled": false }},"#,
            authority.head().to_base32()
        );

        // No chain at all.
        let dir = cache_dir_with("no-chain", &tka_info, UNSIGNED_CACHED_PEER);
        let replayed = super::load_cached_netmap(&ts_control::NetmapCache::new(&dir))
            .await
            .expect("the cached netmap replays");
        assert!(replayed.peer_update.is_none());
        assert!(replayed.node.is_some());
        std::fs::remove_dir_all(&dir).ok();

        // A chain blob that is not a verified chain.
        let dir = cache_dir_with("bad-chain", &tka_info, UNSIGNED_CACHED_PEER);
        let cache = ts_control::NetmapCache::new(&dir);
        cache.store_tka_chain(b"ts-tka-chain-v1\nAQID").await;
        let replayed = super::load_cached_netmap(&cache)
            .await
            .expect("the cached netmap replays");
        assert!(replayed.peer_update.is_none());
        assert!(replayed.node.is_some());
        std::fs::remove_dir_all(&dir).ok();
    }

    /// A chain that is not at the head the cached frame recorded describes a different lock — a
    /// disable-and-re-enable under a fresh genesis, or a crash between writing the netmap and syncing
    /// the chain that followed it. It vouches for nothing.
    #[test]
    fn a_chain_at_another_head_vouches_for_nothing() {
        let signer = SigningKey::from_bytes(&[42u8; 32]);
        let (_store, _oldest, authority) = locked_tailnet(&signer);
        let signed = peer_node(
            "signed-peer",
            SIGNED_PEER_KEY,
            NodeKeySignature::sign_direct(&SIGNED_PEER_KEY, &signer).serialize(),
        );

        // Another lock's head: valid base32, not ours.
        let other = TkaStatus {
            head: AumHash([0x5a; 32]).to_base32(),
            disabled: false,
        };
        assert!(vouch_cached_peers(Some(&authority), &other, vec![signed.clone()]).is_empty());

        // A head control did not send, or one this node cannot parse, is not a match either.
        for head in ["", "not base32"] {
            let malformed = TkaStatus {
                head: head.to_string(),
                disabled: false,
            };
            assert!(
                vouch_cached_peers(Some(&authority), &malformed, vec![signed.clone()]).is_empty(),
                "a head that does not parse must not be treated as matching ({head:?})"
            );
        }
    }
}

#[cfg(test)]
mod persisted_tka_chain_tests {
    //! The lifecycle of the Tailnet-Lock chain persisted for the cold-start replay: when it is
    //! removed, and that a removal can never outlive the write that follows it.

    use super::PersistedTkaChain;

    /// A cache directory this test owns, not yet created — [`PersistedTkaChain::store`] creates it
    /// private to this user, which is the same path the runtime takes.
    fn scratch_dir(label: &str) -> std::path::PathBuf {
        let dir =
            std::env::temp_dir().join(format!("ts-rs-tka-chain-{}-{label}", std::process::id()));
        std::fs::remove_dir_all(&dir).ok();
        dir
    }

    /// An opaque chain blob. The cache stores and returns bytes; only the runtime decodes them, and
    /// nothing in this module does.
    const CHAIN: &[u8] = b"ts-tka-chain-v1\nAQID";

    /// A process that has synced nothing still clears a chain an **earlier** process left behind.
    ///
    /// This is the state a cold start is in: `tka_synced` is `None` because this process has not
    /// spoken to control yet, while the chain the last run persisted is sitting in the cache
    /// directory — and that chain is exactly the one the *next* cold start would load. Gating the
    /// removal on live synced state (as the lock-off paths once did) leaves it there for good on a
    /// node whose lock was turned off between runs.
    #[tokio::test]
    async fn a_chain_an_earlier_process_left_is_cleared_by_one_that_synced_nothing() {
        let dir = scratch_dir("stale");
        let cache = ts_control::NetmapCache::new(&dir);
        cache.store_tka_chain(CHAIN).await;
        assert!(
            cache.load_tka_chain().await.is_some(),
            "the earlier process's chain is on disk"
        );

        // Freshly constructed, exactly as `on_start` builds it: no sync has happened.
        let mut chain = PersistedTkaChain::new(Some(&dir));
        chain.discard().await;

        assert!(
            cache.load_tka_chain().await.is_none(),
            "a lock control reports off clears the chain whether or not this process enforced it"
        );

        std::fs::remove_dir_all(&dir).ok();
    }

    /// A removal is complete when it returns, so a chain synced *after* it survives.
    ///
    /// The removal used to be detached (`tokio::spawn`), so a lock disable could leave one queued
    /// while a later sync persisted a fresh chain, and the queued removal would then delete it. The
    /// yields below are where such a task would get to run.
    #[tokio::test]
    async fn a_removal_never_outlives_the_chain_synced_after_it() {
        let dir = scratch_dir("ordering");
        let cache = ts_control::NetmapCache::new(&dir);
        let mut chain = PersistedTkaChain::new(Some(&dir));

        chain.store(CHAIN).await;
        chain.discard().await;
        assert!(
            cache.load_tka_chain().await.is_none(),
            "the removal has happened by the time it returns; nothing is left pending"
        );

        // The lock comes back on and a sync persists a new chain.
        let resynced = b"ts-tka-chain-v1\nBBBB".as_slice();
        chain.store(resynced).await;
        for _ in 0..8 {
            tokio::task::yield_now().await;
        }

        assert_eq!(
            cache.load_tka_chain().await.as_deref(),
            Some(resynced),
            "the freshly synced chain is what the next cold start finds"
        );

        std::fs::remove_dir_all(&dir).ok();
    }

    /// Repeated lock-off netmaps stay cheap: the removal is idempotent and the chain stays gone.
    #[tokio::test]
    async fn repeating_the_removal_is_a_no_op() {
        let dir = scratch_dir("idempotent");
        let cache = ts_control::NetmapCache::new(&dir);
        let mut chain = PersistedTkaChain::new(Some(&dir));

        chain.store(CHAIN).await;
        for _ in 0..3 {
            chain.discard().await;
            assert!(cache.load_tka_chain().await.is_none());
        }

        std::fs::remove_dir_all(&dir).ok();
    }

    /// An embedder that configured no netmap cache directory writes nothing and removes nothing —
    /// there is no cache, so there is never a chain to vouch with.
    #[tokio::test]
    async fn no_cache_directory_means_no_chain_at_all() {
        let mut chain = PersistedTkaChain::new(None);
        chain.store(CHAIN).await;
        chain.discard().await;
    }
}