vector_core/community/v2/

realtime.rs

//! v2 realtime — the `authors`-based live subscription + dispatch.
//!
//! v2 addresses planes by their group pubkey (CORD-01), not v1's `#z` tag, so
//! the subscription is `{kinds:[1059,21059], authors:[…plane pubkeys…]}`. A
//! received event is routed to the owning community and handed to the shared
//! [`inbound::dispatch_wrap`], which fires the protocol-agnostic
//! `InboundEventHandler` the SDK's `on_message` consumes.
//!
//! The kind-1059 dispatch rule (CLAUDE.md A4): the listen loop tries this v2
//! path for events on the v2 subscription; DM gift wraps and 3313 Direct Invites
//! (both `#p=me`) stay on the DM subscription — the author-set here never
//! includes an identity key, so the two never collide.

use std::collections::HashSet;
use std::sync::{Arc, LazyLock, Mutex as StdMutex};

use nostr_sdk::prelude::{Client, Event, Filter, Kind, PublicKey, RelayStatus, RelayUrl, SubscriptionId, Timestamp};
use tokio::sync::mpsc::UnboundedSender;
use tokio::sync::Mutex;

use super::community::CommunityV2;
use super::stream;
use super::{derive, inbound};
use crate::community::{CommunityId, ConcordProtocol, Epoch};
use crate::event_handler::InboundEventHandler;
use crate::ClientRelayExt;

/// The targeted subscription id (streams on desktop).
static V2_SUB_ID: LazyLock<Mutex<Option<SubscriptionId>>> = LazyLock::new(|| Mutex::new(None));
/// The pool-wide subscription id (the path that streams on Android).
static V2_POOLWIDE_SUB_ID: LazyLock<Mutex<Option<SubscriptionId>>> = LazyLock::new(|| Mutex::new(None));
/// The current author-set (sorted hex) — an unchanged set skips a churny
/// unsubscribe+resubscribe, exactly like v1's `COMMUNITY_SUB_SET`.
static V2_SUB_SET: LazyLock<Mutex<Vec<String>>> = LazyLock::new(|| Mutex::new(Vec::new()));

/// Session generation of the last COMPLETED [`refresh_subscription`] pass, or
/// `u64::MAX` before any. Generation-keyed rather than a bare bool so an account
/// swap invalidates it automatically — a per-account global that survived a swap
/// would report account A's readiness for account B.
static V2_SUB_READY_GEN: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(u64::MAX);

#[inline]
fn mark_subscription_ready() {
    V2_SUB_READY_GEN.store(
        crate::db::current_session_id(),
        std::sync::atomic::Ordering::Release,
    );
}

/// Whether the v2 Community subscription pass has COMPLETED for the active
/// session: live channel events are being delivered (or there is genuinely
/// nothing to subscribe to). `false` during the cold-start window where the
/// relay-connect wait + stream-auth priming are still running — the state a
/// health-checking bot could previously not distinguish from "subscribed to
/// nothing".
pub fn subscription_ready() -> bool {
    V2_SUB_READY_GEN.load(std::sync::atomic::Ordering::Acquire)
        == crate::db::current_session_id()
}
/// Outer-wrap ids already dispatched, so the handler fires EXACTLY ONCE per
/// message. The relay pool delivers the same wrap under both the targeted and
/// pool-wide subs and from every relay independently, so without this a bot's
/// `on_message` (and its reply) would run several times per message — the v1
/// community path and the DM path dedup by outer id for the same reason. Cleared
/// on session swap; coarsely bounded (a message's duplicates all arrive within a
/// short window, so a recent-set suffices).
static V2_SEEN_WRAPS: LazyLock<Mutex<HashSet<[u8; 32]>>> = LazyLock::new(|| Mutex::new(HashSet::new()));
/// Bound on [`V2_SEEN_WRAPS`] before a coarse flush.
const SEEN_WRAPS_CAP: usize = 8192;
/// The per-community follow QUEUE. dispatch, boot catch-up, reconnect, and manual
/// sync all just [`enqueue_follow`] — non-blocking + coalesced. A single spawned
/// worker ([`spawn_follow_worker`]) drains it and runs one combined rekey+control
/// follow per community at a time, so two triggers can never concurrently
/// whole-row-save and clobber each other. This replaces the old gate/rerun/
/// spawn-vs-await machinery: an enqueue never blocks its caller, and a junk-wrap
/// flood coalesces to at most one queued + one running follow per community.
/// Reset on session swap (the worker exits when its `std::sync::Arc<crate::db::Session>` invalidates or
/// its channel closes).
static V2_FOLLOW_TX: LazyLock<StdMutex<Option<UnboundedSender<CommunityId>>>> = LazyLock::new(|| StdMutex::new(None));
/// Community ids currently queued or processing — coalesces a burst to one follow.
static V2_FOLLOW_PENDING: LazyLock<StdMutex<HashSet<[u8; 32]>>> = LazyLock::new(|| StdMutex::new(HashSet::new()));
/// Per-community follow serialization, shared by the queue worker AND the inline
/// (headless) follow path. The worker-vs-inline CHOICE is a benign race
/// (`follow_worker_running` is check-then-act; a worker can spawn right after a
/// `false`), so correctness can't ride on it: whichever path runs, the follow body
/// executes under this lock and two follows of one community can never interleave
/// their whole-row saves. Bounded by the held-community count; reset on swap.
static V2_FOLLOW_LOCKS: LazyLock<StdMutex<std::collections::HashMap<[u8; 32], Arc<Mutex<()>>>>> =
    LazyLock::new(|| StdMutex::new(std::collections::HashMap::new()));

/// The follow lock for one community (created on first use).
pub(crate) fn follow_lock(id: &CommunityId) -> Arc<Mutex<()>> {
    V2_FOLLOW_LOCKS.lock().unwrap().entry(id.0).or_default().clone()
}

/// The author set the live v2 subscription CURRENTLY carries (sorted hex).
/// Diagnostics: comparing this against the freshly-derived plane authors is the
/// one-glance test for "did a rotation leave this client subscribed to a dead
/// epoch" — the failure that otherwise only shows up as silent missing events.
pub async fn subscribed_author_set() -> Vec<String> {
    V2_SUB_SET.lock().await.clone()
}

/// Diagnostics: run the three follow stages once for `id` UNDER the follow lock
/// (so it can't race the worker into a whole-row clobber) and return each
/// stage's outcome as a display string. Reloads freshly-persisted state between
/// stages, exactly like [`follow_community`]. Mutating — same writes a live
/// follow makes; the lock serializes it against the worker.
#[cfg(debug_assertions)]
pub async fn debug_run_follow_stages(id: &CommunityId, session: &std::sync::Arc<crate::db::Session>) -> (String, String, String) {
    let lock = follow_lock(id);
    let _guard = lock.lock().await;
    let transport = crate::community::transport::LiveTransport::with_timeout(std::time::Duration::from_secs(12));

    let Ok(Some(c)) = crate::db::community::load_community_v2(id) else {
        return ("community gone".into(), "-".into(), "-".into());
    };
    let rekeys = match super::service::follow_rekeys(&transport, &c, session).await {
        Ok(f) => format!(
            "Ok(updated={} self_removed={} dissolved={})",
            f.updated.as_ref().map(|u| format!("root_e{}", u.root_epoch.0)).unwrap_or_else(|| "no".into()),
            f.self_removed, f.dissolved
        ),
        Err(e) => format!("ERR: {e}"),
    };
    let Ok(Some(c)) = crate::db::community::load_community_v2(id) else {
        return (rekeys, "community gone".into(), "-".into());
    };
    let control = match super::service::follow_control(&transport, &c).await {
        Ok(v) => format!("Ok(changed={})", v.is_some()),
        Err(e) => format!("ERR: {e}"),
    };
    let Ok(Some(c)) = crate::db::community::load_community_v2(id) else {
        return (rekeys, control, "community gone".into());
    };
    let guestbook = match super::service::sync_guestbook(&transport, &c).await {
        Ok(fresh) => format!("Ok(fresh={})", fresh.len()),
        Err(e) => format!("ERR: {e}"),
    };
    (rekeys, control, guestbook)
}

pub async fn subscription_id() -> Option<SubscriptionId> {
    V2_SUB_ID.lock().await.clone()
}

pub async fn poolwide_subscription_id() -> Option<SubscriptionId> {
    V2_POOLWIDE_SUB_ID.lock().await.clone()
}

/// Clear the v2 realtime state on a session reset (called from `swap_session`
/// alongside v1's clear), so a stale sub id / author-set can't leak across accounts.
pub async fn clear() {
    *V2_SUB_ID.lock().await = None;
    *V2_POOLWIDE_SUB_ID.lock().await = None;
    V2_SUB_SET.lock().await.clear();
    V2_SEEN_WRAPS.lock().await.clear();
    // Drop the queue sender so the worker's channel closes and it exits; the
    // next login spawns a fresh worker.
    *V2_FOLLOW_TX.lock().unwrap() = None;
    V2_FOLLOW_PENDING.lock().unwrap().clear();
    V2_FOLLOW_LOCKS.lock().unwrap().clear();
    // Account A's stream keys must not keep authenticating (or answering relay
    // challenges) once account B is live.
    super::streamauth::clear();
}

/// Every plane pubkey a set of v2 communities publishes under that
/// [`inbound::dispatch_wrap`] handles — the subscription author-set. Per
/// community: the guestbook, the control plane, and each channel's current
/// Chat-Plane address. Pure + deterministic (deduped, sorted) — the testable core.
///
/// The **control plane** rides here so a long-running bot follows metadata +
/// public-channel edits live ([`super::service::follow_control`] re-folds on a
/// recognized wrap), and the next-epoch rekey planes ride via [`rekey_authors`]
/// (each subscribed author has its `dispatch_wrap` arm — never one without the
/// other).
pub fn plane_authors(communities: &[CommunityV2]) -> Vec<PublicKey> {
    let mut out = Vec::new();
    for c in communities {
        out.push(derive::guestbook_group_key(&c.community_root, c.id(), c.root_epoch).pk());
        out.push(control_author(c));
        // The dissolved plane (CORD-02 §9) — so a mid-session dissolution seals live.
        out.push(super::derive::dissolved_group_key(c.id()).pk());
        for ch in &c.channels {
            // A KEYLESS private channel has no readable chat plane — channel_secret's
            // root fallback would subscribe the PUBLIC plane for it. Its rekey plane
            // (below) is still watched, which is how its key arrives.
            if ch.private && ch.key.is_none() {
                continue;
            }
            let (secret, epoch) = c.channel_secret(ch);
            out.push(derive::channel_group_key(&secret, &ch.id, epoch).pk());
        }
        out.extend(rekey_authors(c));
    }
    out.sort_by_key(|p| p.to_hex());
    out.dedup();
    out
}

/// This community's Control Plane address at the current root epoch — the single
/// source of truth shared by [`plane_authors`] (subscribe) and
/// [`inbound::dispatch_wrap`] (recognize) so the two can't drift.
pub(crate) fn control_author(c: &CommunityV2) -> PublicKey {
    // A split epoch's address is the HELD control_pk (CORD-02 §5); a legacy
    // epoch derives it from the community_root as always.
    super::control::ControlPlane::of(c).pk()
}

/// The next-epoch rekey plane addresses for a community: the base rotation
/// (`root_epoch + 1`) and each Private channel's rotation (`channel epoch + 1`),
/// both under the CURRENT community_root. This is the single source of truth for
/// which rekey wraps we subscribe AND recognize (`inbound::dispatch_wrap` calls
/// it), so the two can never drift. A Public channel has no independent rotation
/// (it rides the base), so only Private channels contribute a channel address.
pub(crate) fn rekey_authors(c: &CommunityV2) -> Vec<PublicKey> {
    // saturating: a bundle's epoch isn't covered by the community_id commitment, so
    // it's attacker-influenced — never let `epoch + 1` overflow (a u64::MAX epoch
    // just yields a dead address, never a panic).
    let mut out = vec![derive::base_rekey_group_key(&c.community_root, c.id(), Epoch(c.root_epoch.0.saturating_add(1))).pk()];
    for ch in &c.channels {
        if ch.private {
            out.push(derive::channel_rekey_group_key(&c.community_root, &ch.id, Epoch(ch.epoch.0.saturating_add(1))).pk());
        }
    }
    out
}

/// Load every locally-held, LIVE **v2** community (dispatching each id by its
/// stored protocol). The realtime layer folds these into the subscription +
/// routing, so excluding a DISSOLVED community here is what enforces CORD-02 §9
/// on the receive side: its chat/control/rekey planes stop being subscribed and
/// an arriving wrap for it is `NotOurs` (dropped, never honored). Held keys still
/// open old history through the explicit read paths — this only stops NEW events.
pub fn load_held_v2() -> Vec<CommunityV2> {
    let ids = crate::db::community::list_community_ids().unwrap_or_default();
    ids.iter()
        .filter(|id| matches!(crate::db::community::community_protocol(id).ok().flatten(), Some(ConcordProtocol::V2)))
        .filter(|id| !crate::db::community::get_community_dissolved(&crate::simd::hex::bytes_to_hex_32(&id.0)).unwrap_or(false))
        .filter_map(|id| crate::db::community::load_community_v2(id).ok().flatten())
        .collect()
}

/// Refresh the v2 subscription for the held communities: register
/// `{kinds:[1059,21059], authors:[…]}` on their relays (targeted + pool-wide,
/// mirroring v1). Idempotent on an unchanged author-set.
pub async fn refresh_subscription(client: &Client) {
    // Phase 1, LOCK-FREE: make sure the community relays are added + connected —
    // the slow part (a connect wait of up to ~6s). Holding the sub locks across
    // this stalled every concurrent dispatch/refresh behind one caller's connect.
    {
        let communities = load_held_v2();
        let mut relays: Vec<String> = communities.iter().flat_map(|c| c.relays.iter().cloned()).collect();
        relays.sort();
        relays.dedup();
        if !relays.is_empty() {
            // Community relays ride GOSSIP|PING (warm but excluded from pool-wide DM ops).
            for r in &relays {
                let _ = client.add_managed_relay(r.as_str()).capabilities(crate::community_relay_capabilities()).await;
            }
            client.connect().await;
            // Wait briefly for at least one relay to actually connect (a subscribe
            // against a still-connecting relay silently fails to register — same
            // trap as v1).
            let wanted: Vec<RelayUrl> = relays.iter().filter_map(|r| RelayUrl::parse(r).ok()).collect();
            for _ in 0..24 {
                let pool = client.relays().all().await;
                if wanted.iter().any(|u| pool.get(u).map(|r| r.status() == RelayStatus::Connected).unwrap_or(false)) {
                    break;
                }
                tokio::time::sleep(std::time::Duration::from_millis(250)).await;
            }
            // Register every held plane's key BEFORE subscribing, so the responder
            // can answer any NIP-42 challenge our REQs trigger. Cheap + local.
            for c in &communities {
                super::streamauth::register_community(c);
            }
        }
    }

    // Phase 2, LOCKED + bounded (no connect waits): RE-snapshot the held state
    // INSIDE the sub locks — the follow worker runs concurrently and may have just
    // adopted a rotation, so the LAST locker must read the freshest persisted
    // authors. An out-of-lock read lets a stale caller commit an old author-set
    // over a fresh one and silently mute a rotated community. (A community whose
    // relays appeared between the phases subscribes now and connects on the next
    // refresh — the follow that discovers it always triggers one.)
    let mut sub_guard = V2_SUB_ID.lock().await;
    let mut set_guard = V2_SUB_SET.lock().await;

    let communities = load_held_v2();
    let authors = plane_authors(&communities);
    let mut relays: Vec<String> = communities.iter().flat_map(|c| c.relays.iter().cloned()).collect();
    relays.sort();
    relays.dedup();

    let mut new_set: Vec<String> = authors.iter().map(|p| p.to_hex()).collect();
    new_set.sort();

    // Unchanged-set fast path — but only when BOTH subs actually registered (a
    // failed pool-wide subscribe would otherwise stay absent until the author-set
    // changes, and Android streams via the pool-wide path).
    if sub_guard.is_some() && *set_guard == new_set && (authors.is_empty() || V2_POOLWIDE_SUB_ID.lock().await.is_some()) {
        mark_subscription_ready(); // an unchanged live sub is a completed pass
        return; // the pool re-applies the live subs across reconnects.
    }
    if let Some(old) = sub_guard.take() {
        let _ = client.unsubscribe(&old).await;
    }
    *set_guard = new_set;

    if authors.is_empty() {
        if let Some(old_pw) = V2_POOLWIDE_SUB_ID.lock().await.take() {
            let _ = client.unsubscribe(&old_pw).await;
        }
        // Completed with nothing to hear: READY with zero planes, which is a
        // different observable state from "still connecting".
        mark_subscription_ready();
        return;
    }

    let filter = Filter::new()
        .kinds([Kind::Custom(stream::KIND_WRAP), Kind::Custom(stream::KIND_WRAP_EPHEMERAL)])
        .authors(authors)
        // limit(0) suppresses the stored dump at REQ time, but a relay that
        // ACCEPTS a backfilled old event later pushes it to matching live subs
        // regardless — the `since` floor excludes those at the relay.
        .since(Timestamp::from_secs(
            Timestamp::now().as_secs().saturating_sub(crate::community::REALTIME_FRESH_WINDOW_MS / 1000),
        ))
        .limit(0);

    {
        let mut pw = V2_POOLWIDE_SUB_ID.lock().await;
        if let Some(old) = pw.take() {
            let _ = client.unsubscribe(&old).await;
        }
        if let Ok(out) = client.subscribe(filter.clone()).await {
            *pw = Some(out.value);
        }
    }
    if let Ok(out) = client
        .subscribe(nostr_sdk::prelude::ReqTarget::manual(
            relays.iter().cloned().map(|u| (u, vec![filter.clone()])),
        ))
        .await
    {
        *sub_guard = Some(out.value);
    }
    mark_subscription_ready();
    drop(set_guard);
    drop(sub_guard);

    // AUTH-gating relays serve the planes only to a stream-authenticated
    // connection, and a live subscription isn't auto-retried after the gate. The
    // priming used to run BEFORE the subscribe, which gated the whole registration
    // on the slowest relay: a set with dead or auth-refusing members left the
    // client deaf for a minute while healthy relays sat idle. Prime in the
    // BACKGROUND instead — healthy relays stream from the registration above, and
    // each gating relay's auth completion re-sends the subs (`resubscribe_relay`
    // via the responder), so it joins the moment IT is ready, gating nobody else.
    prime_auth_in_background(client, relays);
}

/// In-flight coalescer for the background auth prime: boot fires many refreshes
/// (dispatch, absorb, reconnect) and each spawning its own prime would stampede
/// the gating relays with duplicate auth fetches.
static PRIME_IN_FLIGHT: std::sync::atomic::AtomicBool = std::sync::atomic::AtomicBool::new(false);

fn prime_auth_in_background(client: &Client, relays: Vec<String>) {
    use std::sync::atomic::Ordering;
    if relays.is_empty() || PRIME_IN_FLIGHT.compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire).is_err() {
        return;
    }
    let client = client.clone();
    crate::db::spawn_bound(async move {
        super::streamauth::prime_auth(&client, &relays).await;
        PRIME_IN_FLIGHT.store(false, Ordering::Release);
    });
}

/// Re-send the CURRENT v2 subscriptions (same ids) to ONE relay. An AUTH-gating
/// relay CLOSEs a sub REQ that raced ahead of the stream AUTHs on a fresh
/// connection, and it never re-challenges once the connection is authenticated —
/// so the moment the streams finish authenticating is exactly when the subs must
/// be re-sent. nostr-sdk re-sends them only when ITS OWN auth completes (it
/// can't see ours), which usually wins by socket ordering; this makes the heal
/// deterministic and independent of that internal. Same-id REQs are idempotent.
pub(crate) async fn resubscribe_relay(client: &Client, relay: &RelayUrl) {
    let targeted = V2_SUB_ID.lock().await.clone();
    let poolwide = V2_POOLWIDE_SUB_ID.lock().await.clone();
    if targeted.is_none() && poolwide.is_none() {
        return; // nothing subscribed yet — the first refresh registers on an authed socket.
    }
    let communities = load_held_v2();
    let authors = plane_authors(&communities);
    if authors.is_empty() {
        return;
    }
    let filter = Filter::new()
        .kinds([Kind::Custom(stream::KIND_WRAP), Kind::Custom(stream::KIND_WRAP_EPHEMERAL)])
        .authors(authors)
        // limit(0) suppresses the stored dump at REQ time, but a relay that
        // ACCEPTS a backfilled old event later pushes it to matching live subs
        // regardless — the `since` floor excludes those at the relay.
        .since(Timestamp::from_secs(
            Timestamp::now().as_secs().saturating_sub(crate::community::REALTIME_FRESH_WINDOW_MS / 1000),
        ))
        .limit(0);
    for id in [targeted, poolwide].into_iter().flatten() {
        let _ = client
            .subscribe(nostr_sdk::prelude::ReqTarget::single(relay.clone(), [filter.clone()]))
            .with_id(id)
            .await;
    }
}

/// Route an arriving v2 wrap: find the held community whose plane it opens under
/// and fire the matching handler callback (via the shared bridge). Persistence to
/// the local DB is deferred (bots deliver via the callback; GUI history is v1 for
/// now). `session` gates against a mid-flight account swap.
pub async fn dispatch_event(event: Event, handler: Arc<dyn InboundEventHandler>) {
    crate::db::scoped(async move {
        let Some(my_pk) = crate::my_public_key() else {
            return;
        };
        // Fire EXACTLY ONCE per wrap: the pool re-delivers the same event under both
        // subs and from every relay. `insert` returns false if already dispatched.
        {
            let mut seen = V2_SEEN_WRAPS.lock().await;
            if !seen.insert(event.id.to_bytes()) {
                return;
            }
            if seen.len() > SEEN_WRAPS_CAP {
                let keep = event.id.to_bytes();
                seen.clear();
                seen.insert(keep);
            }
        }
        let communities = load_held_v2();
        for c in &communities {
            match inbound::dispatch_wrap(&event, c, &my_pk, &*handler) {
                inbound::DispatchedV2::NotOurs => continue,
                // A control OR a rekey wrap: just enqueue a follow for this community.
                // Non-blocking + coalesced — the single follow worker serializes control
                // and rekey per community (no concurrent whole-row clobber) off this hot
                // path, so a junk-wrap flood can't head-of-line-block the notification loop.
                inbound::DispatchedV2::Control { .. } | inbound::DispatchedV2::Rekey { .. } => {
                    enqueue_follow(c.id());
                    return;
                }
                inbound::DispatchedV2::Dissolved { community_id } => {
                    // Death wins (CORD-02 §9): seal read-only + surface the grave, ONCE
                    // (a re-wrapped tombstone with a fresh outer id must not re-fire the
                    // handler). The next load_held_v2 excludes it, so its planes also
                    // stop being subscribed + routed.
                    if crate::db::community::set_community_dissolved(&community_id).unwrap_or(false) {
                        handler.on_community_dissolved(&community_id);
                        if let Some(client) = crate::state::nostr_client() {
                            refresh_subscription(&client).await;
                        }
                    }
                    return;
                }
                // A chat event, opened but NOT yet applied: persist first (dedup by inner
                // id + the author-scoped edit/delete checks), then fire the callback from
                // the outcome — v1's exact model. A re-wrapped duplicate (any keyholder
                // can re-seal a signed rumor into a fresh 1059), the relay echo of our
                // own send, or a forged edit/delete yields no outcome and re-fires
                // nothing.
                inbound::DispatchedV2::Chat { channel_id, event } => {
                    match inbound::persist_chat_event(&event, &channel_id, &my_pk).await {
                        // "New" = first ingest, which a relay-backfill replay of a month-old
                        // event also satisfies — is_new comes from the inner send time so
                        // stale replays persist + surface without ringing anything.
                        Some(inbound::ChatPersist::New(message)) => {
                            let fresh = crate::community::is_realtime_fresh(message.at);
                            handler.on_community_message(&channel_id, &message, fresh);
                        }
                        // A reaction or an edit: the folded TARGET row (its id is the
                        // target's) — the same payload v1 hands this callback. Both come
                        // back out of STATE, whose compact form drops the quoted
                        // attachment's extension, so re-resolve before rendering.
                        Some(inbound::ChatPersist::Updated { mut message, .. }) => {
                            let _ = crate::db::events::populate_reply_context(&mut message).await;
                            handler.on_community_update(&channel_id, &message.id, &message);
                        }
                        // An un-react re-renders the PARENT (its chips changed) — the
                        // same surface a landed reaction drives.
                        Some(inbound::ChatPersist::ReactionRemoved { mut message, .. }) => {
                            let _ = crate::db::events::populate_reply_context(&mut message).await;
                            handler.on_community_update(&channel_id, &message.id, &message);
                        }
                        Some(inbound::ChatPersist::Removed(target_id)) => handler.on_community_removed(&channel_id, &target_id),
                        None => {}
                    }
                    return;
                }
                inbound::DispatchedV2::Presence { .. } => {
                    // Live membership motion: fold it into the persisted Guestbook so
                    // the memberlist stays a local read (the presence callback already
                    // fired inline). Reopen here — the dispatcher stays pure — and
                    // refresh the overview when it lands.
                    let gb = derive::guestbook_group_key(&c.community_root, c.id(), c.root_epoch);
                    if let Ok(opened) = super::stream::open_wrap(&event, &gb) {
                        if let Ok(ev) = super::guestbook::parse_guestbook_event(&opened) {
                            let changed = super::service::ingest_guestbook_event(c, ev, event.created_at.as_secs()).unwrap_or(false);
                            if changed {
                                handler.on_community_refreshed(&crate::simd::hex::bytes_to_hex_32(&c.id().0));
                            }
                        }
                    }
                    return;
                }
                inbound::DispatchedV2::Kick { target } => {
                    let gb = derive::guestbook_group_key(&c.community_root, c.id(), c.root_epoch);
                    let Ok(opened) = super::stream::open_wrap(&event, &gb) else { return };
                    let Ok(ev) = super::guestbook::parse_guestbook_event(&opened) else { return };
                    if !super::service::ingest_guestbook_event(c, ev, event.created_at.as_secs()).unwrap_or(false) {
                        return;
                    }
                    let community_id = crate::simd::hex::bytes_to_hex_32(&c.id().0);
                    // Self-removal rides the AUTHORIZED fold, never the wrap: the store takes
                    // any Kick rumor, so acting on arrival would let any member evict anyone.
                    // Fail-closed — a kick our roster can't justify leaves us in place.
                    //
                    // The COALESCE verdict, not the memberlist: on a rejoin the store starts
                    // empty while the control fold has already re-derived our old ban mark, so
                    // the memberlist legitimately excludes us for the window before the
                    // Guestbook catches up. A stale Kick landing in that window would read as
                    // an authorized eviction. Coalescing asks only what our latest authorized
                    // entry is, so a fresh Join beats the old Kick and an empty store decides
                    // nothing.
                    let evicted = my_pk == target && super::service::stored_kick_verdict(c, &my_pk);
                    if evicted {
                        // WARN, not info: `log_info!` is compiled out in release, and a
                        // self-eviction is exactly the destructive, hard-to-diagnose event
                        // that must leave a trace on a shipped build.
                        crate::log_warn!(
                            "[v2:teardown {}] KICK: the authorized guestbook fold rules us kicked",
                            &community_id[..8.min(community_id.len())]
                        );
                        handler.on_community_self_removed(&community_id);
                    } else {
                        crate::log_debug!(
                            "[v2:kick {}] declined: target={} is not ruled kicked by the fold",
                            &community_id[..8.min(community_id.len())], &target.to_hex()[..8]
                        );
                        handler.on_community_refreshed(&community_id);
                    }
                    return;
                }
                _ => return, // typing (and non-surfaced guestbook kinds) handled inline by the dispatcher.
            }
        }
    })
    .await
}

/// Whether a live follow worker is draining the queue (a `listen()` is running).
/// Headless callers use this to run a follow inline instead of enqueueing into
/// the void.
pub fn follow_worker_running() -> bool {
    V2_FOLLOW_TX.lock().unwrap().as_ref().map(|tx| !tx.is_closed()).unwrap_or(false)
}

/// Queue a follow for `id` — NON-BLOCKING + coalesced. A burst (or a junk-wrap
/// flood) collapses to at most one queued + one running follow per community. A
/// no-op if no worker is running (no live `listen()`). Callers: dispatch,
/// boot/reconnect catch-up, manual sync — none of them block or touch a lock for
/// longer than the enqueue.
pub fn enqueue_follow(id: &CommunityId) {
    let mut pending = V2_FOLLOW_PENDING.lock().unwrap();
    if !pending.insert(id.0) {
        return; // already queued or processing — coalesce.
    }
    match V2_FOLLOW_TX.lock().unwrap().as_ref() {
        Some(tx) if tx.send(*id).is_ok() => {}
        _ => {
            // No worker yet — PARK the id in the pending set instead of
            // dropping it; spawn_follow_worker drains parked ids into its
            // fresh queue. The boot sweep's enqueues race notifs' worker
            // spawn (they fire the moment init completes), so a pre-worker
            // enqueue must defer, never silently vanish — a dropped boot
            // refold leaves an offline-rotated community wedged at its old
            // epoch until some live event happens to trigger a dispatch.
        }
    }
}

/// Spawn the single follow worker for this session. Installs the queue sender and
/// drains it, running one combined follow per community at a time. Replacing the
/// sender (a re-`listen()`) or [`clear`] (a swap) closes the old channel so the old
/// worker exits; the captured `std::sync::Arc<crate::db::Session>` also stops it. Idempotent per session.
pub fn spawn_follow_worker(handler: Arc<dyn InboundEventHandler>) {
    let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<CommunityId>();
    // Re-send every pending id into the fresh queue: parked pre-worker
    // enqueues AND anything a replaced worker's dropped channel still owed.
    // Entries stay in the set — the worker removes each as it starts
    // processing, preserving the coalescing invariant.
    {
        let pending = V2_FOLLOW_PENDING.lock().unwrap();
        for id in pending.iter() {
            let _ = tx.send(CommunityId(*id));
        }
    }
    *V2_FOLLOW_TX.lock().unwrap() = Some(tx);
    let session = crate::db::current_session();
    crate::db::spawn_bound(async move {
        while let Some(id) = rx.recv().await {
            // Remove from pending BEFORE running, so a trigger arriving DURING the
            // follow re-enqueues (and is processed after) rather than being lost.
            V2_FOLLOW_PENDING.lock().unwrap().remove(&id.0);
            follow_community(&session, &id, &*handler).await;
        }
    });
}

/// One combined rekey-then-control follow for a community, each pass against the
/// FRESHLY-RELOADED persisted state (never a stale clone, so the two planes can't
/// lose each other's writes). Rekey runs first: a base adopt moves the control
/// address, and a self-removal tears the community down (skipping control). No-op
/// without a live client — unit tests drive `service::follow_control` /
/// `follow_rekeys` directly.
async fn follow_community(session: &std::sync::Arc<crate::db::Session>, id: &CommunityId, handler: &dyn InboundEventHandler) {
    crate::db::scoped(async move {
        let Some(client) = crate::state::nostr_client() else {
            return;
        };
        // Serialize against an inline (headless) follow of the same community — the
        // queue only serializes triggers routed THROUGH it.
        let lock = follow_lock(id);
        let _guard = lock.lock().await;
        let community_id = crate::simd::hex::bytes_to_hex_32(&id.0);
        let transport = crate::community::transport::LiveTransport::with_timeout(std::time::Duration::from_secs(12));

        // Rekey first (fresh DB state).
        let Ok(Some(current)) = crate::db::community::load_community_v2(id) else {
            return; // community gone (left / removed).
        };
        match super::service::follow_rekeys(&transport, &current, session).await {
            // A tombstone surfaced during catch-up (an offline member learning of a
            // death) — the flag is set; seal + surface, and stop following.
            Ok(follow) if follow.dissolved => {
                crate::log_warn!("[v2:teardown {}] DISSOLVED tombstone", &community_id[..8.min(community_id.len())]);
                handler.on_community_dissolved(&community_id);
                return;
            }
            Ok(follow) if follow.self_removed => {
                crate::log_warn!("[v2:teardown {}] REKEY EXCLUSION: an authorized rotation left us out", &community_id[..8.min(community_id.len())]);
                let _ = crate::db::community::delete_community(&community_id);
                refresh_subscription(&client).await;
                handler.on_community_self_removed(&community_id);
                return;
            }
            Ok(follow) if follow.updated.is_some() => {
                refresh_subscription(&client).await;
                handler.on_community_refreshed(&community_id);
            }
            Ok(_) => {}
            Err(e) => {
                // Surfaced, not swallowed: with EOSE-verified fetches an AUTH-gated
                // or dead rekey plane now reports here instead of masquerading as
                // "no rotation" — the exact signature of an epoch wedge.
                crate::log_warn!("[v2:follow {}] rekey follow failed (will retry on next trigger): {}", &community_id[..8.min(community_id.len())], e);
                return;
            }
        }

        // Control second, on the (possibly new-root) freshly-reloaded state.
        let Ok(Some(current)) = crate::db::community::load_community_v2(id) else {
            return;
        };
        let control_changed = matches!(
            super::service::follow_control(&transport, &current).await,
            Ok(Some(_))
        );
        // Re-judge parked key vends on EVERY pass, not only when the fold moved. A
        // vend that lands AFTER its Grant folded has nothing left to change the
        // control plane, so gating this on `changed` strands it until some unrelated
        // edition happens by (CORD-03 "delivered on grant").
        if let Ok(Some(folded)) = crate::db::community::load_community_v2(id) {
            let keyed = super::service::absorb_parked_channel_keys(&folded, session);
            if !keyed.is_empty() {
                crate::log_info!(
                    "[v2:follow {}] adopted {} vended private-channel key(s)",
                    &community_id[..8.min(community_id.len())],
                    keyed.len()
                );
                // A newly-keyed channel adds its chat plane to the readable set, and
                // the live subscription is addressed BY that set — without a rebuild
                // we hold the key and still never hear the channel. Independent of
                // `control_changed`: adopting a parked vend changes no control state.
                refresh_subscription(&client).await;
            }
            for ch in keyed {
                let hex = crate::simd::hex::bytes_to_hex_32(&ch.0);
                // Read what was said while we were locked out. The live subscription
                // only carries what arrives NEXT, so without this a member granted
                // access walks into a room that looks empty. The count rides the hook:
                // backfilled history is NOT dispatched as live messages, and without
                // the number a handler cannot even tell there is history to go read.
                let backfilled = crate::VectorCore::v2_backfill_channel(
                    id, &hex, 50, 2, None, None,
                    crate::community::transport::Evidence::Fast, 12,
                )
                .await;
                handler.on_channel_keyed(&community_id, &hex, backfilled);
            }
        }
        if control_changed {
            refresh_subscription(&client).await;
            handler.on_community_refreshed(&community_id);
            // A control change can reveal rekey work that predates it — a just-announced
            // private channel's key crate is already sitting on its rekey plane (the key
            // ships BEFORE the vsk-2), and this pass's rekey walk ran before the channel
            // existed. Queue one more pass; it coalesces and converges (an unchanged
            // control fold doesn't re-queue).
            enqueue_follow(id);
        }

        // Banned by the banlist we just folded → self-remove NOW, not when the Refounding
        // eventually lands. A Ban composes three layers (CORD-04 §6) and their guarantees
        // arrive at different speeds: the Banlist edition is instant, the Refounding is
        // "heavy and asynchronous" by design. Waiting on the rotation to notice our own
        // removal left the target sitting in a community that had already dropped them from
        // its member count — silenced, still looking joined. Checked unconditionally: a
        // banlist-only change folds no document, so `follow_control` returns None for it.
        // v1 has done this since it shipped (`am_i_banned` in its own realtime refresh).
        if let Some(me) = crate::my_public_key() {
            if crate::db::community::is_author_banned(&community_id, &me) {
                crate::log_warn!("[v2:teardown {}] SELF-BAN: our npub is in the folded banlist", &community_id[..8.min(community_id.len())]);
                handler.on_community_self_removed(&community_id);
                return;
            }
        }

        // Guestbook third: catch the membership store up from its cursor. Boot and
        // reconnect land here through this same queue, so the memberlist is a local
        // read by the time any panel asks — and every join/leave the catch-up folds
        // surfaces as a presence line (real-rumor-id keyed, so a line each path also
        // saw live inserts exactly once).
        let Ok(Some(current)) = crate::db::community::load_community_v2(id) else {
            return;
        };
        if let Ok(fresh) = super::service::sync_guestbook(&transport, &current).await {
            if fresh.is_empty() {
                return;
            }
            surface_presence(&current, &fresh, handler);
            handler.on_community_refreshed(&community_id);
        }
    })
    .await
}

/// Fire the presence-line surface for freshly-folded guestbook events — the
/// catch-up twin of the live dispatch (same handler, same real-id dedup key,
/// same banned-author drop). Kicks/snapshots shape the memberlist, not the feed.
fn surface_presence(
    community: &CommunityV2,
    fresh: &[super::guestbook::GuestbookEvent],
    handler: &dyn InboundEventHandler,
) {
    use super::guestbook::GuestbookEntry;
    use nostr_sdk::prelude::ToBech32;
    let Some(primary) = community.primary_channel() else {
        return;
    };
    let chat_id = crate::simd::hex::bytes_to_hex_32(&primary.id.0);
    let cid_hex = crate::simd::hex::bytes_to_hex_32(&community.id().0);
    let banned = crate::db::community::banned_set(&cid_hex);
    for ev in fresh {
        let (member, joined, at_ms, invited_by) = match &ev.entry {
            GuestbookEntry::Join { member, at_ms, invited_by } => (member, true, *at_ms, invited_by.clone()),
            GuestbookEntry::Leave { member, at_ms } => (member, false, *at_ms, None),
            GuestbookEntry::Kick { .. } | GuestbookEntry::Snapshot { .. } => continue,
        };
        if banned.contains(&member.to_bytes()) {
            continue;
        }
        let Ok(npub) = member.to_bech32();
        let event_id = crate::simd::hex::bytes_to_hex_32(&ev.rumor_id);
        let (by, label) = match &invited_by {
            Some((c, l)) => (Some(c.as_str()), Some(l.as_str())),
            None => (None, None),
        };
        handler.on_community_presence(&chat_id, &npub, joined, &event_id, at_ms / 1000, by, label);
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use super::super::control::{genesis, CommunityMetadata};
    use crate::community::Epoch;
    use nostr_sdk::prelude::Keys;

    fn a_community(name: &str) -> CommunityV2 {
        let owner = Keys::generate();
        let g = genesis(&owner, CommunityMetadata { name: name.into(), ..Default::default() }, 1_000).unwrap();
        CommunityV2::from_genesis(&g, name, None, vec!["wss://r".into()], 0)
    }

    #[test]
    fn subscription_readiness_is_scoped_to_the_session_that_marked_it() {
        // Generation-keyed on purpose: a bare bool would survive an account swap
        // and report account A's readiness for account B — the exact class of
        // per-account-global bug the swap-cache sweep exists to prevent.
        //
        // Bumping the GLOBAL generation invalidates every live std::sync::Arc<crate::db::Session>, so
        // serialize with the swap-sensitive tests via the same guard they hold.
        let _guard = crate::db::DB_TEST_GUARD.lock().unwrap_or_else(|e| e.into_inner());
        mark_subscription_ready();
        assert!(subscription_ready(), "marked for the current session");
        crate::db::close_database();
        assert!(!subscription_ready(), "an account swap invalidates it");
        mark_subscription_ready();
        assert!(subscription_ready(), "the new session's own pass re-arms it");
    }

    #[test]
    fn plane_authors_covers_the_dispatched_planes_only() {
        let c = a_community("A");
        let authors = plane_authors(std::slice::from_ref(&c));

        // Subscribed: the guestbook, the control plane, and the one public channel
        // (the planes dispatch_wrap handles). Exactly those three.
        let gb = derive::guestbook_group_key(&c.community_root, c.id(), c.root_epoch).pk();
        // The split address (the held control_pk), not the legacy derivation.
        let control = crate::community::v2::control::ControlPlane::of(&c).pk();
        assert_ne!(control, derive::control_group_key(&c.community_root, c.id(), c.root_epoch).pk());
        let general = {
            let (s, e) = c.channel_secret(&c.channels[0]);
            derive::channel_group_key(&s, &c.channels[0].id, e).pk()
        };
        // Plus the next base-rekey address (rekey-follow). A public channel has no
        // independent rotation, so #general contributes no channel-rekey address.
        let next_base = derive::base_rekey_group_key(&c.community_root, c.id(), Epoch(1)).pk();
        // Plus the dissolved plane (CORD-02 §9), so a live dissolution is detected.
        let dissolved = derive::dissolved_group_key(c.id()).pk();
        assert!(
            authors.contains(&gb)
                && authors.contains(&control)
                && authors.contains(&general)
                && authors.contains(&next_base)
                && authors.contains(&dissolved)
        );
        assert_eq!(authors.len(), 5, "guestbook + control + dissolved + chat + base-rekey planes are subscribed");
    }

    #[test]
    fn plane_authors_is_deterministic_deduped_and_multi_community() {
        let a = a_community("A");
        let b = a_community("B");
        let one = plane_authors(std::slice::from_ref(&a));
        // Re-running over the same community is byte-identical (deterministic).
        assert_eq!(plane_authors(std::slice::from_ref(&a)), one);
        // Two distinct communities' planes are all present, none dropped.
        let two = plane_authors(&[a.clone(), b.clone()]);
        assert_eq!(two.len(), one.len() * 2);
        // Order-independent: reversing the input yields the identical sorted set.
        assert_eq!(plane_authors(&[b, a]), two);
    }

    #[tokio::test]
    async fn dispatch_event_routes_a_v2_message_to_the_handler() {
        use crate::community::transport::memory::MemoryRelay;
        use crate::community::transport::{Query, Transport};
        use crate::types::Message;
        use std::sync::Mutex as StdMutex;

        #[derive(Default)]
        struct Recorder {
            got: StdMutex<Vec<(String, String)>>,
        }
        impl InboundEventHandler for Recorder {
            fn on_community_message(&self, chat_id: &str, msg: &Message, _new: bool) {
                self.got.lock().unwrap().push((chat_id.to_string(), msg.content.clone()));
            }
        }

        // Offline DB + identity.
        let _g = crate::db::DB_TEST_GUARD.lock().unwrap_or_else(|e| e.into_inner());
        crate::db::close_database();
        crate::db::clear_id_caches();
        let tmp = tempfile::tempdir().unwrap();
        let acct = {
            const B: &[u8] = b"qpzry9x8gf2tvdw0s3jn54khce6mua7l";
            let mut s = String::from("npub1");
            for i in 0..58 {
                s.push(B[(i * 5 + 1) % 32] as char);
            }
            s
        };
        std::fs::create_dir_all(tmp.path().join(&acct)).unwrap();
        crate::db::set_app_data_dir(crate::db::shared_test_data_dir().to_path_buf());
        crate::db::set_current_account(acct.clone()).unwrap();
        crate::db::init_database(&acct).unwrap();
        let _ = crate::state::take_nostr_client();
        let me = Keys::generate();
        crate::state::MY_SECRET_KEY.store_from_keys(&me, &[]);
        crate::state::set_my_public_key(me.public_key());

        // Create a v2 community (persisted), then ANOTHER member (holds the root)
        // posts — the incoming case a live sub delivers (an OWN send is echoed at
        // send time, so its relay copy correctly dedups instead of firing).
        let relay = MemoryRelay::new();
        let community = super::super::service::create_community(&relay, "Live", vec!["wss://r".into()], None).await.unwrap();
        let general = community.channels[0].id;
        let member = Keys::generate();
        let group = derive::channel_group_key(&community.community_root, &general, community.root_epoch);
        let rumor = super::super::chat::build_message_rumor(member.public_key(), &general, community.root_epoch, "live ping", None, &[], vec![], 5_000);
        let (wrap, _) = super::super::chat::seal_chat_rumor(&rumor, &group, &member, nostr_sdk::prelude::Timestamp::from_secs(5), false).unwrap();
        let _ = relay.publish(&wrap, &community.relays).await;
        let q = Query { kinds: vec![stream::KIND_WRAP], authors: vec![group.pk_hex()], ..Default::default() };
        let wrap = relay.fetch(&q, &community.relays).await.unwrap().into_iter().find(|w| w.pubkey == group.pk()).unwrap();

        // The realtime dispatch (loading held v2 communities from the DB) routes it.
        // Dispatch the SAME wrap TWICE — modelling the pool re-delivering it under
        // the targeted + pool-wide subs (and from multiple relays). The handler
        // must fire EXACTLY ONCE (no duplicate bot replies) — and only AFTER the
        // persist outcome (the callbacks-from-persist model).
        let rec = Arc::new(Recorder::default());
        crate::community::v2::realtime::clear().await; // fresh seen-set for the test
        dispatch_event(wrap.clone(), rec.clone()).await;
        dispatch_event(wrap, rec.clone()).await;

        let got = rec.got.lock().unwrap();
        assert_eq!(got.len(), 1, "a re-delivered wrap fires the handler exactly once");
        assert_eq!(got[0].1, "live ping");
        assert_eq!(got[0].0, crate::simd::hex::bytes_to_hex_32(&general.0));
    }

    #[tokio::test]
    async fn follow_queue_coalesces_a_burst_and_re_enqueues_after_processing() {
        // Install a test channel in place of the worker's, so we can observe what the
        // queue delivers without a live client.
        let (tx, mut rx) = tokio::sync::mpsc::unbounded_channel::<CommunityId>();
        *V2_FOLLOW_TX.lock().unwrap() = Some(tx);
        V2_FOLLOW_PENDING.lock().unwrap().clear();

        let id = CommunityId([0x11; 32]);
        // A burst for one community collapses to a SINGLE queued follow (coalesced).
        enqueue_follow(&id);
        enqueue_follow(&id);
        enqueue_follow(&id);
        assert_eq!(rx.recv().await, Some(id), "first trigger queues a follow");
        assert!(rx.try_recv().is_err(), "the burst coalesced to exactly one");

        // The worker removes it from pending before running; a trigger AFTER that
        // re-queues (so a change during a follow isn't lost).
        V2_FOLLOW_PENDING.lock().unwrap().remove(&id.0);
        enqueue_follow(&id);
        assert_eq!(rx.recv().await, Some(id), "a trigger after processing re-queues");

        // A different community is independent (not coalesced against the first).
        let id2 = CommunityId([0x22; 32]);
        enqueue_follow(&id2);
        assert_eq!(rx.recv().await, Some(id2));

        *V2_FOLLOW_TX.lock().unwrap() = None;
        V2_FOLLOW_PENDING.lock().unwrap().clear();
    }

    #[tokio::test]
    async fn a_dissolved_community_honors_no_new_events_and_fires_death_once() {
        use super::super::service;
        use crate::community::transport::memory::MemoryRelay;
        use crate::community::transport::Transport;
        use crate::types::Message;
        use std::sync::Mutex as StdMutex;

        #[derive(Default)]
        struct Recorder {
            messages: StdMutex<Vec<String>>,
            deaths: StdMutex<Vec<String>>,
        }
        impl InboundEventHandler for Recorder {
            fn on_community_message(&self, _chat: &str, msg: &Message, _new: bool) {
                self.messages.lock().unwrap().push(msg.content.clone());
            }
            fn on_community_dissolved(&self, community_id: &str) {
                self.deaths.lock().unwrap().push(community_id.to_string());
            }
        }

        let _g = crate::db::DB_TEST_GUARD.lock().unwrap_or_else(|e| e.into_inner());
        crate::db::close_database();
        crate::db::clear_id_caches();
        let tmp = tempfile::tempdir().unwrap();
        let acct = {
            const B: &[u8] = b"qpzry9x8gf2tvdw0s3jn54khce6mua7l";
            let mut s = String::from("npub1");
            for i in 0..58 {
                s.push(B[(i * 3 + 2) % 32] as char);
            }
            s
        };
        std::fs::create_dir_all(tmp.path().join(&acct)).unwrap();
        crate::db::set_app_data_dir(crate::db::shared_test_data_dir().to_path_buf());
        crate::db::set_current_account(acct.clone()).unwrap();
        crate::db::init_database(&acct).unwrap();
        let _ = crate::state::take_nostr_client();
        let me = Keys::generate();
        crate::state::MY_SECRET_KEY.store_from_keys(&me, &[]);
        crate::state::set_my_public_key(me.public_key());

        let relay = MemoryRelay::new();
        let community = service::create_community(&relay, "Doomed", vec!["wss://r".into()], None).await.unwrap();
        let general = community.channels[0].id;

        // The owner's tombstone arrives as a MEMBER sees it (local flag still 0 —
        // build + publish it directly rather than via dissolve_community, which
        // would seal our own DB first and make it the owner-published case). `me`
        // is the owner, so the seal verifies.
        let rumor = super::super::dissolution::dissolved_tombstone_rumor(me.public_key(), community.id(), 8_000);
        let tombstone = super::super::dissolution::seal_dissolved(&rumor, community.id(), &me, nostr_sdk::prelude::Timestamp::from_secs(8_000)).unwrap();
        let _ = relay.publish(&tombstone, &community.relays).await;
        assert!(!crate::db::community::get_community_dissolved(&crate::simd::hex::bytes_to_hex_32(&community.id().0)).unwrap(), "not yet locally sealed");

        let rec = Arc::new(Recorder::default());
        clear().await;
        // Fire the SAME tombstone twice AND a fresh re-wrap of its verified seal
        // (distinct outer id) — death must be announced exactly once.
        dispatch_event(tombstone.clone(), rec.clone()).await;
        dispatch_event(tombstone, rec.clone()).await;
        assert!(crate::db::community::get_community_dissolved(&crate::simd::hex::bytes_to_hex_32(&community.id().0)).unwrap());

        // A member posts a fresh message to #general AFTER the tombstone. It must
        // not be honored (the community is excluded from load_held_v2, so its
        // plane is NotOurs).
        let member = Keys::generate();
        let cgroup = derive::channel_group_key(&community.community_root, &general, community.root_epoch);
        let rumor = super::super::chat::build_message_rumor(member.public_key(), &general, community.root_epoch, "into the grave", None, &[], vec![], 9_000);
        let (mw, _) = super::super::chat::seal_chat_rumor(&rumor, &cgroup, &member, nostr_sdk::prelude::Timestamp::from_secs(9), false).unwrap();
        dispatch_event(mw, rec.clone()).await;

        assert_eq!(rec.deaths.lock().unwrap().len(), 1, "death is announced exactly once");
        assert!(rec.messages.lock().unwrap().is_empty(), "a post-tombstone message is never honored (CORD-02 §9)");
    }

    #[test]
    fn a_private_channel_subscribes_to_its_own_chat_plane() {
        let mut c = a_community("Priv");
        c.channels.push(super::super::community::ChannelV2 {
            id: crate::community::ChannelId([0x33; 32]),
            name: "mods".into(),
            private: true,
            key: Some([0x44; 32]),
            epoch: Epoch(1),
            voice: None,
            meta_custom: None,
            meta_extra: Default::default(),
        });
        let authors = plane_authors(std::slice::from_ref(&c));
        // A private channel is read under its OWN key/epoch (not the root).
        let priv_chat = derive::channel_group_key(&[0x44; 32], &c.channels[1].id, Epoch(1)).pk();
        assert!(authors.contains(&priv_chat), "a private channel subscribes to its own chat plane");
        // Its next-rekey address IS subscribed (rekey-follow), keyed by the current
        // root at the channel's next epoch — so a rotation is delivered.
        let next_rekey = derive::channel_rekey_group_key(&c.community_root, &c.channels[1].id, Epoch(2)).pk();
        assert!(authors.contains(&next_rekey), "a private channel's next rekey plane is subscribed");
    }
}