tidaldb/tidal-server/tests/cluster_region.rs

//! m8p10 in-process multi-process-cluster tests.
//!
//! Each test builds TWO `RegionClusterState`s in ONE test process — distinct
//! topologies pointing at each other's REAL loopback gRPC addresses — and drives
//! them over real HTTP (axum on loopback) + real `GrpcTransport` replication.
//! Unlike `cluster_e2e.rs` (tier-3, spawns OS processes), these run in the
//! default test build with no OS processes, exactly like `cluster_grpc.rs`.
//!
//! They prove the multi-process region node: convergence over real loopback
//! gRPC with decay parity, the typed `NotLeader` rejection, partition→heal with
//! idempotent re-heal, and that promote flips roles with the ALWAYS-ON receiver
//! (the demoted node applies the new leader's ships).
#![allow(
    clippy::unwrap_used,
    clippy::missing_panics_doc,
    clippy::too_many_lines,
    clippy::doc_markdown
)]

use std::{
    net::{SocketAddr, TcpListener},
    sync::Arc,
    time::{Duration, Instant},
};

use tidal_server::cluster::{
    RegionClusterState, RegionSpec, ReplicationSpec, TimeoutsSpec, TopologySpec, WalSpec,
    build_region_router,
};
use tidaldb::schema::{DecaySpec, EntityKind, Schema, SchemaBuilder, Window};

/// A single-`view`-signal schema with a `hide` hard-negative signal (so the
/// `/hardnegs` route's `signal_with_context("hide", …)` resolves).
fn region_schema() -> Schema {
    let mut builder = SchemaBuilder::new();
    let _ = builder
        .signal(
            "view",
            EntityKind::Item,
            DecaySpec::Exponential {
                half_life: Duration::from_secs(7 * 24 * 3600),
            },
        )
        .windows(&[Window::OneHour])
        .velocity(false)
        .add();
    let _ = builder
        .signal("hide", EntityKind::Item, DecaySpec::Permanent)
        .velocity(false)
        .add();
    builder.build().unwrap()
}

/// Reserve a free loopback port and return its address.
fn free_addr() -> SocketAddr {
    TcpListener::bind("127.0.0.1:0")
        .unwrap()
        .local_addr()
        .unwrap()
}

/// A pair of fully-declared (grpc + http) region specs that point at each
/// other. `leader` is the first region's name.
struct Pair {
    leader_name: String,
    follower_name: String,
    leader_grpc: SocketAddr,
    follower_grpc: SocketAddr,
    leader_http: SocketAddr,
    follower_http: SocketAddr,
}

impl Pair {
    fn new() -> Self {
        Self {
            leader_name: "us-east".into(),
            follower_name: "eu-west".into(),
            leader_grpc: free_addr(),
            follower_grpc: free_addr(),
            leader_http: free_addr(),
            follower_http: free_addr(),
        }
    }

    /// The shared topology both processes parse (same declaration order ⇒ same
    /// RegionIds in both). HTTP addrs are the in-test axum binds.
    fn topology(&self) -> TopologySpec {
        TopologySpec {
            regions: vec![
                RegionSpec {
                    name: self.leader_name.clone(),
                    grpc_addr: Some(self.leader_grpc.to_string()),
                    http_addr: Some(self.leader_http.to_string()),
                    grpc_tls: None,
                    metrics_addr: None,
                },
                RegionSpec {
                    name: self.follower_name.clone(),
                    grpc_addr: Some(self.follower_grpc.to_string()),
                    http_addr: Some(self.follower_http.to_string()),
                    grpc_tls: None,
                    metrics_addr: None,
                },
            ],
            leader: self.leader_name.clone(),
            write_workers: None,
            timeouts: TimeoutsSpec::default(),
            replication: ReplicationSpec::default(),
            wal: WalSpec::default(),
        }
    }
}

/// A node's persistent data dir (m11p2: the durable WAL is the replication
/// stream, so multi-process cluster mode requires one).
///
/// DECLARE THE DIR BEFORE ANYTHING THAT CAN HOLD THE NODE (including the
/// tokio runtime): locals — and bindings within one tuple pattern — drop in
/// reverse declaration order, and a dir deleted while its node still runs
/// wedges fjall's flush worker on `NotFound` (the sealed memtable then never
/// drains and `rotate_memtable_and_wait` polls forever — observed as a
/// permanently hung test on the panic-unwind path).
fn region_dir() -> tempfile::TempDir {
    tempfile::tempdir().expect("create per-region data dir")
}

/// Build one region node off the reactor (GrpcTransport::new blocks on its own
/// runtime, so it must run on a plain thread). `dir` is the node's data dir;
/// see [`region_dir`] for the declaration-order contract.
fn build_region(
    topology: TopologySpec,
    region: &str,
    dir: &tempfile::TempDir,
) -> RegionClusterState {
    let region = region.to_string();
    let data_dir = dir.path().to_path_buf();
    std::thread::spawn(move || {
        RegionClusterState::new(
            &topology,
            &region,
            region_schema(),
            Vec::new(),
            Some(data_dir),
            0,
        )
    })
    .join()
    .unwrap()
    .expect("region node builds with real gRPC transport")
}

/// Serve `router` on `addr` using `rt`; returns once the listener is bound.
fn serve(rt: &tokio::runtime::Runtime, router: axum::Router, addr: SocketAddr) {
    let listener = rt
        .block_on(tokio::net::TcpListener::bind(addr))
        .unwrap_or_else(|e| panic!("bind {addr}: {e}"));
    rt.spawn(async move {
        let _ = axum::serve(listener, router).await;
    });
}

/// Poll `GET /cluster/status/local` on `base` until `pred(applied, lag)` holds
/// or the deadline elapses.
fn poll_status(
    client: &reqwest::blocking::Client,
    base: &str,
    pred: impl Fn(u64, u64) -> bool,
) -> serde_json::Value {
    let deadline = Instant::now() + Duration::from_secs(5);
    loop {
        let status: serde_json::Value = client
            .get(format!("{base}/cluster/status/local"))
            .send()
            .unwrap()
            .json()
            .unwrap();
        let applied = status["applied_events"].as_u64().unwrap_or(0);
        let lag = status["lag_events"].as_u64().unwrap_or(u64::MAX);
        if pred(applied, lag) {
            return status;
        }
        assert!(
            Instant::now() <= deadline,
            "status predicate not met within 5s: {status}"
        );
        std::thread::sleep(Duration::from_millis(20));
    }
}

/// Read entity `entity`'s trending feed score on `base` (0.0 if absent).
fn feed_score(client: &reqwest::blocking::Client, base: &str, entity: u64) -> f64 {
    let feed: serde_json::Value = client
        .get(format!("{base}/feed?profile=trending&limit=10"))
        .send()
        .unwrap()
        .json()
        .unwrap();
    feed["items"]
        .as_array()
        .unwrap()
        .iter()
        .find(|it| it["entity_id"].as_u64() == Some(entity))
        .and_then(|it| it["score"].as_f64())
        .unwrap_or(0.0)
}

/// Two region nodes converge over real loopback gRPC; the follower's feed scores
/// match the leader's to 1e-6 (decay parity).
#[test]
fn region_node_replicates_over_grpc() {
    let pair = Pair::new();

    let leader_dir = region_dir();
    let leader = build_region(pair.topology(), &pair.leader_name, &leader_dir);
    let follower_dir = region_dir();
    let follower = build_region(pair.topology(), &pair.follower_name, &follower_dir);

    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(leader), None),
        pair.leader_http,
    );
    serve(
        &rt,
        build_region_router(Arc::new(follower), None),
        pair.follower_http,
    );

    let client = reqwest::blocking::Client::new();
    let leader_base = format!("http://{}", pair.leader_http);
    let follower_base = format!("http://{}", pair.follower_http);

    // Broadcast items to BOTH nodes (items are not WAL-replicated in this task),
    // then write signals on the leader (replicated to the follower over gRPC).
    for i in 1..=8u64 {
        for base in [&leader_base, &follower_base] {
            let resp = client
                .post(format!("{base}/items"))
                .json(&serde_json::json!({
                    "entity_id": i,
                    "metadata": { "title": format!("item {i}") }
                }))
                .send()
                .unwrap();
            assert!(resp.status().is_success(), "POST /items: {}", resp.status());
        }
        let resp = client
            .post(format!("{leader_base}/signals"))
            .json(&serde_json::json!({ "entity_id": i, "signal": "view", "weight": 1.0 }))
            .send()
            .unwrap();
        assert!(
            resp.status().is_success(),
            "POST /signals on leader: {}",
            resp.status()
        );
    }

    // Follower converges: applied reaches 8 and lag returns to 0.
    poll_status(&client, &follower_base, |applied, lag| {
        applied >= 8 && lag == 0
    });

    // Decay parity: the leader and follower feeds rank the same items with
    // scores equal to 1e-6 (the follower replayed the exact same WAL events).
    let leader_feed: serde_json::Value = client
        .get(format!("{leader_base}/feed?profile=trending&limit=8"))
        .send()
        .unwrap()
        .json()
        .unwrap();
    let follower_feed: serde_json::Value = client
        .get(format!("{follower_base}/feed?profile=trending&limit=8"))
        .send()
        .unwrap()
        .json()
        .unwrap();
    let l_items = leader_feed["items"].as_array().unwrap();
    let f_items = follower_feed["items"].as_array().unwrap();
    assert!(!f_items.is_empty(), "follower must serve replicated items");
    assert_eq!(l_items.len(), f_items.len(), "same number of ranked items");

    let mut l_scores: std::collections::HashMap<u64, f64> = std::collections::HashMap::new();
    for it in l_items {
        l_scores.insert(
            it["entity_id"].as_u64().unwrap(),
            it["score"].as_f64().unwrap(),
        );
    }
    for it in f_items {
        let id = it["entity_id"].as_u64().unwrap();
        let f = it["score"].as_f64().unwrap();
        let l = *l_scores.get(&id).expect("follower item also on leader");
        assert!((l - f).abs() < 1e-6, "entity {id}: leader={l} follower={f}");
    }

    rt.shutdown_timeout(Duration::from_secs(2));
}

/// A write to the FOLLOWER (a non-leader node) is FORWARDED to the leader; with
/// the leader process not running, the forward fails and degrades to a 503 whose
/// body names the (unreachable) leader — the task-03 leader-unreachable contract.
#[test]
fn region_node_rejects_writes_when_not_leader() {
    let pair = Pair::new();

    let follower_dir = region_dir();
    let follower = build_region(pair.topology(), &pair.follower_name, &follower_dir);
    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(follower), None),
        pair.follower_http,
    );

    let client = reqwest::blocking::Client::new();
    let follower_base = format!("http://{}", pair.follower_http);

    let resp = client
        .post(format!("{follower_base}/signals"))
        .json(&serde_json::json!({ "entity_id": 1, "signal": "view", "weight": 1.0 }))
        .send()
        .unwrap();
    assert_eq!(
        resp.status().as_u16(),
        503,
        "a non-leader write must be 503 NotLeader"
    );
    let body: serde_json::Value = resp.json().unwrap();
    assert_eq!(
        body["leader"].as_str(),
        Some(pair.leader_name.as_str()),
        "the 503 body must name the leader: {body}"
    );

    rt.shutdown_timeout(Duration::from_secs(2));
}

/// Partition the follower → leader ships are skipped (follower lags) → heal →
/// the leader redelivers over gRPC and the follower converges. A SECOND heal is
/// a no-op (idempotent): the follower's scores are unchanged.
#[test]
fn region_node_partition_heal() {
    let pair = Pair::new();

    let leader_dir = region_dir();
    let leader = build_region(pair.topology(), &pair.leader_name, &leader_dir);
    let follower_dir = region_dir();
    let follower = build_region(pair.topology(), &pair.follower_name, &follower_dir);
    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(leader), None),
        pair.leader_http,
    );
    serve(
        &rt,
        build_region_router(Arc::new(follower), None),
        pair.follower_http,
    );

    let client = reqwest::blocking::Client::new();
    let leader_base = format!("http://{}", pair.leader_http);
    let follower_base = format!("http://{}", pair.follower_http);

    let post_signal = |entity: u64| {
        let resp = client
            .post(format!("{leader_base}/signals"))
            .json(&serde_json::json!({ "entity_id": entity, "signal": "view", "weight": 1.0 }))
            .send()
            .unwrap();
        assert!(
            resp.status().is_success(),
            "leader signal: {}",
            resp.status()
        );
    };
    let post_item = |entity: u64| {
        let resp = client
            .post(format!("{leader_base}/items"))
            .json(&serde_json::json!({ "entity_id": entity, "metadata": { "t": entity.to_string() } }))
            .send()
            .unwrap();
        assert!(resp.status().is_success(), "leader item: {}", resp.status());
    };

    // Create items 1..=5 on the LEADER up front, so heal can backfill the ones the
    // follower misses while partitioned (BUG 3). Write 2 signals, let the follower
    // catch up.
    for e in 1..=5u64 {
        post_item(e);
    }
    post_signal(1);
    post_signal(2);
    // m11p2: items ride the WAL too — 5 item records + 2 signals = 7 seqnos.
    poll_status(&client, &follower_base, |applied, _| applied >= 7);

    // Partition the follower from the leader, then write 3 more.
    let resp = client
        .post(format!("{leader_base}/cluster/partition"))
        .json(&serde_json::json!({ "region": pair.follower_name }))
        .send()
        .unwrap();
    assert!(resp.status().is_success(), "partition: {}", resp.status());
    for e in 3..=5u64 {
        post_signal(e);
    }
    // While partitioned the follower is STUCK at applied=7: the leader's
    // ships were skipped, so no segments arrive. (Its locally-reported lag stays
    // 0 because a follower cannot observe the leader's progress while
    // partitioned — cross-node lag aggregation is task 03; the load-bearing
    // proof here is that `applied` does NOT advance past 2.)
    let lagging = poll_status(&client, &follower_base, |applied, _| applied == 7);
    assert_eq!(lagging["applied_events"].as_u64(), Some(7));
    // Give the leader a beat to (not) ship — applied must remain 7, proving the
    // partition truly skips the eager ships rather than racing convergence.
    std::thread::sleep(Duration::from_millis(200));
    let still: serde_json::Value = client
        .get(format!("{follower_base}/cluster/status/local"))
        .send()
        .unwrap()
        .json()
        .unwrap();
    assert_eq!(
        still["applied_events"].as_u64(),
        Some(7),
        "partitioned follower must NOT receive the leader's post-partition writes"
    );

    // Heal: the leader redelivers the missed segments over gRPC.
    let heal = |base: &str| {
        let resp = client
            .post(format!("{base}/cluster/heal"))
            .json(&serde_json::json!({ "region": pair.follower_name }))
            .send()
            .unwrap();
        assert!(resp.status().is_success(), "heal: {}", resp.status());
    };
    heal(&leader_base);
    poll_status(&client, &follower_base, |applied, lag| {
        applied >= 10 && lag == 0
    });

    // Capture the follower's converged decay for entity 5. Heal redelivered the
    // signal AND backfilled the item (BUG 3), so the follower can rank entity 5
    // WITHOUT the test posting anything. Poll until the score stabilizes (two equal
    // consecutive non-zero reads) so the capture does not race the async item
    // index; after it settles the score is stable, so re-heal idempotence is exact.
    let score_of = |entity: u64| -> f64 {
        let deadline = Instant::now() + Duration::from_secs(10);
        let mut last = -1.0f64;
        loop {
            let s = feed_score(&client, &follower_base, entity);
            if s > 0.0 && (s - last).abs() < 1e-12 {
                return s;
            }
            last = s;
            assert!(
                Instant::now() <= deadline,
                "heal must deliver signal+item for entity {entity} so the follower ranks it"
            );
            std::thread::sleep(Duration::from_millis(50));
        }
    };
    let before = score_of(5);

    // Second heal is idempotent: re-ships nothing new (follower already applied
    // through seq 10) and re-broadcasts the same items, so the score is unchanged.
    heal(&leader_base);
    poll_status(&client, &follower_base, |applied, lag| {
        applied == 10 && lag == 0
    });
    let after = score_of(5);
    assert!(
        (before - after).abs() < 1e-9,
        "idempotent re-heal must not change scores: before={before} after={after}"
    );

    rt.shutdown_timeout(Duration::from_secs(2));
}

/// Promote flips the leadership view: the OLD leader now rejects writes
/// (NotLeader), the NEW leader accepts and ships, and the always-on receiver on
/// the demoted node applies the new leader's segments.
#[test]
fn region_node_promote_local() {
    let pair = Pair::new();

    let leader_dir = region_dir();
    let leader = build_region(pair.topology(), &pair.leader_name, &leader_dir);
    let follower_dir = region_dir();
    let follower = build_region(pair.topology(), &pair.follower_name, &follower_dir);
    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(leader), None),
        pair.leader_http,
    );
    serve(
        &rt,
        build_region_router(Arc::new(follower), None),
        pair.follower_http,
    );

    let client = reqwest::blocking::Client::new();
    let leader_base = format!("http://{}", pair.leader_http);
    let follower_base = format!("http://{}", pair.follower_http);

    // Promote the FOLLOWER to leader on BOTH nodes (each node holds its own view).
    for base in [&leader_base, &follower_base] {
        let resp = client
            .post(format!("{base}/cluster/promote"))
            .json(&serde_json::json!({ "region": pair.follower_name }))
            .send()
            .unwrap();
        assert!(resp.status().is_success(), "promote: {}", resp.status());
    }

    // The OLD leader (us-east), now demoted, transparently FORWARDS a write to the
    // new leader (eu-west) and relays its 204 (task 03 replaces the standalone
    // NotLeader 503 with leader forwarding). Item 1 must exist on eu-west to be
    // rankable, but the forwarded signal itself proves the forward path.
    let _ = client
        .post(format!("{follower_base}/items"))
        .json(&serde_json::json!({ "entity_id": 1, "metadata": {} }))
        .send()
        .unwrap();
    let resp = client
        .post(format!("{leader_base}/signals"))
        .json(&serde_json::json!({ "entity_id": 1, "signal": "view", "weight": 1.0 }))
        .send()
        .unwrap();
    assert_eq!(
        resp.status().as_u16(),
        204,
        "demoted node must FORWARD the write to the new leader (204), not 503: {}",
        resp.status()
    );

    // The NEW leader (eu-west) accepts and ships to the demoted node (us-east),
    // whose ALWAYS-ON receiver applies the segments. Broadcast items to both.
    for e in 1..=4u64 {
        for base in [&leader_base, &follower_base] {
            let _ = client
                .post(format!("{base}/items"))
                .json(&serde_json::json!({ "entity_id": e, "metadata": {} }))
                .send()
                .unwrap();
        }
        let resp = client
            .post(format!("{follower_base}/signals"))
            .json(&serde_json::json!({ "entity_id": e, "signal": "view", "weight": 1.0 }))
            .send()
            .unwrap();
        assert!(
            resp.status().is_success(),
            "new-leader signal: {}",
            resp.status()
        );
    }

    // The demoted node (us-east) applies the new leader's stream — proving the
    // always-on receiver runs on every node, not just the initial follower.
    poll_status(&client, &leader_base, |applied, lag| {
        applied >= 4 && lag == 0
    });

    rt.shutdown_timeout(Duration::from_secs(2));
}

/// `POST /hardnegs` records a hide on the local node (node-local by design).
#[test]
fn region_node_records_hardneg() {
    let pair = Pair::new();

    let leader_dir = region_dir();
    let leader = build_region(pair.topology(), &pair.leader_name, &leader_dir);
    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(leader), None),
        pair.leader_http,
    );

    let client = reqwest::blocking::Client::new();
    let leader_base = format!("http://{}", pair.leader_http);

    let resp = client
        .post(format!("{leader_base}/hardnegs"))
        .json(&serde_json::json!({ "user_id": 42, "item_id": 7 }))
        .send()
        .unwrap();
    assert_eq!(
        resp.status().as_u16(),
        204,
        "POST /hardnegs must record a hide: {}",
        resp.status()
    );

    rt.shutdown_timeout(Duration::from_secs(2));
}

// ── BUG 1: lag gauge across a leadership change ──────────────────────────────

/// A fully-declared THREE-region topology pointing at three loopback addresses.
struct Trio {
    names: [String; 3],
    grpc: [SocketAddr; 3],
    http: [SocketAddr; 3],
}

impl Trio {
    fn new() -> Self {
        Self {
            names: ["us-east".into(), "eu-west".into(), "ap-south".into()],
            grpc: [free_addr(), free_addr(), free_addr()],
            http: [free_addr(), free_addr(), free_addr()],
        }
    }

    fn topology(&self) -> TopologySpec {
        TopologySpec {
            regions: (0..3)
                .map(|i| RegionSpec {
                    name: self.names[i].clone(),
                    grpc_addr: Some(self.grpc[i].to_string()),
                    http_addr: Some(self.http[i].to_string()),
                    grpc_tls: None,
                    metrics_addr: None,
                })
                .collect(),
            leader: self.names[0].clone(),
            write_workers: None,
            timeouts: TimeoutsSpec::default(),
            replication: ReplicationSpec::default(),
            wal: WalSpec::default(),
        }
    }
}

/// BUG 3 reproduction: items are HTTP-broadcast (not WAL-relayed), so a follower
/// that is partitioned during the item broadcast misses the items forever — and
/// heal (which only re-ships signal segments) must ALSO backfill the item
/// metadata + embeddings, so heal is the single recovery verb that leaves the
/// follower with EXACTLY the leader's data.
///
/// Before the fix, the leader's `/items` broadcast to a partitioned follower
/// landed in the `failed` list and nothing backfilled it; the follower's feed
/// could not rank items it never received, even after heal closed the signal gap.
#[test]
fn region_node_heal_backfills_missed_items() {
    let pair = Pair::new();

    let leader_dir = region_dir();
    let leader = build_region(pair.topology(), &pair.leader_name, &leader_dir);
    let follower_node_dir = region_dir();
    let follower = Arc::new(build_region(
        pair.topology(),
        &pair.follower_name,
        &follower_node_dir,
    ));
    // Build the runtime AFTER the nodes (and their data-dir guards):
    // locals drop in reverse order, so the runtime — which owns the
    // nodes via the serve tasks — tears down BEFORE the dirs delete,
    // on the panic-unwind path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(2)
        .enable_all()
        .build()
        .unwrap();
    serve(
        &rt,
        build_region_router(Arc::new(leader), None),
        pair.leader_http,
    );
    // The follower's HTTP server is NOT serving yet — it models a region that is
    // DOWN during the leader's item/signal broadcast. The leader's best-effort
    // HTTP broadcast to it will fail (connection refused) and land in `failed`,
    // exactly as it would for a crashed/restarting region. The follower's gRPC
    // receiver IS running (started in RegionClusterState::new), so once the
    // leader heals it, the relay re-ships signals — but the item broadcast that
    // failed during downtime is what heal must backfill.

    let client = reqwest::blocking::Client::new();
    let leader_base = format!("http://{}", pair.leader_http);
    let follower_base = format!("http://{}", pair.follower_http);

    // Partition the follower at the RELAY level too, so the eager signal ships
    // are skipped while it is down (mirrors the real partition: no gRPC either).
    let resp = client
        .post(format!("{leader_base}/cluster/partition"))
        .json(&serde_json::json!({ "region": pair.follower_name }))
        .send()
        .unwrap();
    assert!(resp.status().is_success(), "partition: {}", resp.status());

    // Leader writes items + signals for entities 6,7,8 while the follower is DOWN.
    // The item broadcast to the follower fails (connection refused); the signal
    // ship is skipped (partitioned).
    for e in 6..=8u64 {
        let resp = client
            .post(format!("{leader_base}/items"))
            .json(&serde_json::json!({
                "entity_id": e, "metadata": { "title": format!("item {e}") }
            }))
            .send()
            .unwrap();
        assert!(resp.status().is_success(), "leader item: {}", resp.status());
        let resp = client
            .post(format!("{leader_base}/signals"))
            .json(&serde_json::json!({ "entity_id": e, "signal": "view", "weight": 1.0 }))
            .send()
            .unwrap();
        assert!(
            resp.status().is_success(),
            "leader signal: {}",
            resp.status()
        );
    }

    // The follower comes back up (HTTP server starts serving). Its store is empty
    // for these items — the broadcast during its downtime was lost.
    serve(
        &rt,
        build_region_router(Arc::clone(&follower), None),
        pair.follower_http,
    );

    // The recovered follower knows none of these items: its feed is empty.
    let pre_feed: serde_json::Value = client
        .get(format!("{follower_base}/feed?profile=trending&limit=10"))
        .send()
        .unwrap()
        .json()
        .unwrap();
    assert!(
        pre_feed["items"].as_array().unwrap().is_empty(),
        "recovered follower must not know the leader's items yet (broadcast lost during \
         downtime): {pre_feed}"
    );

    // Heal: the leader re-ships the missed signal segments AND must backfill the
    // missed item metadata. After heal, the follower has EXACTLY the leader's
    // data — its feed ranks items 6,7,8 WITHOUT the test re-posting them.
    let resp = client
        .post(format!("{leader_base}/cluster/heal"))
        .json(&serde_json::json!({ "region": pair.follower_name }))
        .send()
        .unwrap();
    assert!(resp.status().is_success(), "heal: {}", resp.status());

    // Converge on the signal HWM (3 segments), then assert item parity.
    poll_status(&client, &follower_base, |applied, lag| {
        applied >= 3 && lag == 0
    });

    // Poll the follower's feed until the backfilled items appear (heal's item
    // re-broadcast is async over HTTP).
    let deadline = Instant::now() + Duration::from_secs(5);
    loop {
        let feed: serde_json::Value = client
            .get(format!("{follower_base}/feed?profile=trending&limit=10"))
            .send()
            .unwrap()
            .json()
            .unwrap();
        let ids: std::collections::HashSet<u64> = feed["items"]
            .as_array()
            .unwrap()
            .iter()
            .filter_map(|it| it["entity_id"].as_u64())
            .collect();
        if [6u64, 7, 8].iter().all(|e| ids.contains(e)) {
            break;
        }
        assert!(
            Instant::now() <= deadline,
            "heal must backfill items 6,7,8 to the follower's feed; saw {ids:?}"
        );
        std::thread::sleep(Duration::from_millis(20));
    }

    rt.shutdown_timeout(Duration::from_secs(2));
}

/// BUG 1 reproduction: after the leader stream populates a follower's lag gauge
/// (HWM = N for the OLD leader's shard), promoting a DIFFERENT region to leader
/// must NOT leave the now-non-leader follower reporting a permanent stale lag.
///
/// Before the fix, `local_status` computed
/// `lag = lag_gauge.leader_seqno() − applied_seqno(NEW-leader-shard)`. The gauge's
/// `leader_seqno` is a single monotonic scalar fed by EVERY source stream, so it
/// still held the OLD leader's HWM (N), while `applied_seqno(new-leader-shard)`
/// was 0 (the new leader had not shipped). A fully-converged follower then
/// reported `lag = N − 0 = N` forever.
#[test]
fn region_node_lag_honest_across_promote() {
    // Number of signals the old leader ships before the leadership change.
    const N: u64 = 6;

    let trio = Trio::new();

    // Three nodes: us-east (leader, shard 0), eu-west (shard 1), ap-south
    // (shard 2). Dirs FIRST (see `region_dir` for the drop-order contract).
    let dirs: Vec<tempfile::TempDir> = (0..3).map(|_| region_dir()).collect();
    let nodes: Vec<Arc<RegionClusterState>> = (0..3)
        .map(|i| Arc::new(build_region(trio.topology(), &trio.names[i], &dirs[i])))
        .collect();
    // Build the runtime AFTER the nodes (and their data-dir guards): locals
    // drop in reverse order, so the runtime — which owns the nodes via the
    // serve tasks — tears down BEFORE the dirs delete, on the panic-unwind
    // path too (m11p2: nodes are persistent now).
    let rt = tokio::runtime::Builder::new_multi_thread()
        .worker_threads(3)
        .enable_all()
        .build()
        .unwrap();
    for (i, node) in nodes.iter().enumerate() {
        serve(
            &rt,
            build_region_router(Arc::clone(node), None),
            trio.http[i],
        );
    }

    let client = reqwest::blocking::Client::new();
    let bases: Vec<String> = (0..3).map(|i| format!("http://{}", trio.http[i])).collect();

    // us-east (shard 0) leads: write N signals, replicated to eu-west + ap-south.
    // This populates each follower's lag gauge HWM for the OLD leader's shard.
    for e in 1..=N {
        for base in &bases {
            let _ = client
                .post(format!("{base}/items"))
                .json(&serde_json::json!({ "entity_id": e, "metadata": {} }))
                .send()
                .unwrap();
        }
        let resp = client
            .post(format!("{}/signals", bases[0]))
            .json(&serde_json::json!({ "entity_id": e, "signal": "view", "weight": 1.0 }))
            .send()
            .unwrap();
        assert!(
            resp.status().is_success(),
            "leader signal: {}",
            resp.status()
        );
    }

    // eu-west (the future non-leader) fully converges on the old leader's stream.
    poll_status(&client, &bases[1], |applied, lag| applied >= N && lag == 0);

    // Promote ap-south (shard 2) to leader on ALL nodes. ap-south has shipped
    // NOTHING yet, so every node's applied-seqno for shard 2 is 0.
    for base in &bases {
        let resp = client
            .post(format!("{base}/cluster/promote"))
            .json(&serde_json::json!({ "region": trio.names[2] }))
            .send()
            .unwrap();
        assert!(resp.status().is_success(), "promote: {}", resp.status());
    }

    // eu-west is now a NON-leader, fully converged (nothing new to apply). Its lag
    // MUST be 0 — it is not behind the new leader, which has shipped nothing.
    // Before the fix this read N (stale gauge HWM from shard 0 − applied(shard2)=0).
    let status: serde_json::Value = client
        .get(format!("{}/cluster/status/local", bases[1]))
        .send()
        .unwrap()
        .json()
        .unwrap();
    assert_eq!(
        status["lag_events"].as_u64(),
        Some(0),
        "a converged non-leader must report lag 0 against a new leader that shipped \
         nothing — stale-gauge lag is bug 1: {status}"
    );

    rt.shutdown_timeout(Duration::from_secs(2));
}