diff --git a/crates/buzz-db/src/lib.rs b/crates/buzz-db/src/lib.rs
index 55bff9759..37977a454 100644
--- a/crates/buzz-db/src/lib.rs
+++ b/crates/buzz-db/src/lib.rs
@@ -6562,6 +6562,65 @@ mod tests {
         .await;
     }
 
+    /// Migrations must outlive the runtime caps — an index build or an
+    /// `ACCESS EXCLUSIVE` wait routinely exceeds them, and startup treats a
+    /// migration failure as fatal — and the relaxed session must not survive
+    /// into the pool afterwards.
+    #[tokio::test]
+    #[ignore = "requires Postgres"]
+    async fn migrations_ignore_runtime_timeouts_and_leak_no_relaxed_session() {
+        const TIGHT: &str = "50ms";
+
+        let admin = PgPool::connect(&admin_url().await)
+            .await
+            .expect("connect admin");
+        let name = format!("migration_timeouts_{}", Uuid::new_v4().simple());
+        sqlx::query(sqlx::AssertSqlSafe(format!("CREATE DATABASE {name}")))
+            .execute(&admin)
+            .await
+            .expect("create scratch db");
+        let base = admin_url().await;
+        let idx = base.rfind('/').expect("db url has a path segment");
+        let scratch_url = format!("{}/{}", &base[..idx], name);
+
+        // Far shorter than the migration suite needs, ample for a pooled query.
+        let db = Db::new(&DbConfig {
+            database_url: scratch_url,
+            max_connections: 2,
+            min_connections: 2,
+            statement_timeout: TIGHT.to_string(),
+            lock_timeout: TIGHT.to_string(),
+            ..DbConfig::default()
+        })
+        .await
+        .expect("connect Db against the unmigrated scratch db");
+
+        db.migrate()
+            .await
+            .expect("migrations must not inherit the runtime caps");
+
+        // Hold every connection at once so a leaked relaxed session cannot hide
+        // behind a freshly dialed one.
+        let mut held = Vec::new();
+        for _ in 0..2 {
+            let mut connection = db.pool.acquire().await.expect("acquire pooled connection");
+            let statement_timeout: String = sqlx::query_scalar("SHOW statement_timeout")
+                .fetch_one(&mut *connection)
+                .await
+                .expect("SHOW statement_timeout");
+            let lock_timeout: String = sqlx::query_scalar("SHOW lock_timeout")
+                .fetch_one(&mut *connection)
+                .await
+                .expect("SHOW lock_timeout");
+            assert_eq!(statement_timeout, TIGHT);
+            assert_eq!(lock_timeout, TIGHT);
+            held.push(connection);
+        }
+        drop(held);
+
+        drop_scratch_db(&admin, db.pool.clone(), &name).await;
+    }
+
     /// Insert identical community + channel rows into a database so the same
     /// (community, channel) ids resolve in both writer and replica.
     async fn seed_community_channel(
diff --git a/crates/buzz-db/src/migration.rs b/crates/buzz-db/src/migration.rs
index 30eee9df3..09a993518 100644
--- a/crates/buzz-db/src/migration.rs
+++ b/crates/buzz-db/src/migration.rs
@@ -38,18 +38,31 @@ pub async fn run_migrations(pool: &PgPool) -> Result<()> {
 
 async fn run_migrator_without_runtime_timeouts(pool: &PgPool) -> Result<()> {
     let mut connection = pool.acquire().await?;
+    lift_runtime_timeouts(&mut connection).await?;
+    let migrated = MIGRATOR.run(&mut *connection).await;
+    // Retire the connection either way; report the migration outcome first so a
+    // close failure cannot mask it.
+    let retired = retire_connection(connection).await;
+    migrated?;
+    retired?;
+    Ok(())
+}
+
+/// Remove both runtime limits from one connection's session.
+async fn lift_runtime_timeouts(connection: &mut sqlx::PgConnection) -> Result<()> {
     crate::apply_runtime_connection_timeouts(
-        &mut connection,
+        connection,
         crate::TIMEOUT_DISABLED,
         crate::TIMEOUT_DISABLED,
     )
     .await?;
-    let migrated = MIGRATOR.run(&mut *connection).await;
-    // Retire the connection either way; report the migration outcome first so a
-    // close failure cannot mask it.
-    let closed = connection.detach().close().await;
-    migrated?;
-    closed?;
+    Ok(())
+}
+
+/// Close a connection instead of returning it to the pool, so a session that
+/// carries lifted limits can never serve traffic.
+async fn retire_connection(connection: sqlx::pool::PoolConnection<sqlx::Postgres>) -> Result<()> {
+    connection.detach().close().await?;
     Ok(())
 }
 
@@ -1149,6 +1162,69 @@ mod tests {
         .expect("read applied migrations")
     }
 
+    /// The migration connection must have both limits lifted, and it must not
+    /// come back to the pool afterwards — a session with no statement timeout
+    /// serving traffic is the failure this exemption trades against.
+    #[tokio::test]
+    #[ignore = "requires Postgres"]
+    async fn migration_connection_is_unbounded_and_is_retired_not_reused() {
+        const TIGHT: &str = "50ms";
+
+        let database_url = std::env::var("BUZZ_TEST_DATABASE_URL")
+            .or_else(|_| std::env::var("DATABASE_URL"))
+            .unwrap_or_else(|_| TEST_DB_URL.to_owned());
+        // One slot: a reused connection would be handed straight back below.
+        let pool = sqlx::postgres::PgPoolOptions::new()
+            .max_connections(1)
+            .after_connect(|connection, _meta| {
+                Box::pin(crate::apply_runtime_connection_timeouts(
+                    connection, TIGHT, TIGHT,
+                ))
+            })
+            .connect(&database_url)
+            .await
+            .expect("connect to test DB");
+
+        let mut connection = pool.acquire().await.expect("acquire");
+        assert_eq!(
+            show_timeout(&mut connection, "statement_timeout").await,
+            TIGHT
+        );
+
+        lift_runtime_timeouts(&mut connection)
+            .await
+            .expect("lift runtime timeouts");
+        for setting in ["statement_timeout", "lock_timeout"] {
+            assert_eq!(
+                show_timeout(&mut connection, setting).await,
+                "0",
+                "{setting} must be lifted for the migrator"
+            );
+        }
+
+        retire_connection(connection)
+            .await
+            .expect("retire migration connection");
+
+        let mut fresh = pool.acquire().await.expect("re-acquire");
+        for setting in ["statement_timeout", "lock_timeout"] {
+            assert_eq!(
+                show_timeout(&mut fresh, setting).await,
+                TIGHT,
+                "the pool must not hand out the relaxed migration session"
+            );
+        }
+        drop(fresh);
+        pool.close().await;
+    }
+
+    async fn show_timeout(connection: &mut sqlx::PgConnection, setting: &str) -> String {
+        sqlx::query_scalar(sqlx::AssertSqlSafe(format!("SHOW {setting}")))
+            .fetch_one(connection)
+            .await
+            .unwrap_or_else(|error| panic!("SHOW {setting}: {error}"))
+    }
+
     #[tokio::test]
     #[ignore = "requires Postgres"]
     async fn pre_0007_ambiguous_nip_rs_data_blocks_without_mutation_and_allows_retry() {