diff --git a/desktop/src-tauri/src/commands/mesh_llm.rs b/desktop/src-tauri/src/commands/mesh_llm.rs
index 03fc7c0ac4c..ad55d950601 100644
--- a/desktop/src-tauri/src/commands/mesh_llm.rs
+++ b/desktop/src-tauri/src/commands/mesh_llm.rs
@@ -64,6 +64,17 @@ async fn query_mesh_discovery_events(state: &AppState) -> Result<Vec<nostr::Even
     let mut events = relay::query_relay(state, &[mesh_llm::relay_membership_filter()]).await?;
     let member_pubkeys = mesh_llm::current_member_pubkeys(&events);
     if member_pubkeys.is_empty() {
+        // Distinguish "relay returned a membership snapshot listing zero
+        // members" (authoritative empty — allowed to shrink the roster to
+        // self-only) from "no membership snapshot came back at all" (a
+        // transient gap / replication lag). The relay publishes an explicit
+        // kind:13534 event even for a zero-member community, so its absence
+        // means the query is incomplete: surface it as an error so the
+        // reconcile loop keeps the current allowlist instead of flapping the
+        // node down to self-only on a successful-but-empty response.
+        if !mesh_llm::has_membership_snapshot(&events) {
+            return Err("relay returned no membership snapshot".to_string());
+        }
         return Ok(events);
     }
     let mut status_filter = mesh_llm::mesh_status_filter();
@@ -88,12 +99,25 @@ async fn query_mesh_discovery_events(state: &AppState) -> Result<Vec<nostr::Even
 }
 
 /// Resolve the admission roster by intersecting member-signed mesh status
-/// reporters with the current NIP-43 direct-member list. Missing membership or
-/// a failed query returns an empty roster, which the runtime normalizes to
-/// self-only admission.
-pub(crate) async fn resolve_trusted_owner_ids(state: &AppState) -> Vec<String> {
-    match query_mesh_discovery_events(state).await {
-        Ok(events) => mesh_llm::owner_ids_from_events(&events),
+/// reporters with the current NIP-43 direct-member list.
+///
+/// Returns `Err` when the relay query fails. Callers MUST distinguish this from
+/// an `Ok(empty)` roster (a genuinely empty community): a failed query must
+/// never be collapsed into "self-only", or a transient relay blip de-admits
+/// every other member. `reconcile_roster` relies on this to keep the current
+/// allowlist on error instead of restarting the node down to self-only.
+pub(crate) async fn resolve_trusted_owner_ids(state: &AppState) -> Result<Vec<String>, String> {
+    let events = query_mesh_discovery_events(state).await?;
+    Ok(mesh_llm::owner_ids_from_events(&events))
+}
+
+/// Resolve the roster for an initial node *start*, failing closed to self-only
+/// (an empty roster) when the relay query fails. This is safe only at start:
+/// there is no established allowlist to preserve yet. The periodic
+/// `reconcile_roster` path must NOT use this — it has a live roster to keep.
+pub(crate) async fn resolve_trusted_owner_ids_or_self_only(state: &AppState) -> Vec<String> {
+    match resolve_trusted_owner_ids(state).await {
+        Ok(owners) => owners,
         Err(error) => {
             eprintln!("buzz-mesh: roster query failed; allowing only this node: {error}");
             Vec::new()
@@ -117,7 +141,7 @@ pub(crate) async fn restore_mesh_sharing(app: &AppHandle, state: &AppState) -> C
         model_id: Some(config.model_id),
         max_vram_gb: config.max_vram_gb,
         join_token: None,
-        trusted_owner_ids: Some(resolve_trusted_owner_ids(state).await),
+        trusted_owner_ids: Some(resolve_trusted_owner_ids_or_self_only(state).await),
     };
     let started = mesh_llm::DesktopMeshRuntime::start(request)
         .await
@@ -137,7 +161,7 @@ pub async fn mesh_start_node(
     // Frontend requests never carry a roster; resolve it here so every
     // UI-started node enforces the member allowlist.
     if request.trusted_owner_ids.is_none() {
-        request.trusted_owner_ids = Some(resolve_trusted_owner_ids(&state).await);
+        request.trusted_owner_ids = Some(resolve_trusted_owner_ids_or_self_only(&state).await);
     }
     let mut runtime = state.mesh_llm_runtime.lock().await;
     if runtime.is_some() {
@@ -268,7 +292,7 @@ pub(crate) async fn ensure_client_node_for_model(
         model_id: None,
         max_vram_gb: None,
         join_token: Some(join_token),
-        trusted_owner_ids: Some(resolve_trusted_owner_ids(state).await),
+        trusted_owner_ids: Some(resolve_trusted_owner_ids_or_self_only(state).await),
     };
     let mut runtime = state.mesh_llm_runtime.lock().await;
     if runtime.is_some() {
diff --git a/desktop/src-tauri/src/mesh_llm/coordinator.rs b/desktop/src-tauri/src/mesh_llm/coordinator.rs
index ad67fd62f5d..e83bf8b2807 100644
--- a/desktop/src-tauri/src/mesh_llm/coordinator.rs
+++ b/desktop/src-tauri/src/mesh_llm/coordinator.rs
@@ -51,10 +51,13 @@ pub async fn start_coordinator(app: AppHandle) {
     });
     let roster_app = app.clone();
     let roster_watcher = tokio::spawn(async move {
+        // Carries a shrink awaiting confirmation across polls (hysteresis):
+        // a reduced roster must be seen twice in a row before we tear down.
+        let mut pending_shrink: Option<Vec<String>> = None;
         loop {
             tokio::time::sleep(ROSTER_POLL_INTERVAL).await;
             let state = roster_app.state::<AppState>();
-            if let Err(error) = reconcile_roster(&state).await {
+            if let Err(error) = reconcile_roster(&state, &mut pending_shrink).await {
                 eprintln!("buzz-mesh: roster reconcile failed: {error}");
             }
         }
@@ -73,21 +76,108 @@ pub async fn start_coordinator(app: AppHandle) {
     }
 }
 
-async fn reconcile_roster(state: &AppState) -> Result<(), String> {
+/// Outcome of a roster reconcile decision.
+#[derive(Debug, PartialEq, Eq)]
+enum RosterReconcileAction {
+    /// Keep the running allowlist untouched (no-op, or a failure we ride out).
+    Keep,
+    /// Restart the node with a freshly resolved roster.
+    Restart(Vec<String>),
+    /// Observed a *shrink* (or empty) once. Hold the current allowlist and
+    /// require the same reduced roster on the next poll before tearing down,
+    /// so a single transient short-read never drops a member mid-inference.
+    AwaitConfirm(Vec<String>),
+}
+
+/// Whether `fresh` removes any owner present in `current` (a shrink), as
+/// opposed to purely adding owners or leaving the set unchanged.
+fn roster_shrinks(current: &[String], fresh: &[String]) -> bool {
+    current.iter().any(|owner| !fresh.contains(owner))
+}
+
+/// Pure decision for `reconcile_roster`, extracted so the transient-failure
+/// and hysteresis invariants are unit-testable without a live relay.
+///
+/// `pending_shrink` is the reduced roster we are waiting to re-confirm (from a
+/// prior poll's [`RosterReconcileAction::AwaitConfirm`]), if any.
+///
+/// Rules:
+/// - query failed (`Err`)              → `Keep` (never de-admit on a relay blip)
+/// - resolved roster == current        → `Keep` (no-op)
+/// - grows (only additions)            → `Restart` immediately (fast admission)
+/// - shrinks/empties, first observation → `AwaitConfirm` (hold, re-check next poll)
+/// - shrinks/empties, confirmed         → `Restart` (same reduced roster twice)
+fn roster_reconcile_action(
+    current_owners: &[String],
+    pending_shrink: Option<&[String]>,
+    query: Result<Vec<String>, String>,
+) -> RosterReconcileAction {
+    let fresh = match query {
+        Err(error) => {
+            eprintln!(
+                "buzz-mesh: roster reconcile query failed; keeping current allowlist: {error}"
+            );
+            return RosterReconcileAction::Keep;
+        }
+        Ok(fresh) => fresh,
+    };
+
+    if fresh == current_owners {
+        return RosterReconcileAction::Keep;
+    }
+
+    // Growth (pure additions) is safe to apply immediately.
+    if !roster_shrinks(current_owners, &fresh) {
+        return RosterReconcileAction::Restart(fresh);
+    }
+
+    // A shrink (including down to empty) must be confirmed across two
+    // consecutive polls with the *same* reduced roster before we tear down.
+    match pending_shrink {
+        Some(pending) if pending == fresh => RosterReconcileAction::Restart(fresh),
+        _ => RosterReconcileAction::AwaitConfirm(fresh),
+    }
+}
+
+async fn reconcile_roster(
+    state: &AppState,
+    pending_shrink: &mut Option<Vec<String>>,
+) -> Result<(), String> {
     let current_request = {
         let runtime = state.mesh_llm_runtime.lock().await;
         match runtime.as_ref() {
             Some(runtime) => runtime.start_request().clone(),
-            None => return Ok(()),
+            None => {
+                *pending_shrink = None;
+                return Ok(());
+            }
         }
     };
     let Some(current_owners) = current_request.trusted_owner_ids.as_ref() else {
+        *pending_shrink = None;
         return Ok(());
     };
-    let fresh = crate::commands::mesh_llm::resolve_trusted_owner_ids(state).await;
-    if &fresh == current_owners {
-        return Ok(());
-    }
+    // A failed roster query must NOT be treated as "the roster became empty":
+    // doing so would restart the node down to self-only and de-admit every
+    // other member on a transient relay blip (the flapping restart loop). Keep
+    // the current allowlist and try again on the next poll. A shrink is held
+    // for one extra poll (hysteresis) so a single short-read never tears down.
+    let query = crate::commands::mesh_llm::resolve_trusted_owner_ids(state).await;
+    let fresh = match roster_reconcile_action(current_owners, pending_shrink.as_deref(), query) {
+        RosterReconcileAction::Keep => {
+            *pending_shrink = None;
+            return Ok(());
+        }
+        RosterReconcileAction::AwaitConfirm(reduced) => {
+            eprintln!("buzz-mesh: roster shrink observed; awaiting confirmation before restart");
+            *pending_shrink = Some(reduced);
+            return Ok(());
+        }
+        RosterReconcileAction::Restart(fresh) => {
+            *pending_shrink = None;
+            fresh
+        }
+    };
 
     let mut request = current_request;
     request.trusted_owner_ids = Some(fresh);
@@ -227,6 +317,91 @@ mod tests {
 
     use super::*;
 
+    // Regression: a transient roster-query failure must never restart the node
+    // down to self-only. Before the fix, `resolve_trusted_owner_ids` returned
+    // an empty Vec on error, which `reconcile_roster` read as "roster changed
+    // to empty" and restarted — de-admitting every other member and flapping
+    // the node on each relay blip. See #2000 follow-up.
+    #[test]
+    fn failed_roster_query_keeps_current_allowlist() {
+        let current = vec!["owner-a".to_string(), "owner-b".to_string()];
+        let action = roster_reconcile_action(&current, None, Err("relay returned 503".to_string()));
+        assert_eq!(
+            action,
+            RosterReconcileAction::Keep,
+            "a failed query must keep the running allowlist, never de-admit members"
+        );
+    }
+
+    #[test]
+    fn unchanged_roster_is_a_noop() {
+        let current = vec!["owner-a".to_string()];
+        let action = roster_reconcile_action(&current, None, Ok(vec!["owner-a".to_string()]));
+        assert_eq!(action, RosterReconcileAction::Keep);
+    }
+
+    // Growth (pure additions) applies immediately — fast admission is fine.
+    #[test]
+    fn roster_growth_restarts_immediately() {
+        let current = vec!["owner-a".to_string()];
+        let fresh = vec!["owner-a".to_string(), "owner-c".to_string()];
+        let action = roster_reconcile_action(&current, None, Ok(fresh.clone()));
+        assert_eq!(action, RosterReconcileAction::Restart(fresh));
+    }
+
+    // A shrink is NOT applied on first observation — it must be confirmed.
+    #[test]
+    fn roster_shrink_awaits_confirmation_first() {
+        let current = vec!["owner-a".to_string(), "owner-b".to_string()];
+        let reduced = vec!["owner-a".to_string()];
+        let action = roster_reconcile_action(&current, None, Ok(reduced.clone()));
+        assert_eq!(action, RosterReconcileAction::AwaitConfirm(reduced));
+    }
+
+    // The same reduced roster on two consecutive polls confirms the shrink.
+    #[test]
+    fn roster_shrink_restarts_once_confirmed() {
+        let current = vec!["owner-a".to_string(), "owner-b".to_string()];
+        let reduced = vec!["owner-a".to_string()];
+        let action = roster_reconcile_action(&current, Some(&reduced), Ok(reduced.clone()));
+        assert_eq!(action, RosterReconcileAction::Restart(reduced));
+    }
+
+    // A shrink that changes between polls is not confirmed — it re-holds with
+    // the newly observed reduced roster instead of tearing down.
+    #[test]
+    fn roster_shrink_reconfirms_when_it_changes() {
+        let current = vec!["a".to_string(), "b".to_string(), "c".to_string()];
+        let first_reduced = vec!["a".to_string(), "b".to_string()];
+        let second_reduced = vec!["a".to_string()];
+        let action =
+            roster_reconcile_action(&current, Some(&first_reduced), Ok(second_reduced.clone()));
+        assert_eq!(action, RosterReconcileAction::AwaitConfirm(second_reduced));
+    }
+
+    // A genuinely empty community (Ok(empty), distinct from a failed query)
+    // still shrinks to self-only — but only after confirmation.
+    #[test]
+    fn genuinely_empty_roster_awaits_then_restarts_to_self_only() {
+        let current = vec!["owner-a".to_string()];
+        let first = roster_reconcile_action(&current, None, Ok(Vec::new()));
+        assert_eq!(first, RosterReconcileAction::AwaitConfirm(Vec::new()));
+        let empty: Vec<String> = Vec::new();
+        let confirmed = roster_reconcile_action(&current, Some(&empty), Ok(Vec::new()));
+        assert_eq!(confirmed, RosterReconcileAction::Restart(Vec::new()));
+    }
+
+    // A shrink followed by recovery to the full roster cancels the teardown.
+    #[test]
+    fn roster_shrink_then_recovery_keeps_allowlist() {
+        let current = vec!["owner-a".to_string(), "owner-b".to_string()];
+        let reduced = vec!["owner-a".to_string()];
+        let held = roster_reconcile_action(&current, None, Ok(reduced.clone()));
+        assert_eq!(held, RosterReconcileAction::AwaitConfirm(reduced.clone()));
+        let recovered = roster_reconcile_action(&current, Some(&reduced), Ok(current.clone()));
+        assert_eq!(recovered, RosterReconcileAction::Keep);
+    }
+
     #[test]
     fn member_heartbeat_leaves_room_before_admission_status_expires() {
         assert!(
diff --git a/desktop/src-tauri/src/mesh_llm/discovery.rs b/desktop/src-tauri/src/mesh_llm/discovery.rs
index b4624ca7dc1..d7dca6b11bd 100644
--- a/desktop/src-tauri/src/mesh_llm/discovery.rs
+++ b/desktop/src-tauri/src/mesh_llm/discovery.rs
@@ -93,6 +93,20 @@ pub(crate) fn current_member_pubkeys(events: &[nostr::Event]) -> Vec<String> {
         .unwrap_or_default()
 }
 
+/// Whether the relay actually returned a NIP-43 membership snapshot (kind
+/// 13534) in `events`.
+///
+/// The relay publishes an explicit membership event even for a zero-member
+/// community, so its presence is what makes an empty roster *authoritative*.
+/// Callers use this to distinguish "the community genuinely has no members"
+/// (snapshot present, zero `member` tags) from "no snapshot came back at all"
+/// (a transient relay gap / replication lag). Only the former may shrink the
+/// admission roster; the latter must be surfaced as an error so the reconcile
+/// loop keeps the current allowlist instead of restarting to self-only.
+pub(crate) fn has_membership_snapshot(events: &[nostr::Event]) -> bool {
+    events.iter().any(|event| event.kind.as_u16() == 13_534)
+}
+
 fn owner_id_from_status_event(event: &nostr::Event) -> Option<String> {
     let content = serde_json::from_str::<serde_json::Value>(&event.content).ok()?;
     let owner_id = content
@@ -200,11 +214,12 @@ pub fn availability_from_events(events: Vec<nostr::Event>) -> MeshAvailability {
             .unwrap_or_default()
             .into_iter()
             .filter_map(|mut target| {
-                let endpoint_id =
+                let validated =
                     super::transport_policy::validate_advertised_endpoint(&target.endpoint_addr)
                         .ok()?;
+                target.endpoint_addr = validated.join_token;
                 if target.endpoint_id.is_none() {
-                    target.endpoint_id = Some(endpoint_id);
+                    target.endpoint_id = Some(validated.endpoint_id);
                 }
                 if target.device_id.is_none() {
                     target.device_id = target.endpoint_id.clone();
@@ -322,7 +337,9 @@ pub(super) fn device_name_from_status(
 }
 
 fn endpoint_id_from_invite_token(invite_token: &str) -> Option<String> {
-    super::transport_policy::validate_advertised_endpoint(invite_token).ok()
+    super::transport_policy::validate_advertised_endpoint(invite_token)
+        .ok()
+        .map(|validated| validated.endpoint_id)
 }
 
 fn string_value(value: &serde_json::Value, key: &str) -> Option<String> {
diff --git a/desktop/src-tauri/src/mesh_llm/mod.rs b/desktop/src-tauri/src/mesh_llm/mod.rs
index 59e46492223..80d887bec7d 100644
--- a/desktop/src-tauri/src/mesh_llm/mod.rs
+++ b/desktop/src-tauri/src/mesh_llm/mod.rs
@@ -8,7 +8,9 @@ mod discovery;
 pub use discovery::{
     availability_from_events, mesh_status_filter, owner_ids_from_events, relay_membership_filter,
 };
-pub(crate) use discovery::{current_member_pubkeys, MESH_STATUS_PAGE_SIZE};
+pub(crate) use discovery::{
+    current_member_pubkeys, has_membership_snapshot, MESH_STATUS_PAGE_SIZE,
+};
 use discovery::{device_name_from_status, endpoint_id_from_status, enrich_status_payload_identity};
 
 mod catalog;
@@ -270,8 +272,8 @@ async fn ensure_model_downloaded(model: &str) -> anyhow::Result<()> {
 }
 
 impl DesktopMeshRuntime {
-    pub async fn start(request: StartMeshNodeRequest) -> anyhow::Result<Self> {
-        validate_no_leak_request(&request)?;
+    pub async fn start(mut request: StartMeshNodeRequest) -> anyhow::Result<Self> {
+        sanitize_no_leak_request(&mut request)?;
         initialize_mesh_native_runtime().await?;
         let model_id = request
             .model_id
@@ -420,8 +422,8 @@ impl DesktopMeshRuntime {
 
     pub async fn dial_endpoint_addr(&self, endpoint_addr: impl Into<String>) -> anyhow::Result<()> {
         let endpoint_addr = endpoint_addr.into();
-        validate_advertised_endpoint(&endpoint_addr)?;
-        self.handle.join_token(endpoint_addr).await
+        let validated = validate_advertised_endpoint(&endpoint_addr)?;
+        self.handle.join_token(validated.join_token).await
     }
 
     pub async fn installed_models(&self) -> anyhow::Result<Vec<MeshModelOption>> {
@@ -503,9 +505,9 @@ fn normalized_roster(
     Some(owners)
 }
 
-fn validate_no_leak_request(request: &StartMeshNodeRequest) -> anyhow::Result<()> {
-    if let Some(join_token) = request.join_token.as_deref() {
-        validate_advertised_endpoint(join_token)?;
+fn sanitize_no_leak_request(request: &mut StartMeshNodeRequest) -> anyhow::Result<()> {
+    if let Some(join_token) = request.join_token.as_mut() {
+        *join_token = validate_advertised_endpoint(join_token)?.join_token;
     }
     Ok(())
 }
diff --git a/desktop/src-tauri/src/mesh_llm/mod_tests.rs b/desktop/src-tauri/src/mesh_llm/mod_tests.rs
index 32cfae113a1..e169af88864 100644
--- a/desktop/src-tauri/src/mesh_llm/mod_tests.rs
+++ b/desktop/src-tauri/src/mesh_llm/mod_tests.rs
@@ -214,6 +214,34 @@ fn signed_membership_event(members: &[String]) -> nostr::Event {
         .expect("test membership event signs")
 }
 
+#[test]
+fn has_membership_snapshot_distinguishes_empty_from_missing() {
+    // A zero-member community still publishes an explicit kind:13534 event, so
+    // its presence — not the member count — is what makes an empty roster
+    // authoritative. No snapshot at all means the query was incomplete.
+    let zero_member_snapshot = signed_membership_event(&[]);
+    assert!(
+        super::has_membership_snapshot(std::slice::from_ref(&zero_member_snapshot)),
+        "an explicit zero-member snapshot counts as present"
+    );
+
+    let member = nostr::Keys::parse(&"1".repeat(64))
+        .unwrap()
+        .public_key()
+        .to_hex();
+    let populated = signed_membership_event(std::slice::from_ref(&member));
+    assert!(super::has_membership_snapshot(std::slice::from_ref(
+        &populated
+    )));
+
+    // A response with only status events (or nothing) has no snapshot.
+    assert!(!super::has_membership_snapshot(&[]));
+    let status_only = signed_reporter_status(&"2".repeat(64), "owner-x");
+    assert!(!super::has_membership_snapshot(std::slice::from_ref(
+        &status_only
+    )));
+}
+
 #[test]
 fn owner_ids_from_events_collects_sorted_deduped_roster() {
     let secret_a = "1".repeat(64);
diff --git a/desktop/src-tauri/src/mesh_llm/transport_policy.rs b/desktop/src-tauri/src/mesh_llm/transport_policy.rs
index e4e7ca9fe1e..122bffd60e4 100644
--- a/desktop/src-tauri/src/mesh_llm/transport_policy.rs
+++ b/desktop/src-tauri/src/mesh_llm/transport_policy.rs
@@ -72,7 +72,15 @@ fn parse_configured_relay_url(raw: &str) -> anyhow::Result<RelayUrl> {
         .map_err(|error| anyhow::anyhow!("invalid iroh relay URL {raw:?}: {error}"))
 }
 
-pub(super) fn validate_advertised_endpoint(invite_token: &str) -> anyhow::Result<String> {
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(super) struct ValidatedEndpoint {
+    pub endpoint_id: String,
+    pub join_token: String,
+}
+
+pub(super) fn validate_advertised_endpoint(
+    invite_token: &str,
+) -> anyhow::Result<ValidatedEndpoint> {
     let mode = iroh_relay_mode()?;
     validate_advertised_endpoint_with_mode(invite_token, &mode)
 }
@@ -80,7 +88,7 @@ pub(super) fn validate_advertised_endpoint(invite_token: &str) -> anyhow::Result
 pub(super) fn validate_advertised_endpoint_with_mode(
     invite_token: &str,
     mode: &IrohRelayMode,
-) -> anyhow::Result<String> {
+) -> anyhow::Result<ValidatedEndpoint> {
     let token = invite_token.trim();
     if token.is_empty() {
         anyhow::bail!("mesh invite token is empty");
@@ -92,59 +100,139 @@ pub(super) fn validate_advertised_endpoint_with_mode(
         .decode(token)
         .map_err(|error| anyhow::anyhow!("invalid mesh invite encoding: {error}"))?;
 
-    let addrs = if let Ok(addr) = serde_json::from_slice::<EndpointAddr>(&payload) {
-        vec![addr]
-    } else {
-        let signed = serde_json::from_slice::<SignedBootstrapToken>(&payload)
-            .map_err(|error| anyhow::anyhow!("invalid mesh invite payload: {error}"))?;
-        signed
-            .verify()
-            .map_err(|reason| anyhow::anyhow!("invalid signed mesh invite: {}", reason.code()))?;
-        if signed.serialized_addrs.is_empty() || signed.serialized_addrs.len() > MAX_BOOTSTRAP_ADDRS
-        {
-            anyhow::bail!(
-                "signed mesh invite must contain 1..={MAX_BOOTSTRAP_ADDRS} endpoint addresses"
-            );
-        }
-        signed
-            .serialized_addrs
-            .iter()
-            .map(|bytes| {
-                serde_json::from_slice::<EndpointAddr>(bytes)
-                    .map_err(|error| anyhow::anyhow!("invalid signed endpoint address: {error}"))
-            })
-            .collect::<anyhow::Result<Vec<_>>>()?
-    };
+    if let Ok(mut addr) = serde_json::from_slice::<EndpointAddr>(&payload) {
+        retain_usable_transports(&mut addr, mode)?;
+        let endpoint_id = addr.id.to_string();
+        let join_token = URL_SAFE_NO_PAD.encode(serde_json::to_vec(&addr)?);
+        return Ok(ValidatedEndpoint {
+            endpoint_id,
+            join_token,
+        });
+    }
 
+    let signed = serde_json::from_slice::<SignedBootstrapToken>(&payload)
+        .map_err(|error| anyhow::anyhow!("invalid mesh invite payload: {error}"))?;
+    signed
+        .verify()
+        .map_err(|reason| anyhow::anyhow!("invalid signed mesh invite: {}", reason.code()))?;
+    if signed.serialized_addrs.is_empty() || signed.serialized_addrs.len() > MAX_BOOTSTRAP_ADDRS {
+        anyhow::bail!(
+            "signed mesh invite must contain 1..={MAX_BOOTSTRAP_ADDRS} endpoint addresses"
+        );
+    }
+    let addrs = signed
+        .serialized_addrs
+        .iter()
+        .map(|bytes| {
+            serde_json::from_slice::<EndpointAddr>(bytes)
+                .map_err(|error| anyhow::anyhow!("invalid signed endpoint address: {error}"))
+        })
+        .collect::<anyhow::Result<Vec<_>>>()?;
+
+    // Rewriting a signed token would invalidate its signature. Keep the signed
+    // envelope intact only when every advertised transport is policy-approved;
+    // mixed signed tokens fail closed rather than leaking rejected dial targets.
     for addr in &addrs {
-        validate_endpoint_addr(addr, mode)?;
+        validate_signed_transports(addr, mode)?;
     }
-    Ok(addrs[0].id.to_string())
+    Ok(ValidatedEndpoint {
+        endpoint_id: addrs[0].id.to_string(),
+        join_token: token.to_string(),
+    })
 }
 
-fn validate_endpoint_addr(addr: &EndpointAddr, mode: &IrohRelayMode) -> anyhow::Result<()> {
-    if addr.addrs.is_empty() || addr.addrs.len() > MAX_ENDPOINT_TRANSPORT_ADDRS {
+fn retain_usable_transports(addr: &mut EndpointAddr, mode: &IrohRelayMode) -> anyhow::Result<()> {
+    validate_transport_count(addr)?;
+    let mut rejections = Vec::new();
+    addr.addrs
+        .retain(|transport| match validate_transport(transport, mode) {
+            Ok(()) => true,
+            Err(error) => {
+                rejections.push(error.to_string());
+                false
+            }
+        });
+    if addr.addrs.is_empty() {
         anyhow::bail!(
-            "mesh endpoint must contain 1..={MAX_ENDPOINT_TRANSPORT_ADDRS} transport addresses"
+            "mesh endpoint has no usable transport address (all rejected: {})",
+            rejections.join("; ")
         );
     }
+    Ok(())
+}
+
+fn validate_signed_transports(addr: &EndpointAddr, mode: &IrohRelayMode) -> anyhow::Result<()> {
+    validate_transport_count(addr)?;
+    let mut usable = 0usize;
     for transport in &addr.addrs {
-        match transport {
-            TransportAddr::Relay(relay) if relay_allowed(relay, mode) => {}
-            TransportAddr::Relay(relay) => {
-                anyhow::bail!("mesh endpoint advertises unapproved relay URL {relay}")
-            }
-            TransportAddr::Ip(socket) => validate_direct_socket(*socket)?,
-            _ => anyhow::bail!("mesh endpoint contains an unsupported transport address"),
+        match validate_transport(transport, mode) {
+            Ok(()) => usable += 1,
+            // mesh-llm currently signs its own port-0 placeholder alongside a
+            // valid relay. It is non-dialable, so preserving it is safe and
+            // necessary for stock signed tokens to remain usable.
+            Err(_) if matches!(transport, TransportAddr::Ip(socket) if socket.port() == 0) => {}
+            Err(error) => return Err(error),
         }
     }
+    if usable == 0 {
+        anyhow::bail!("signed mesh endpoint has no usable transport address");
+    }
     Ok(())
 }
 
+fn validate_transport_count(addr: &EndpointAddr) -> anyhow::Result<()> {
+    if addr.addrs.is_empty() || addr.addrs.len() > MAX_ENDPOINT_TRANSPORT_ADDRS {
+        anyhow::bail!(
+            "mesh endpoint must contain 1..={MAX_ENDPOINT_TRANSPORT_ADDRS} transport addresses"
+        );
+    }
+    Ok(())
+}
+
+fn validate_transport(transport: &TransportAddr, mode: &IrohRelayMode) -> anyhow::Result<()> {
+    match transport {
+        TransportAddr::Relay(relay) if relay_allowed(relay, mode) => Ok(()),
+        TransportAddr::Relay(relay) => anyhow::bail!("unapproved relay URL {relay}"),
+        TransportAddr::Ip(socket) => validate_direct_socket(*socket),
+        _ => anyhow::bail!("unsupported transport address"),
+    }
+}
+
+/// mesh-llm's default public relay set (`RelayPolicy::DefaultPublic` in
+/// `mesh-llm-host-runtime`). A stock mesh-llm server with no custom relay
+/// config advertises endpoints on exactly these relays, so buzz's `Default`
+/// mode MUST accept them — otherwise shared compute rejects every out-of-the-box
+/// mesh-llm serving node (they are not in iroh's own prod relay map).
+///
+/// Kept in sync with `effective_relay_urls(RelayPolicy::DefaultPublic, &[])`.
+const MESH_LLM_DEFAULT_RELAYS: &[&str] = &[
+    "https://usw1-2.relay.michaelneale.mesh-llm.iroh.link./",
+    "https://aps1-1.relay.michaelneale.mesh-llm.iroh.link./",
+];
+
+/// Whether `relay` is one of mesh-llm's baked-in default public relays.
+/// Parses each known URL to a `RelayUrl` so comparison is normalization-safe
+/// (matches regardless of trailing-dot / trailing-slash formatting).
+fn is_mesh_llm_default_relay(relay: &RelayUrl) -> bool {
+    MESH_LLM_DEFAULT_RELAYS.iter().any(|candidate| {
+        candidate
+            .parse::<RelayUrl>()
+            .map(|known| &known == relay)
+            .unwrap_or(false)
+    })
+}
+
 fn relay_allowed(relay: &RelayUrl, mode: &IrohRelayMode) -> bool {
     match mode {
         IrohRelayMode::Disabled => false,
-        IrohRelayMode::Default => iroh::defaults::prod::default_relay_map().contains(relay),
+        // `Default` covers both iroh's own production relays AND mesh-llm's
+        // default public relays. Without the latter, a stock mesh-llm serving
+        // node is unreachable by default and shared compute silently fails with
+        // "no live member is serving this model" even though discovery found it.
+        IrohRelayMode::Default => {
+            iroh::defaults::prod::default_relay_map().contains(relay)
+                || is_mesh_llm_default_relay(relay)
+        }
         IrohRelayMode::Custom(urls) => urls.contains(relay),
     }
 }
@@ -213,6 +301,76 @@ mod tests {
         assert!(validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Disabled).is_err());
     }
 
+    #[test]
+    fn default_mode_accepts_meshllm_default_relays() {
+        // Regression: a stock mesh-llm serving node advertises endpoints on
+        // mesh-llm's OWN default public relays (not iroh's prod relay map).
+        // Under `Default` mode these MUST be accepted, or shared compute rejects
+        // every out-of-the-box mesh-llm server with "no live member is serving
+        // this model" even though discovery found it. See mesh-llm
+        // effective_relay_urls(RelayPolicy::DefaultPublic, &[]).
+        for relay_url in MESH_LLM_DEFAULT_RELAYS {
+            let relay: RelayUrl = relay_url
+                .parse()
+                .unwrap_or_else(|e| panic!("mesh-llm default relay {relay_url:?} must parse: {e}"));
+            assert!(
+                relay_allowed(&relay, &IrohRelayMode::Default),
+                "Default mode must accept mesh-llm default relay {relay_url}"
+            );
+
+            // And end-to-end through the advertised-endpoint validator.
+            let token = endpoint_token_for_test([TransportAddr::Relay(relay)]);
+            assert!(
+                validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default).is_ok(),
+                "Default mode must validate an endpoint on mesh-llm default relay {relay_url}"
+            );
+        }
+    }
+
+    #[test]
+    fn endpoint_with_one_good_and_one_junk_candidate_is_sanitized() {
+        // A mesh-llm endpoint can advertise a usable relay alongside an
+        // unusable direct IP. Keep the endpoint reachable, but never pass the
+        // rejected candidate through to iroh's parallel dialer.
+        let good_relay: RelayUrl = MESH_LLM_DEFAULT_RELAYS[0].parse().unwrap();
+        let unsafe_socket = "169.254.169.254:80".parse().unwrap();
+        let token = endpoint_token_for_test([
+            TransportAddr::Relay(good_relay.clone()),
+            TransportAddr::Ip(unsafe_socket),
+        ]);
+        let validated =
+            validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default).unwrap();
+        let payload = URL_SAFE_NO_PAD.decode(validated.join_token).unwrap();
+        let sanitized: EndpointAddr = serde_json::from_slice(&payload).unwrap();
+        assert_eq!(
+            sanitized.addrs,
+            [TransportAddr::Relay(good_relay)].into_iter().collect()
+        );
+    }
+
+    #[test]
+    fn endpoint_with_all_junk_candidates_is_rejected() {
+        // Guard: if EVERY candidate is unusable, the endpoint must still fail.
+        let token = endpoint_token_for_test([
+            TransportAddr::Ip("180.181.228.108:0".parse().unwrap()), // port 0
+            TransportAddr::Ip("127.0.0.1:9337".parse().unwrap()),    // loopback
+        ]);
+        assert!(
+            validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default).is_err(),
+            "endpoint with no usable candidate must be rejected"
+        );
+    }
+
+    #[test]
+    fn default_mode_still_rejects_unknown_relay() {
+        // Guard the fix doesn't over-open: a relay that is neither iroh-prod nor
+        // a mesh-llm default must still be rejected under Default mode.
+        let unknown: RelayUrl = "https://not-a-real-relay.example".parse().unwrap();
+        assert!(!relay_allowed(&unknown, &IrohRelayMode::Default));
+        let token = endpoint_token_for_test([TransportAddr::Relay(unknown)]);
+        assert!(validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default).is_err());
+    }
+
     #[test]
     fn remote_endpoint_rejects_unsafe_direct_targets_and_oversized_tokens() {
         for socket in ["127.0.0.1:9337", "169.254.169.254:80", "0.0.0.0:1"] {
@@ -263,10 +421,59 @@ mod tests {
         .expect("sign test bootstrap token");
         let token = URL_SAFE_NO_PAD.encode(serde_json::to_vec(&signed).unwrap());
         assert_eq!(
-            validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default).unwrap(),
+            validate_advertised_endpoint_with_mode(&token, &IrohRelayMode::Default)
+                .unwrap()
+                .endpoint_id,
             endpoint.id.to_string()
         );
 
+        let good_relay: RelayUrl = MESH_LLM_DEFAULT_RELAYS[0].parse().unwrap();
+        let placeholder_endpoint = EndpointAddr {
+            id: endpoint.id,
+            addrs: [
+                TransportAddr::Relay(good_relay.clone()),
+                TransportAddr::Ip("180.181.228.108:0".parse().unwrap()),
+            ]
+            .into_iter()
+            .collect(),
+        };
+        let placeholder_signed = SignedBootstrapToken::sign(
+            vec![serde_json::to_vec(&placeholder_endpoint).unwrap()],
+            &signed_policy,
+            None,
+            &owner,
+        )
+        .expect("sign placeholder test bootstrap token");
+        let placeholder_token =
+            URL_SAFE_NO_PAD.encode(serde_json::to_vec(&placeholder_signed).unwrap());
+        assert!(
+            validate_advertised_endpoint_with_mode(&placeholder_token, &IrohRelayMode::Default)
+                .is_ok(),
+            "signed stock token with a port-0 placeholder must remain usable"
+        );
+
+        let mixed_endpoint = EndpointAddr {
+            id: endpoint.id,
+            addrs: [
+                TransportAddr::Relay(good_relay),
+                TransportAddr::Ip("169.254.169.254:80".parse().unwrap()),
+            ]
+            .into_iter()
+            .collect(),
+        };
+        let mixed_signed = SignedBootstrapToken::sign(
+            vec![serde_json::to_vec(&mixed_endpoint).unwrap()],
+            &signed_policy,
+            None,
+            &owner,
+        )
+        .expect("sign mixed test bootstrap token");
+        let mixed_token = URL_SAFE_NO_PAD.encode(serde_json::to_vec(&mixed_signed).unwrap());
+        assert!(
+            validate_advertised_endpoint_with_mode(&mixed_token, &IrohRelayMode::Default).is_err(),
+            "signed mixed-candidate tokens must fail closed because they cannot be rewritten"
+        );
+
         let mut tampered = signed;
         tampered.serialized_addrs[0] = serde_json::to_vec(&EndpointAddr {
             id: endpoint.id,