diff --git a/tests/e2e/topologies/grid-glb-demo/resources/provider-workload-east-secondary.yaml b/tests/e2e/topologies/grid-glb-demo/resources/provider-workload-east-secondary.yaml
index a42a7181..b9826860 100644
--- a/tests/e2e/topologies/grid-glb-demo/resources/provider-workload-east-secondary.yaml
+++ b/tests/e2e/topologies/grid-glb-demo/resources/provider-workload-east-secondary.yaml
@@ -31,6 +31,8 @@ spec:
           image: "{{ cluster.properties.vcrImage }}"
           imagePullPolicy: "{{ cluster.properties.imagePullPolicy }}"
           env:
+            - name: MOCK_PROVIDER_SITE
+              value: "east-provider-secondary"
             - name: MODEL
               value: "Qwen/Qwen3-0.6B"
             - name: MOCK_PD_ROLE
@@ -52,6 +54,11 @@ spec:
           ports:
             - name: http
               containerPort: 8000
+          args:
+            - --provider
+            - openai
+            - --port
+            - "8000"
           readinessProbe:
             httpGet:
               path: /health
@@ -60,7 +67,7 @@ spec:
             periodSeconds: 5
           startupProbe:
             httpGet:
-              path: /v1/models
+              path: /health
               port: 8000
             periodSeconds: 10
             failureThreshold: 60
diff --git a/tests/e2e/topologies/grid-glb-demo/resources/provider-workloads.yaml b/tests/e2e/topologies/grid-glb-demo/resources/provider-workloads.yaml
index 977caaa2..118db4ed 100644
--- a/tests/e2e/topologies/grid-glb-demo/resources/provider-workloads.yaml
+++ b/tests/e2e/topologies/grid-glb-demo/resources/provider-workloads.yaml
@@ -30,6 +30,8 @@ spec:
           image: "{{ cluster.properties.vcrImage }}"
           imagePullPolicy: "{{ cluster.properties.imagePullPolicy }}"
           env:
+            - name: MOCK_PROVIDER_SITE
+              value: "{{ cluster.name }}"
             - name: MODEL
               value: "Qwen/Qwen3-0.6B"
             - name: MOCK_PD_ROLE
@@ -51,6 +53,11 @@ spec:
           ports:
             - name: http
               containerPort: 8000
+          args:
+            - --provider
+            - openai
+            - --port
+            - "8000"
           readinessProbe:
             httpGet:
               path: /health
@@ -59,7 +66,7 @@ spec:
             periodSeconds: 5
           startupProbe:
             httpGet:
-              path: /v1/models
+              path: /health
               port: 8000
             periodSeconds: 10
             failureThreshold: 60
diff --git a/xtask/src/env/glb.rs b/xtask/src/env/glb.rs
index 3d689225..8691368b 100644
--- a/xtask/src/env/glb.rs
+++ b/xtask/src/env/glb.rs
@@ -1566,6 +1566,12 @@ fn run_steps(ctx: &PrereqContext, mode: DemoMode, ingress_mode: IngressMode, res
         return;
     }
 
+    // Workload mode has no global ingress or GTM path. Its direct withdrawal
+    // and fallback proof runs in glb_demo with workload-specific attribution.
+    if ingress_mode == IngressMode::Workload {
+        return;
+    }
+
     // Session affinity initial bind.
     proof_banner("checking session affinity bind");
     let provider_a = match check_session_bind(EDGE_PORT) {
@@ -1830,6 +1836,7 @@ fn check_backend_network_policy(mode: IngressMode) -> Result<String, Box<dyn std
     let mut evidence = Vec::new();
     for provider in PROVIDER_CLUSTERS {
         let context = kubectl_context(provider);
+        load_network_probe_image(provider)?;
         let backends: &[(&str, &str)] = if *provider == "east-provider" {
             &[
                 ("primary", "vcr-inference.grid-system.svc.cluster.local:8000"),
@@ -1851,6 +1858,30 @@ fn check_backend_network_policy(mode: IngressMode) -> Result<String, Box<dyn std
     Ok(evidence.join("; "))
 }
 
+/// Load the restricted probe image before using `imagePullPolicy: Never`.
+fn load_network_probe_image(provider: &str) -> Result<(), Box<dyn std::error::Error>> {
+    let cluster = format!("{CLUSTER_PREFIX}-{provider}");
+    let output = Command::new("timeout")
+        .args([
+            "120s",
+            "kind",
+            "load",
+            "docker-image",
+            "curlimages/curl:8.10.1",
+            "--name",
+            &cluster,
+        ])
+        .output()?;
+    if output.status.success() {
+        return Ok(());
+    }
+    Err(format!(
+        "failed to load network-policy probe image into {cluster}: {}",
+        safe_truncate_str(String::from_utf8_lossy(&output.stderr).trim(), 240)
+    )
+    .into())
+}
+
 /// Prove network and credential rejection behavior for one private backend.
 #[expect(
     clippy::too_many_lines,
@@ -3322,6 +3353,81 @@ fn verify_edge_accepted_revision(edge: &str, revision: &str) -> Result<(), Box<d
     )
 }
 
+/// Maximum time to wait for an edge to serve a withdrawal revision.
+const EDGE_SERVING_TIMEOUT: Duration = Duration::from_secs(90);
+
+/// Poll until the edge accepts and serves `revision` in two consecutive
+/// observations without a pod replacement or container restart.
+#[expect(
+    clippy::too_many_lines,
+    reason = "the bounded serving barrier keeps polling, identity, and diagnostic state together"
+)]
+pub(crate) fn verify_edge_serving_revision(edge: &str, revision: &str) -> Result<(), Box<dyn std::error::Error>> {
+    let deadline = Instant::now() + EDGE_SERVING_TIMEOUT;
+    let initial = edge_pod_identity(edge)?;
+    let mut stable_observations = 0_u8;
+    loop {
+        let observation = edge_serving_observation(edge)?;
+        let current = observation.identity;
+        let logs = observation.logs;
+        let last = observation.last;
+        if current.uid != initial.uid || current.restart_count != initial.restart_count {
+            return Err(format!(
+                "{edge} edge pod changed while waiting for serving revision {}; initial uid/restarts={}/{}, observed={}/{}, {}",
+                safe_truncate_str(revision, 16),
+                safe_truncate_str(&initial.uid, 12),
+                initial.restart_count,
+                safe_truncate_str(&current.uid, 12),
+                current.restart_count,
+                last,
+            )
+            .into());
+        }
+        if logs_prove_serving(&logs, revision) {
+            stable_observations = stable_observations.saturating_add(1);
+            if stable_observations >= 2 {
+                return Ok(());
+            }
+        } else {
+            stable_observations = 0;
+        }
+        if Instant::now() >= deadline {
+            return Err(format!(
+                "{edge} did not serve revision {} within {EDGE_SERVING_TIMEOUT:?}; last observed {}",
+                safe_truncate_str(revision, 16),
+                last
+            )
+            .into());
+        }
+        thread::park_timeout(EDGE_ACCEPTANCE_POLL);
+    }
+}
+
+/// A single atomic observation of edge identity and revision log state.
+struct EdgeServingObservation {
+    /// Pod identity and restart state observed alongside the logs.
+    identity: EdgePodIdentity,
+    /// Gateway logs used to prove accepted and serving revisions.
+    logs: String,
+    /// Sanitized revision summary retained for timeout diagnostics.
+    last: String,
+}
+
+/// Read one complete edge observation for the serving-revision barrier.
+fn edge_serving_observation(edge: &str) -> Result<EdgeServingObservation, Box<dyn std::error::Error>> {
+    let logs = edge_gateway_logs(edge)?;
+    let last = format!(
+        "accepted_revision={} serving_revision={}",
+        extract_last_accepted_revision(&logs).unwrap_or_else(|| "none".to_owned()),
+        extract_last_serving_revision(&logs).unwrap_or_else(|| "none".to_owned())
+    );
+    Ok(EdgeServingObservation {
+        identity: edge_pod_identity(edge)?,
+        logs,
+        last,
+    })
+}
+
 /// Testable core of the edge acceptance barrier.
 ///
 /// `log_source` is injected so unit tests can supply synthetic logs without
@@ -3361,9 +3467,18 @@ fn wait_for_edge_revision_acceptance(
 /// must appear as an exact quoted value, not as a substring of a longer field
 /// or an unrelated context.
 fn logs_prove_acceptance(logs: &str, revision: &str) -> bool {
-    logs.lines().any(|line| {
+    strip_csi_sgr(logs).lines().any(|line| {
+        (line.contains("overlay snapshot initialized") || line.contains("overlay reloaded"))
+            && accepted_revision_from_line(line) == Some(revision)
+    })
+}
+
+/// Check whether one log line proves the exact accepted and serving revision.
+fn logs_prove_serving(logs: &str, revision: &str) -> bool {
+    strip_csi_sgr(logs).lines().any(|line| {
         (line.contains("overlay snapshot initialized") || line.contains("overlay reloaded"))
             && accepted_revision_from_line(line) == Some(revision)
+            && serving_revision_from_line(line) == Some(revision)
     })
 }
 
@@ -3426,6 +3541,14 @@ fn extract_last_accepted_revision(logs: &str) -> Option<String> {
         .map(str::to_owned)
 }
 
+/// Extract the most recent exact `serving_revision` value from logs.
+fn extract_last_serving_revision(logs: &str) -> Option<String> {
+    logs.lines()
+        .rev()
+        .find_map(serving_revision_from_line)
+        .map(str::to_owned)
+}
+
 /// Kubernetes pod identity used to prove hot reload did not restart the edge.
 #[derive(Debug)]
 struct EdgePodIdentity {
@@ -4049,20 +4172,38 @@ fn wait_for_candidate_admission(
     Err(format!("timeout waiting for {provider} admission_state={expected} on {edge}; observed={observed:?}").into())
 }
 
-/// Wait until an unavailable provider is absent from the edge overlay.
-#[expect(
-    clippy::disallowed_methods,
-    reason = "synchronous polling in xtask; no async runtime is active"
-)]
-fn wait_for_candidate_absent(edge: &str, provider: &str) -> Result<String, Box<dyn std::error::Error>> {
-    let deadline = Instant::now() + PROVIDER_STATE_WAIT;
-    while Instant::now() < deadline {
-        if overlay_candidate(edge, provider)?.is_none() {
-            return Ok(format!("provider={provider} absent from operator-rendered overlay"));
-        }
-        thread::sleep(Duration::from_secs(1));
+/// Wait until the withdrawn provider is absent and the edge serves that exact
+/// projected overlay revision.
+pub(crate) fn wait_for_withdrawal_serving_revision(
+    edge: &str,
+    provider: &str,
+) -> Result<String, Box<dyn std::error::Error>> {
+    wait_for_check(
+        "withdrawal overlay candidate removal",
+        EDGE_SERVING_TIMEOUT,
+        EDGE_ACCEPTANCE_POLL,
+        || {
+            if overlay_candidate(edge, provider)?.is_some() {
+                return Err(format!("{edge} still contains withdrawn provider {provider}").into());
+            }
+            let revision = overlay_revision(edge)?;
+            verify_edge_serving_revision(edge, &revision)?;
+            Ok(revision)
+        },
+    )
+}
+
+/// Require every edge in the request path to serve its own withdrawn overlay.
+pub(crate) fn wait_for_all_edges_withdrawal_serving_revisions(
+    edges: &[&str],
+    provider: &str,
+) -> Result<BTreeMap<String, String>, Box<dyn std::error::Error>> {
+    let mut revisions = BTreeMap::new();
+    for edge in edges {
+        let revision = wait_for_withdrawal_serving_revision(edge, provider)?;
+        revisions.insert((*edge).to_owned(), revision);
     }
-    Err(format!("timeout waiting for {provider} removal from {edge} overlay").into())
+    Ok(revisions)
 }
 
 /// Return the `InferenceProvider` resource for one demo provider identity.
@@ -4343,8 +4484,8 @@ fn withdraw_provider(edge: &str, provider: &str) -> Result<(String, ProviderWith
         scale_provider_backend(provider, 0)?;
         refresh_provider(provider)?;
         let phase = wait_for_provider_phase(provider, "Unavailable")?;
-        let overlay = wait_for_candidate_absent(edge, provider)?;
-        Ok(format!("{phase}; {overlay}"))
+        let overlays = wait_for_all_edges_withdrawal_serving_revisions(EDGE_CLUSTERS, provider)?;
+        Ok(format!("{phase}; serving revisions={overlays:?}"))
     })();
     match result {
         Ok(evidence) => Ok((evidence, state)),
@@ -4484,12 +4625,17 @@ fn parse_header_dump(text: &str) -> BTreeMap<String, String> {
     map
 }
 
-/// Extract the `X-AI-Demo-Provider-Gateway` header from a response.
+/// Extract the most specific trusted provider identity from a response.
 fn extract_provider(resp: &verify::HttpResponse) -> Result<String, Box<dyn std::error::Error>> {
+    // Prefer the backend candidate identity when the provider gateway
+    // forwards it.  The gateway-site header is intentionally coarser: one
+    // site can host multiple candidates, so using it for withdrawal checks
+    // would confuse a withdrawn backend with a still-eligible sibling.
     resp.headers
-        .get("x-ai-demo-provider-gateway")
+        .get("x-grid-demo-provider")
+        .or_else(|| resp.headers.get("x-ai-demo-provider-gateway"))
         .cloned()
-        .ok_or_else(|| "missing x-ai-demo-provider-gateway header".into())
+        .ok_or_else(|| "missing provider attribution headers".into())
 }
 
 // ---------------------------------------------------------------------------
@@ -5314,6 +5460,19 @@ clusters:
         assert_eq!(result.ok().as_deref(), Some("east-provider"), "should extract provider");
     }
 
+    #[test]
+    fn extract_provider_prefers_backend_candidate_identity() {
+        let resp = verify::HttpResponse {
+            status: 200,
+            body: String::new(),
+            headers: BTreeMap::from([
+                ("x-ai-demo-provider-gateway".to_owned(), "east-provider".to_owned()),
+                ("x-grid-demo-provider".to_owned(), "east-provider-secondary".to_owned()),
+            ]),
+        };
+        assert_eq!(extract_provider(&resp).ok().as_deref(), Some("east-provider-secondary"));
+    }
+
     #[test]
     fn extract_provider_missing() {
         let resp = verify::HttpResponse {
@@ -5571,6 +5730,27 @@ clusters:
         assert!(logs_prove_acceptance(&cleaned, "rev_abc"));
     }
 
+    #[test]
+    fn serving_proof_requires_exact_accepted_and_serving_revision() {
+        let revision = "rev_abc";
+        let matching = format!("overlay reloaded accepted_revision=\"{revision}\" serving_revision=\"{revision}\"");
+        assert!(logs_prove_serving(&matching, revision));
+        assert!(!logs_prove_serving(
+            "overlay reloaded accepted_revision=\"rev_abc\" serving_revision=\"old\"",
+            revision
+        ));
+        assert!(!logs_prove_serving(
+            "overlay reloaded previous_serving_revision=\"rev_abc\"",
+            revision
+        ));
+    }
+
+    #[test]
+    fn serving_proof_accepts_ansi_and_unquoted_fields() {
+        let raw = "\x1b[32moverlay reloaded\x1b[0m \x1b[3maccepted_revision\x1b[0m=rev_abc serving_revision=rev_abc";
+        assert!(logs_prove_serving(raw, "rev_abc"));
+    }
+
     #[test]
     fn convergence_timeout_reports_no_accepted_revision() {
         let result = wait_for_edge_revision_acceptance(
diff --git a/xtask/src/env/glb_demo.rs b/xtask/src/env/glb_demo.rs
index 53c5710b..0380e57c 100644
--- a/xtask/src/env/glb_demo.rs
+++ b/xtask/src/env/glb_demo.rs
@@ -14,7 +14,7 @@ use serde::Serialize;
 use super::{
     DemoMode, GlbDemoModeOptions, GlbDemoOptions, IngressMode, certs,
     external_provider::{self, ExternalProviderDescriptor},
-    glb, gtm_emulator, image_overrides, kubectl, operator, workload,
+    glb, gtm_emulator, image_overrides, kubectl, operator, safe_truncate_str, workload,
 };
 
 #[cfg(test)]
@@ -1361,27 +1361,6 @@ fn read_overlay_candidates(edge: &str) -> Result<Vec<serde_json::Value>, Box<dyn
     Ok(candidates)
 }
 
-/// Poll until no candidates from `site` appear in the edge overlay.
-fn wait_for_overlay_candidate_absent(edge: &str, site: &str) -> Result<(), Box<dyn std::error::Error>> {
-    let deadline = Instant::now() + OVERLAY_CONVERGENCE_TIMEOUT;
-    loop {
-        let candidates = read_overlay_candidates(edge)?;
-        let has_site = candidates
-            .iter()
-            .any(|c| c.get("site").and_then(serde_json::Value::as_str) == Some(site));
-        if !has_site {
-            return Ok(());
-        }
-        if Instant::now() >= deadline {
-            return Err(format!(
-                "timeout: {edge} overlay still lists {site} candidates after {OVERLAY_CONVERGENCE_TIMEOUT:?}"
-            )
-            .into());
-        }
-        std::thread::park_timeout(OVERLAY_POLL_INTERVAL);
-    }
-}
-
 /// Poll until at least one candidate from `site` appears in the edge overlay.
 fn wait_for_overlay_candidate_present(edge: &str, site: &str) -> Result<(), Box<dyn std::error::Error>> {
     let deadline = Instant::now() + OVERLAY_CONVERGENCE_TIMEOUT;
@@ -1427,23 +1406,40 @@ fn wait_for_rollout(context: &str, name: &str) -> Result<(), Box<dyn std::error:
     Ok(())
 }
 
-/// Confirm overlay convergence, hot-reload, and remote fallback to a
-/// provider outside the drained site. Returns the fallback provider name.
+/// Confirm overlay convergence, hot-reload, and routing away from the exact
+/// withdrawn backend candidate. A sibling candidate in the same provider
+/// site remains valid; the assertion therefore uses backend attribution, not
+/// only the coarser provider-gateway site name.
+#[expect(
+    clippy::too_many_lines,
+    reason = "the fallback proof keeps the serving barrier and bounded request assertion together"
+)]
 fn verify_drain_fallback(
     consumer_edge: &str,
     reload_before: usize,
     narrator: &mut Narrator,
 ) -> Result<String, Box<dyn std::error::Error>> {
-    wait_for_overlay_candidate_absent(consumer_edge, "east-provider")?;
+    let withdrawal_revisions =
+        glb::wait_for_all_edges_withdrawal_serving_revisions(&["east-edge", "west-edge"], "east-provider")?;
+    let withdrawal_revision = withdrawal_revisions
+        .get(consumer_edge)
+        .ok_or_else(|| format!("missing withdrawal revision for {consumer_edge}"))?;
     narrator.narrate(&format!(
-        "  [OK] {consumer_edge} overlay no longer lists east-provider candidates"
+        "  [OK] {consumer_edge} serves withdrawal revision {} with no east-provider candidates",
+        safe_truncate_str(withdrawal_revision, 16)
     ));
 
     glb::check_hot_reload_observed(consumer_edge, reload_before)?;
     narrator.narrate(&format!("  [OK] {consumer_edge} gateway reloaded overlay after drain"));
 
-    let resp = glb::wait_for_data_plane_convergence("remote fallback routing", || {
-        let resp = workload::send_workload_request(consumer_edge, WORKLOAD_REQUEST_BODY, None)?;
+    let resp = glb::wait_for_data_plane_convergence("post-withdrawal routing", || {
+        let session = format!(
+            "withdrawal-proof-{}",
+            std::time::SystemTime::now()
+                .duration_since(std::time::UNIX_EPOCH)
+                .map_or(0, |duration| duration.as_nanos())
+        );
+        let resp = workload::send_workload_request(consumer_edge, WORKLOAD_REQUEST_BODY, Some(&session))?;
         if resp.status != 200 {
             return Err(format!(
                 "remote fallback: {consumer_edge} returned status {} after drain",
@@ -1454,9 +1450,9 @@ fn verify_drain_fallback(
         if resp.provider.is_empty() {
             return Err(format!("remote fallback: {consumer_edge} response missing provider header").into());
         }
-        if !resp.provider.starts_with("west") {
+        if resp.provider == "east-provider" {
             return Err(format!(
-                "remote fallback: {consumer_edge} routed to {}, expected a west-provider identity",
+                "post-withdrawal routing: {consumer_edge} routed to withdrawn backend {}",
                 resp.provider
             )
             .into());
@@ -1725,7 +1721,7 @@ fn restart_grid_operator(cluster: &str) -> Result<(), Box<dyn std::error::Error>
     if !output.status.success() {
         return Err(format!(
             "failed to restart {cluster} Grid operator: {}",
-            super::safe_truncate_str(&String::from_utf8_lossy(&output.stderr), 160)
+            safe_truncate_str(&String::from_utf8_lossy(&output.stderr), 160)
         )
         .into());
     }