Guard large unaligned Qwen expert gathers

src/engine/metal/qwen_mtplx/session_checkpoint.rs

@@ -32,6 +32,8 @@ struct Manifest {
     #[serde(default)]
     verify_gate: bool,
     #[serde(default)]
+    sorted_gather_guard: bool,
+    #[serde(default)]
     images: Vec<crate::engine::qwen_vision::Identity>,
     tokens: Vec<i32>,
     lazy_kv: bool,
@@ -119,6 +121,7 @@ impl LiveSession {
             let offset = file.stream_position().map_err(|e| e.to_string())?;
             let manifest = serde_json::to_vec(&Manifest {
                 verify_gate: true,
+                sorted_gather_guard: true,
                 identity,
                 tag,
                 images: self.current_images.clone(),
@@ -193,6 +196,14 @@ impl LiveSession {
                 "Qwen checkpoint uses obsolete MTP attention-gate math; rebuild the cache".into(),
             );
         }
+        // Old files do not record prefill chunk sizes or producer hardware.
+        // At ten routes/token this length can contain an affected chunk;
+        // shorter sessions cannot. Reject before allocating/restoring state.
+        if !manifest.sorted_gather_guard && manifest.tokens.len() >= 3277 {
+            return Err(
+                "Qwen checkpoint predates the sorted-gather row guard; rebuild the cache".into(),
+            );
+        }
         if !manifest.payload["mtp_history_snapshot"].is_null() != identity.mtp {
             return Err("Qwen checkpoint MTP payload differs".into());
         }
@@ -480,6 +491,40 @@ fn qwen_checkpoint_roundtrip_and_failed_restore_preserve_live_state() {
             .unwrap()
     );
     assert_eq!(fingerprint(&live), ar_original);
+    let bytes = fs::read(&path).unwrap();
+    let offset = u64::from_le_bytes(bytes[8..16].try_into().unwrap()) as usize;
+    let mut manifest: serde_json::Value = serde_json::from_slice(&bytes[offset..]).unwrap();
+    manifest
+        .as_object_mut()
+        .unwrap()
+        .remove("sorted_gather_guard");
+    let write_manifest = |manifest: &serde_json::Value| {
+        let mut old = bytes[..offset].to_vec();
+        old.extend(serde_json::to_vec(manifest).unwrap());
+        fs::write(&broken, old).unwrap();
+    };
+    write_manifest(&manifest);
+    assert!(
+        live.load_checkpoint(&broken, ar_identity, &streams, stream, |_| {}, || false)
+            .unwrap()
+    );
+    assert_eq!(fingerprint(&live), ar_original);
+    manifest["tokens"] = serde_json::json!(vec![1; 3277]);
+    manifest["identity"]["context"] = 4096.into();
+    write_manifest(&manifest);
+    let large_identity = Identity {
+        context: 4096,
+        ..ar_identity
+    };
+    assert!(
+        live.load_checkpoint(&broken, large_identity, &streams, stream, |_| {}, || false)
+            .unwrap_err()
+            .contains("sorted-gather row guard")
+    );
+    assert_eq!(fingerprint(&live), ar_original);
+    println!(
+        "Qwen sorted-gather checkpoint policy: old short state exact, possibly affected long state rejected before restore"
+    );
     println!(
         "Qwen checkpoint: exact roundtrip, cancellation/corruption controls, old MTP rejected and old AR restored"
     );