fix(kvbm): prevent scheduler crash from race between abort and request completion

lib/bindings/kvbm/src/block_manager/vllm/connector/leader.rs

@@ -549,7 +549,21 @@ impl Leader for KvConnectorLeader {
                 "request_finished called for request_id: {request_id} but slot is not found"
             );
             self.inflight_requests.remove(&request_id);
-            return Ok(false);
+            // We must return `true` here even though the leader slot is gone.
+            //
+            // Within a single call to vLLM's `update_from_output()`, two things
+            // happen in sequence:
+            //   1. Stopped requests call _free_request() → request_finished() (here)
+            //   2. Worker's finished_sending is processed by _update_from_kv_xfer_finished()
+            //
+            // The worker's get_finished() ran during the forward pass and may have
+            // reported this request in finished_sending for this same step. If we
+            // return `false`, vLLM deletes the request from self.requests at step 1.
+            // Then step 2 hits `assert req_id in self.requests` and crashes.
+            //
+            // Returning `true` keeps the request in self.requests so step 2 can
+            // process finished_sending and call _free_blocks() properly.
+            return Ok(true);
         }
 
         // grab the slot

lib/bindings/kvbm/src/block_manager/vllm/connector/worker.rs

@@ -409,11 +409,15 @@ impl Worker for KvConnectorWorker {
                     tracing::debug!(request_id, "request slot is not finished");
                 }
             } else {
-                // made this condition more strict slot existence checks were added as a prerequesite
-                // to be added to the maybe_finished_offloading set.
-                panic!(
-                    "request slot missing for {request_id}; however, it was present when added to the maybe finished offloading set"
+                // Slot was removed between when we added it to
+                // maybe_finished_offloading and now. Signal completion so
+                // vLLM can free the request via _free_blocks().
+                tracing::warn!(
+                    request_id,
+                    "request slot missing from maybe_finished_offloading set; \
+                     signaling completion"
                 );
+                is_finished_offloading.insert(request_id.clone());
             }
         }
 
@@ -443,9 +447,15 @@ impl Worker for KvConnectorWorker {
                     tracing::debug!(request_id, "request slot is not finished");
                 }
             } else {
-                panic!(
-                    "request slot missing for {request_id}; however, it was present when added to the maybe finished onboarding set"
+                // Slot was removed between when we added it to
+                // maybe_finished_onboarding and now. Signal completion so
+                // vLLM can free the request via _free_blocks().
+                tracing::warn!(
+                    request_id,
+                    "request slot missing from maybe_finished_onboarding set; \
+                     signaling completion"
                 );
+                is_finished_onboarding.insert(request_id.clone());
             }
         }