diff --git a/src/ros2_medkit_fault_manager/src/correlation/correlation_engine.cpp b/src/ros2_medkit_fault_manager/src/correlation/correlation_engine.cpp
index f405375d..a6cea5d5 100644
--- a/src/ros2_medkit_fault_manager/src/correlation/correlation_engine.cpp
+++ b/src/ros2_medkit_fault_manager/src/correlation/correlation_engine.cpp
@@ -139,15 +139,47 @@ ProcessClearResult CorrelationEngine::process_clear(const std::string & fault_co
   if (cluster_it != fault_to_cluster_.end()) {
     const std::string cluster_id = cluster_it->second;
 
-    // TODO(#127): Clean up pending_clusters_ when fault is cleared.
-    // Currently, if a fault is cleared before cluster reaches min_count,
-    // the pending cluster retains a phantom reference. Low impact since
-    // pending clusters expire after window_ms, but could cause brief
-    // inconsistency if cluster becomes active with cleared fault in list.
-    // Fix requires: iteration (pending_clusters_ keyed by rule_id),
-    // representative reassignment if cleared fault was representative.
-
-    // Remove fault from cluster
+    // Clean up pending_clusters_ when fault is cleared before min_count (#127)
+    for (auto pending_it = pending_clusters_.begin(); pending_it != pending_clusters_.end();) {
+      auto & pending_cluster = pending_it->second.data;
+      if (pending_cluster.cluster_id != cluster_id) {
+        ++pending_it;
+        continue;
+      }
+
+      auto & codes = pending_cluster.fault_codes;
+      codes.erase(std::remove(codes.begin(), codes.end(), fault_code), codes.end());
+
+      if (codes.empty()) {
+        pending_it = pending_clusters_.erase(pending_it);
+        continue;
+      }
+
+      // Reassign representative if the cleared fault was the representative
+      if (pending_cluster.representative_code == fault_code) {
+        for (const auto & rule : config_.rules) {
+          if (rule.id == pending_it->first) {
+            switch (rule.representative) {
+              case Representative::FIRST:
+              case Representative::HIGHEST_SEVERITY:
+                // TODO(#213): HIGHEST_SEVERITY reassignment is approximate —
+                // PendingCluster lacks per-fault severity, so we fall back to
+                // first remaining fault. Store severities to fix.
+                pending_cluster.representative_code = codes.front();
+                break;
+              case Representative::MOST_RECENT:
+                pending_cluster.representative_code = codes.back();
+                break;
+            }
+            break;
+          }
+        }
+      }
+
+      ++pending_it;
+    }
+
+    // Remove fault from active cluster
     auto active_it = active_clusters_.find(cluster_id);
     if (active_it != active_clusters_.end()) {
       auto & codes = active_it->second.fault_codes;
diff --git a/src/ros2_medkit_fault_manager/test/test_correlation_engine.cpp b/src/ros2_medkit_fault_manager/test/test_correlation_engine.cpp
index e6d9c88f..b0da6800 100644
--- a/src/ros2_medkit_fault_manager/test/test_correlation_engine.cpp
+++ b/src/ros2_medkit_fault_manager/test/test_correlation_engine.cpp
@@ -541,6 +541,90 @@ TEST_F(CorrelationEngineTest, AutoClusterRetroactiveMuting) {
   EXPECT_EQ("SENSOR_001", clusters[0].representative_code);  // FIRST policy
 }
 
+// ============================================================================
+// Pending cluster cleanup on clear (#127)
+// ============================================================================
+
+TEST_F(CorrelationEngineTest, ClearFaultRemovesFromPendingCluster) {
+  auto config = create_auto_cluster_config();
+  CorrelationEngine engine(config);
+
+  auto t0 = std::chrono::steady_clock::now();
+
+  // Add 2 faults (below min_count=3), creating a pending cluster
+  engine.process_fault("MOTOR_COMM_FL", "ERROR", t0);
+  engine.process_fault("SENSOR_TIMEOUT", "ERROR", t0 + 10ms);
+  EXPECT_EQ(0u, engine.get_cluster_count());  // Still pending
+
+  // Clear one of the faults
+  engine.process_clear("MOTOR_COMM_FL");
+
+  // Now add a third fault - should NOT activate the cluster because
+  // the cleared fault was removed from pending, so only 2 faults total
+  engine.process_fault("DRIVE_COMM_ERROR", "ERROR", t0 + 20ms);
+  EXPECT_EQ(0u, engine.get_cluster_count());  // Still not enough
+}
+
+TEST_F(CorrelationEngineTest, ClearRepresentativeReassignsPendingCluster) {
+  const std::string yaml = R"(
+correlation:
+  enabled: true
+  patterns:
+    errors:
+      codes: ["*_ERROR"]
+  rules:
+    - id: test_cluster
+      mode: auto_cluster
+      match:
+        - pattern: errors
+      min_count: 3
+      window_ms: 500
+      show_as_single: true
+      representative: first
+)";
+  auto config = parse_config_string(yaml);
+  CorrelationEngine engine(config);
+
+  auto t0 = std::chrono::steady_clock::now();
+
+  // Add 2 faults, creating a pending cluster
+  engine.process_fault("FIRST_ERROR", "CRITICAL", t0);
+  engine.process_fault("SECOND_ERROR", "ERROR", t0 + 10ms);
+
+  // Clear the representative (first fault)
+  engine.process_clear("FIRST_ERROR");
+
+  // Add 2 more faults to reach min_count (SECOND_ERROR + 2 new = 3)
+  engine.process_fault("THIRD_ERROR", "ERROR", t0 + 20ms);
+  engine.process_fault("FOURTH_ERROR", "ERROR", t0 + 30ms);
+
+  EXPECT_EQ(1u, engine.get_cluster_count());
+  auto clusters = engine.get_clusters();
+  ASSERT_EQ(1u, clusters.size());
+  // SECOND_ERROR should be the new representative (first remaining after clear)
+  EXPECT_EQ("SECOND_ERROR", clusters[0].representative_code);
+}
+
+TEST_F(CorrelationEngineTest, ClearAllFaultsRemovesPendingCluster) {
+  auto config = create_auto_cluster_config();
+  CorrelationEngine engine(config);
+
+  auto t0 = std::chrono::steady_clock::now();
+
+  // Add 2 faults
+  engine.process_fault("MOTOR_COMM_FL", "ERROR", t0);
+  engine.process_fault("SENSOR_TIMEOUT", "ERROR", t0 + 10ms);
+
+  // Clear both
+  engine.process_clear("MOTOR_COMM_FL");
+  engine.process_clear("SENSOR_TIMEOUT");
+
+  // Adding 2 new faults should start a fresh pending cluster, not join old one
+  engine.process_fault("DRIVE_COMM_NEW", "ERROR", t0 + 100ms);
+  engine.process_fault("MOTOR_COMM_NEW", "ERROR", t0 + 110ms);
+  EXPECT_EQ(0u, engine.get_cluster_count());  // Still 2 faults, below min_count=3
+}
+
 int main(int argc, char ** argv) {
   ::testing::InitGoogleTest(&argc, argv);
   return RUN_ALL_TESTS();