diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index d349a2f61..b447d49a5 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -7957,9 +7957,12 @@ impl GpuDqnTrainer {
         if ring.snapshots.len() >= crate::cuda_pipeline::q_snapshot::MAX_SNAPSHOTS {
             let worst_idx = ring.snapshots.iter()
                 .enumerate()
-                .min_by(|(_, a), (_, b)| a.q_gap.partial_cmp(&b.q_gap).unwrap())
+                .min_by(|(_, a), (_, b)| {
+                    // NaN-safe: NaN q_gaps compare as Equal instead of panicking.
+                    a.q_gap.partial_cmp(&b.q_gap).unwrap_or(std::cmp::Ordering::Equal)
+                })
                 .map(|(i, _)| i)
-                .unwrap();
+                .expect("MAX_SNAPSHOTS >= 1 and len >= MAX_SNAPSHOTS, so at least one snapshot exists");
             ring.snapshots.swap_remove(worst_idx);
         }
 
diff --git a/crates/ml/src/cuda_pipeline/q_snapshot.rs b/crates/ml/src/cuda_pipeline/q_snapshot.rs
index 7cae71a92..6b0ab6fd9 100644
--- a/crates/ml/src/cuda_pipeline/q_snapshot.rs
+++ b/crates/ml/src/cuda_pipeline/q_snapshot.rs
@@ -107,9 +107,13 @@ impl SnapshotRing {
         if self.snapshots.len() >= MAX_SNAPSHOTS {
             let worst_idx = self.snapshots.iter()
                 .enumerate()
-                .min_by(|(_, a), (_, b)| a.q_gap.partial_cmp(&b.q_gap).unwrap())
+                .min_by(|(_, a), (_, b)| {
+                    // NaN-safe: treat NaN q_gaps as equal so a snapshot with NaN
+                    // does not panic the swap-out path.
+                    a.q_gap.partial_cmp(&b.q_gap).unwrap_or(std::cmp::Ordering::Equal)
+                })
                 .map(|(i, _)| i)
-                .unwrap();
+                .expect("MAX_SNAPSHOTS >= 1 and len >= MAX_SNAPSHOTS, so at least one snapshot exists");
             self.snapshots.swap_remove(worst_idx);
         }
 
@@ -120,7 +124,7 @@ impl SnapshotRing {
     /// Best (highest q_gap) snapshot raw device pointer, if any.
     pub fn best_raw_ptr(&self) -> Option<u64> {
         self.snapshots.iter()
-            .max_by(|a, b| a.q_gap.partial_cmp(&b.q_gap).unwrap())
+            .max_by(|a, b| a.q_gap.partial_cmp(&b.q_gap).unwrap_or(std::cmp::Ordering::Equal))
             .map(|s| s.weights.raw_ptr())
     }
 }