diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index d7727f39e..1f3a1205f 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -7617,18 +7617,16 @@ impl GpuDqnTrainer {
         unsafe { (*self.q_readback_pinned.add(6)).clamp(0.0, 1.0) }
     }
 
-    /// Coarse spectral-gap approximation on the last Q readback.
+    /// Coarse Q-collapse detector derived from the single-sample Q readback at
+    /// `q_readback_pinned[7..7+total_actions]`.
     ///
-    /// Reads per-action Q values from q_readback_pinned[7..7+total_actions] (single
-    /// sample from the previous reduce_current_q_stats call) and returns
-    /// 1 + 1/range as a cheap proxy for sigma_1/sigma_2.
+    /// Returns a LARGE value (~100) when all Q values are (nearly) equal — i.e.
+    /// rank-1 / Q-collapse — and a value close to 1.0 when the Q spread is wide.
+    /// Intentionally coarse: a full SVD across a batch of Q samples would require
+    /// a separate larger buffer and is out of scope for the host-side health path.
     ///
-    /// NOTE: This is intentionally a coarse approximation rather than a full SVD.
-    /// A proper sigma_1/sigma_2 would require a batch-sized Q matrix and cuSOLVER,
-    /// which adds complexity and a sync. The proxy rises when Q is dominated by a
-    /// single action (rank-1-like), which is the collapse signal we care about.
-    /// The spectral_gap thresholds in NormalizedComponents::from_raw (2..10) are
-    /// calibrated for this proxy, not for true singular values.
+    /// The downstream `NormalizedComponents::from_raw` thresholds are calibrated
+    /// for this behaviour (`spectral_gap_norm` drops toward 0 as this value rises).
     pub fn compute_q_spectral_gap(&self) -> f32 {
         if self.q_readback_pinned.is_null() { return 1.0; }
         // Layout: q_readback_pinned[0..7] = stats, [7..7+total_actions] = per-action Q values.
diff --git a/crates/ml/src/trainers/dqn/trainer/training_loop.rs b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
index 7996577a5..71fa8c51c 100644
--- a/crates/ml/src/trainers/dqn/trainer/training_loop.rs
+++ b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
@@ -1785,8 +1785,10 @@ impl DQNTrainer {
 
         // ── LearningHealth computation (Layer 1) ──────────────────────────────
         // Source signal adaptations vs. plan spec (A3 known deviations):
-        //   - q_gap: mean of per_branch_q_gap_ema (4 branches) via fused_ctx — synchronous
-        //     DtoH at epoch boundary (not on the hot path; acceptable).
+        //   - q_gap: mean of per_branch_q_gap_ema (4 branches) via fused_ctx —
+        //     NOTE: per_branch_q_gap_ema() performs a synchronous stream flush + DtoH
+        //     (cuStreamSynchronize + cuMemcpyDtoH_v2). Acceptable at epoch boundary
+        //     because it runs once per epoch; DO NOT move into a per-step loop.
         //   - q_var: epoch_q_max - epoch_q_min proxy instead of QValueStatsResult.q_variance
         //     (q_variance is not separately tracked as a scalar field; range is a sufficient proxy).
         //   - grad_consistency: scalar proxy (successive grad_norm delta) instead of full Adam
@@ -1804,8 +1806,14 @@ impl DQNTrainer {
                 0.0
             };
 
-            // Q-var proxy from epoch min/max (accumulated by update_epoch_q_stats each step).
-            let q_var_signal = (self.epoch_q_max - self.epoch_q_min).max(0.0);
+            // q_var proxy: range (max - min) of per-step Q values this epoch. If no
+            // steps ran (epoch_q_min still INFINITY), fall back to 0.0 rather than
+            // propagating NaN/-inf into health normalization.
+            let q_var_signal = if self.epoch_q_min.is_finite() && self.epoch_q_max.is_finite() {
+                (self.epoch_q_max - self.epoch_q_min).max(0.0)
+            } else {
+                0.0
+            };
 
             // Atom utilization from q_readback_pinned[6] (one-step lag).
             let atom_util = self.fused_ctx.as_ref()