fix(rl): remove double-divide on l_q reading (cosmetic, matches l_pi)

dqn_distributional_q.cu:285 atomicAdds ce/B into ss_q_loss (kernel-side
B divide, matching ppo_clipped_surrogate.cu:269-271's b_inv pattern).
Host previously divided by B AGAIN at integrated.rs:3944 and :5196 —
true mean CE (≈ 3.0 for 21-atom softmax) appeared as 3.0/1024 ≈ 0.003.

The "baseline l_q=1.45 at step 500" was actually a missing-zero-out
artifact: ss_q_loss accumulated 500 × 0.003 = 1.46 across steps before
bf8887400 fixed the zero-out, exposing the underlying double-divide.

Q learning has been working correctly the whole time — only the
diagnostic display was wrong.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

crates/ml-alpha/src/trainer/integrated.rs

@@ -3941,7 +3941,11 @@ impl IntegratedTrainer {
             self.last_pi_loss = l_pi_host;
 
             let l_q_host = unsafe { std::ptr::read_volatile(self.ss_q_loss_mapped.host_ptr) };
-            self.last_q_loss = l_q_host / (b_size as f32);
+            // Kernel already divides by B in atomicAdd (dqn_distributional_q.cu:285).
+            // The previous host divide was a double-divide bug — true mean CE
+            // appeared as mean/B ≈ 0.003 instead of ≈ 3.0. Matches l_pi convention
+            // (ppo_clipped_surrogate.cu:269-271 also kernel-side divides).
+            self.last_q_loss = l_q_host;
 
             let l_v_sum_host = unsafe { std::ptr::read_volatile(self.ss_v_loss_sum_mapped.host_ptr) };
             let l_v_host = l_v_sum_host / (b_size as f32);
@@ -5193,7 +5197,8 @@ impl IntegratedTrainer {
         // is zero = sentinel, and the per-branch LR controller's cold-
         // start gate handles that.
         let l_q_host = unsafe { std::ptr::read_volatile(self.ss_q_loss_mapped.host_ptr) };
-        let l_q = l_q_host / (b_size as f32).max(1.0);
+        // Kernel divides by B (see step_synthetic_body comment). No host divide.
+        let l_q = l_q_host;
 
         self.dqn_head
             .backward_gemm(