perf(rl): scale default batch size to 128 + auto-size PER capacity

n-backtests default in Argo template: 16 → 128.
PER capacity: max(configured, 4 × b_size) so replay buffer is always
large enough for useful prioritized sampling at any batch size.

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

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

@@ -1676,8 +1676,11 @@ impl IntegratedTrainer {
         // owned by each `Transition`). Sample → gather populates the
         // trainer-owned sampled_* buffers via per-batch DtoD copies
         // out of the buffer's transitions.
+        // Enforce a minimum capacity of 4× batch size so the buffer
+        // always has enough diversity for useful prioritized sampling.
+        let per_capacity = cfg.per_capacity.max(4 * b_size);
         let replay = crate::rl::replay::ReplayBuffer::new(
-            cfg.per_capacity,
+            per_capacity,
             cfg.per_seed,
         );
         let sampled_h_t_d = stream

infra/k8s/argo/alpha-rl-template.yaml

@@ -9,7 +9,7 @@
 #   argo submit -n foxhunt --from=wftmpl/alpha-rl \
 #     -p git-branch=ml-alpha-phase-a
 #   argo submit -n foxhunt --from=wftmpl/alpha-rl \
-#     -p git-branch=ml-alpha-phase-a -p n-steps=50000 -p n-backtests=16
+#     -p git-branch=ml-alpha-phase-a -p n-steps=50000 -p n-backtests=128
 apiVersion: argoproj.io/v1alpha1
 kind: WorkflowTemplate
 metadata:
@@ -42,7 +42,7 @@ spec:
       - name: n-steps
         value: "50000"
       - name: n-backtests
-        value: "16"
+        value: "128"
       - name: seed
         value: "16962"
       - name: seq-len