fix(dqn): invert adaptive gamma direction on low atom utilization

Val-Flat-collapse fix #4 (task #94, 2026-04-24). Research-agent audit
of train-bv2n5 identified a positive-feedback loop driving atom
utilisation to 1%:

    util < 0.4 → γ -= 0.01 (toward 0.90 floor)
    γ × delta_z shrinks → TD targets concentrate on fewer bins
    → more collapse → util lower → γ lower ...

The original controller had the right structure but the wrong
direction on the low-util branch. Flipping the sign (`-=` → `+=`)
breaks the loop: when atoms collapse, raising γ widens the effective
TD-target support `γ × atom_support`, pushing targets across more
bins of the C51 grid and giving the network more distributional
signal to learn from.

Evidence from train-bv2n5 (pre-fix): atom_util stuck at 0.01 for 16
epochs; γ slowly drifted toward 0.90. No mechanism recovered.

Change localised to a single `else if util < 0.4` branch at
`training_loop.rs:660-675`. Same step size (0.005) as the healthy
branch avoids overshoot on the recovery path. Clamps retained
[0.90, 0.95] at this layer; the fused trainer re-clamps to
[0.90, 0.995] after regime adjustment downstream.

Other pending atom-util fixes from the audit (deferred pending
validation):
- Atom-entropy regularisation term in C51 loss (medium risk)
- Absolute v_half floor in update_eval_v_range (low risk)
- TD-target dithering before Bellman projection (low risk)
This single-line fix addresses the dominant mechanism — research
agent labelled it "lowest risk, highest leverage". If util remains
low after this, the others stack additively.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -654,7 +654,29 @@ impl DQNTrainer {
                 }
             }
 
-            // G4: Gamma annealing — adaptive horizon based on atom utilization
+            // G4: Gamma annealing — adaptive horizon based on atom utilization.
+            //
+            // Prior direction (pre-2026-04-24 val-Flat-collapse fix #4):
+            //   util > 0.6 → γ += 0.005  (extend horizon when healthy)
+            //   util < 0.4 → γ -= 0.010  (shorten horizon when collapsed)
+            //
+            // That second arm creates a positive-feedback loop: atom
+            // collapse (low util) → lower γ → `γ × delta_z` shrinks →
+            // TD targets concentrate on fewer bins → deeper collapse.
+            // Observed on train-bv2n5: atom util stuck at 1% for 16
+            // epochs while this controller quietly nudged γ toward its
+            // 0.90 floor, reinforcing the symptom it was meant to
+            // mitigate. Per research-agent audit (2026-04-24).
+            //
+            // New direction: when atoms are collapsed, RAISE γ so
+            // `γ × atom_support` spans a larger fraction of the grid,
+            // spreading TD targets across more bins and breaking the
+            // loop. Slower step (0.005) than the healthy-branch to
+            // avoid overshoot.
+            //
+            // Clamps retained: [0.90, 0.95] at this layer; the fused
+            // trainer re-clamps to [0.90, 0.995] after its regime
+            // adjustment.
             {
                 let util = self.fused_ctx.as_ref()
                     .map(|f| f.utilization_ema())
@@ -662,7 +684,9 @@ impl DQNTrainer {
                 if util > 0.6 {
                     self.adaptive_gamma = (self.adaptive_gamma + 0.005).min(0.95);
                 } else if util < 0.4 {
-                    self.adaptive_gamma = (self.adaptive_gamma - 0.01).max(0.90);
+                    // Direction flipped (atom-collapse fix): RAISE γ to
+                    // widen the TD-target spread across the atom grid.
+                    self.adaptive_gamma = (self.adaptive_gamma + 0.005).min(0.95);
                 }
                 // C4/P4: Wire adaptive gamma into GPU trainer with temporal-coupled adjustment.
                 // apply_adaptive_gamma reads health+regime from ISV, adjusts, clamps [0.9, 0.995],