fix(dqn): unstick eval Kelly cap — health-coupled warmup_floor never collapses to zero

The eval-mode policy was producing diverse Boltzmann picks (verified by the
new `val_dir_dist` HEALTH_DIAG line) but every active-direction pick (Long /
Short) was collapsing to `actual_dir = Flat` because the Kelly cap forced
`target_position = 0` at cold start.

Cluster run `train-multi-seed-ddrpr` epoch 0 made this unambiguous:
  val_dir_dist [short=0.0000 hold=0.1953 long=0.0001 flat=0.8047]

Boltzmann fired correctly (sum hold + flat ≈ 100% of bars, with Hold ~ 20% =
the share of bars where the policy explicitly picked Hold; the other 80%
were active-direction picks all rerouted to Flat by `target_position = 0`).

Root cause in `trade_physics.cuh::kelly_position_cap`:
  warmup_floor = clamp(conviction, 0, 1)            // ← can hit 0
  effective_kelly = maturity*kelly_f + (1-maturity)*warmup_floor
                  = 0 + 1*0 = 0   at cold start with low conviction
  cap = effective_kelly * max_position * safety = 0 → no exposure permitted

The `safety_multiplier` was already protected by a `health_safety = 0.5 + 0.5×h`
floor, but `warmup_floor` had no such floor. Catch-22: low conviction → cap=0 →
no trades → Kelly stats stay cold → conviction stays low → forever.

The same bootstrap-deadlock pattern as the IQN trunk SAXPY readiness gate
(commit f86353840), and the fix is structurally identical — apply a non-zero
adaptive floor sourced from the same training-stability signal:

  warmup_floor = max(conviction, health_floor)

where `health_floor = 0.5 + 0.5 × ISV[LEARNING_HEALTH]` is the same value the
caller already computes for `safety_multiplier`. Both signals are adaptive
and ISV-driven; no tuned constants. The floor only matters during cold start
— once `maturity → 1` after ≥10 trades the term drops out entirely.

Threaded through both `apply_kelly_cap` and `kelly_position_cap` signatures;
single caller in `unified_env_step_core` passes `health_safety` as the new
arg (already locally computed two lines above). Build clean at 11-warning
baseline.

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

crates/ml/src/cuda_pipeline/trade_physics.cuh

@@ -269,7 +269,8 @@ __device__ __forceinline__ float kelly_position_cap(
     float sum_losses,
     float max_position,
     float safety_multiplier,
-    float conviction
+    float conviction,
+    float health_floor
 ) {
     const float prior_wins = 2.0f;
     const float prior_losses = 2.0f;
@@ -316,7 +317,22 @@ __device__ __forceinline__ float kelly_position_cap(
      * doesn't swing wildly on small samples. */
     float total_trades = win_count + loss_count;
     float maturity = fminf(1.0f, total_trades * 0.1f);
-    float warmup_floor = fminf(1.0f, fmaxf(0.0f, conviction));
+    /* warmup_floor must never collapse to zero: at cold start (maturity=0) the
+     * effective Kelly cap reduces to `(1 - maturity) * warmup_floor = warmup_floor`,
+     * so a zero floor pins target_position to 0 and the policy can never open a
+     * position to accumulate the trades that would graduate Kelly stats — the
+     * same bootstrap deadlock pattern that gated the IQN trunk SAXPY. The
+     * adaptive floor is therefore the max of (a) per-sample policy conviction
+     * and (b) the caller's training-stability floor (`health_floor` = 0.5 +
+     * 0.5 × health, ∈ [0.5, 1.0], computed from ISV[LEARNING_HEALTH]). When
+     * the policy is confident, conviction wins; when training is healthy but
+     * Q-spread is tight, the health floor wins. Both are adaptive ISV-driven
+     * signals; no tuned constants. The floor only matters during cold start —
+     * once `maturity → 1` (≥10 completed trades) the term drops out entirely. */
+    float warmup_floor = fmaxf(
+        fminf(1.0f, fmaxf(0.0f, conviction)),
+        fminf(1.0f, fmaxf(0.0f, health_floor))
+    );
     float effective_kelly = maturity * kelly_f + (1.0f - maturity) * warmup_floor;
 
     return effective_kelly * max_position * safety_multiplier;
@@ -334,10 +350,12 @@ __device__ __forceinline__ float apply_kelly_cap(
     float sum_losses,
     float max_position,
     float safety_multiplier,
-    float conviction
+    float conviction,
+    float health_floor
 ) {
     float cap = kelly_position_cap(win_count, loss_count, sum_wins, sum_losses,
-                                   max_position, safety_multiplier, conviction);
+                                   max_position, safety_multiplier, conviction,
+                                   health_floor);
     return fmaxf(-cap, fminf(cap, target_position));
 }
 
@@ -648,7 +666,8 @@ __device__ __forceinline__ void unified_env_step_core(
             *win_count, *loss_count,
             *sum_wins,  *sum_losses,
             max_position_physics, safety,
-            conviction
+            conviction,
+            health_safety
         );
     }
 

docs/dqn-wire-up-audit.md

@@ -2,6 +2,22 @@
 
 **Status:** Populated during Plan 1 Task 6 (A.5 orphan audit). Updated on every commit per Invariant 7.
 
+Kelly cap warmup_floor health-coupled (2026-04-26): `trade_physics.cuh::kelly_position_cap`
+previously computed `warmup_floor = clamp(conviction, 0, 1)`. At cold start
+(maturity = 0) `effective_kelly = warmup_floor`, so a zero conviction collapsed
+the cap to zero — eval-mode could never open a position to graduate Kelly stats,
+the same bootstrap deadlock pattern the IQN trunk SAXPY exhibited. Confirmed by
+cluster run `train-multi-seed-ddrpr` epoch 0: new `val_dir_dist [short=0.0000
+hold=0.1953 long=0.0001 flat=0.8047]` shows Boltzmann fired correctly (hold +
+flat ≈ 100% of bars) but every active-direction pick (Long/Short) was forced
+to actual_dir = Flat by `target_position = 0`. Fix: `warmup_floor = max(conviction,
+health_floor)` where `health_floor = 0.5 + 0.5 × health` is the same training-
+stability floor already used by `safety_multiplier` (composed in
+`unified_env_step_core` from `ISV[LEARNING_HEALTH]`). Both signals are adaptive
+and ISV-driven; the floor only matters during cold start (maturity → 1 after
+≥10 trades drops the term out entirely). Threaded through both `apply_kelly_cap`
+and `kelly_position_cap` signatures.
+
 Eval-mode action selection unified with Boltzmann softmax (2026-04-26):
 `experience_kernels.cu` — the four factored action heads (direction, magnitude,
 order, urgency) all dropped their `else if (eval_mode)` strict-argmax + ISV-