From efb4e529a4af97a03a1194bb9671900e97f001fd Mon Sep 17 00:00:00 2001
From: jgrusewski <jgrusewski@localhost>
Date: Sat, 25 Apr 2026 00:51:10 +0200
Subject: [PATCH] plan(dqn-v2): mark Task 6a, 6b LANDED with iteration history

---
 .../plans/2026-04-24-dqn-v2-plan-3-behavioural.md      | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/docs/superpowers/plans/2026-04-24-dqn-v2-plan-3-behavioural.md b/docs/superpowers/plans/2026-04-24-dqn-v2-plan-3-behavioural.md
index 6486b2a43..033a04643 100644
--- a/docs/superpowers/plans/2026-04-24-dqn-v2-plan-3-behavioural.md
+++ b/docs/superpowers/plans/2026-04-24-dqn-v2-plan-3-behavioural.md
@@ -945,7 +945,11 @@ Fingerprint shifts to `[79,80]`; `ISV_TOTAL_DIM` grows from 78 (post-Task-4) to
 
 > **IMPORTANT — task ordering**. Task 6 must land AFTER Task 4. If Task 4 is skipped, Task 6's tail-start index drops to 75. Recompute slot constants from the current `ISV_TOTAL_DIM` at implementation time — never hard-code from this plan.
 
-### Task 6a: Persistence credit (held through drawdown, took profit)
+### Task 6a: Persistence credit — ✅ LANDED `777341776`
+
+> **Actual outcome.** `PS_INTRA_TRADE_MIN_PNL = 39` added; `PS_STRIDE` grew 39 → 40 with 6 hardcoded-stride sites + 7 comment mentions migrated in lockstep. MIN_PNL tracked per bar alongside MAX_PNL; reset at 5 MAX_PNL reset sites. Consumer fires on `segment_complete && reward > 0 && drawdown_depth > 1e-6f`, adds `shaping × conviction × |min_pnl| × tanh(reward/|min_pnl|)` to `reward` and `rc[5]`. Smoke fold-2 best Sharpe jumped 84.44 → 100.10; HEALTH_DIAG `reward_split bonus=17.21` confirms firing.
+
+#### Original design (for reference)
 
 **Files:**
 - Modify: `crates/ml/src/cuda_pipeline/state_layout.cuh` — add 1 PS slot `PS_INTRA_TRADE_MIN_PNL`; grow `PS_STRIDE` 39 → 40.
@@ -966,6 +970,10 @@ Fingerprint shifts to `[79,80]`; `ISV_TOTAL_DIM` grows from 78 (post-Task-4) to
   ```
 - Self-scaling rationale: `tanh(reward / drawdown_depth)` saturates at ±1 when reward ≫ drawdown (full credit) and approaches 0 when reward ≈ drawdown (no surprise). No tuned coefficients.
 
+### Task 6b: Regime-shift penalty — ✅ LANDED `5ebebc564`
+
+> **Actual outcome.** `PS_REGIME_SHIFT_BAR = 40` added; `PS_STRIDE` grew 40 → 41 via 6 stride sites. Detector fires first bar where `|regime_now - PS_PLAN_ENTRY_REGIME|` exceeds `clamp(0.25 × |clamp(sharpe, −2, 2)|, 0.05, 0.5)`. Consumer on exit: `penalty = shaping × conviction × bars_late_frac × |reward|` (bounded by |reward|). Smoke fold-2 best Sharpe: 117.92. **Iteration note:** first formula multiplied by `ISV[Q_DIR_ABS_REF] × |reward|` simultaneously — absolute Q-magnitude (5–50) amplified |reward| by 20×, flipping trade outcomes and destabilising training. Fixed by dropping `q_scale` multiplier. Pearl: pick exactly ONE unbounded scaling signal per reward term; bound everything else.
+
 ### Task 6b: Regime-shift penalty (held past regime flip)
 
 **Files:**