spec: OFI momentum v3 — add raw_open + mid_price_open from MBP-10

target_dim 4→6: adds raw_open (OHLCV) and mid_price_open (MBP-10
best bid/ask midpoint at bar formation) to the targets buffer.

Micro-reward now uses real intra-bar move (close - open) / atr
instead of close-to-close approximation. Also adds spread cost
awareness via |open - mid| penalty. Requires fxcache recomputation.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs/superpowers/specs/2026-04-19-ofi-momentum-dense-reward-design.md

@@ -52,7 +52,18 @@ concat_ofi_features reads batch_states[b * SD + 66] for order branch (was 42)
 
 **Location:** `experience_kernels.cu` env_step_batch, replacing lines 1715-1721.
 
-**Formula (revised — no open_price, uses close-to-close):**
+**Targets buffer expansion (target_dim 4→6):**
+```
+Current: tgt[0]=preproc_close, tgt[1]=preproc_next, tgt[2]=raw_close, tgt[3]=raw_next
+New:     tgt[0]=preproc_close, tgt[1]=preproc_next, tgt[2]=raw_close, tgt[3]=raw_next,
+         tgt[4]=raw_open, tgt[5]=mid_price_open
+```
+- `raw_open`: OHLCV open price (already parsed in data_loading.rs:650)
+- `mid_price_open`: MBP-10 best_bid/best_ask midpoint at bar formation — the true entry price
+
+Requires fxcache recomputation (one-time Argo job: `precompute_features` binary).
+
+**Formula:**
 ```c
 if (!segment_complete && fabsf(position) > 0.001f && micro_reward_scale > 0.0f) {
     /* Read current OFI from batch_states (wired in Phase A) */
@@ -67,16 +78,22 @@ if (!segment_complete && fabsf(position) > 0.001f && micro_reward_scale > 0.0f)
 
     float flow_momentum = sign_pos * 0.25f * (delta_depth + delta_vpin + delta_queue + delta_tarr);
 
-    /* Price confirmation: close-to-close return (no lookahead) */
-    float prev_close = ps[PREV_CLOSE_SLOT];  /* stored from previous bar */
-    float bar_return = (prev_close > 0.0f) ? (raw_close - prev_close) / fmaxf(prev_close, 1.0f) : 0.0f;
-    float price_confirm = sign_pos * bar_return / fmaxf(atr_pct, 1e-6f);
+    /* Price confirmation: intra-bar directional move from MBP-10 data */
+    float raw_open = tgt[4];           /* bar open price from OHLCV */
+    float mid_open = tgt[5];           /* mid-price at bar formation from MBP-10 */
+    float bar_move = (raw_close - raw_open) / fmaxf(atr_abs, 1e-6f);
+    float price_confirm = sign_pos * bar_move;
+
+    /* Spread cost awareness: distance from mid to open = half-spread at entry */
+    float spread_at_entry = fabsf(raw_open - mid_open);
 
     /* Informed flow intensity */
     float vol_informed = ofi_cur[6] * ofi_cur[7];  /* VPIN × Kyle's_lambda */
 
-    /* Composite quality */
-    float quality = flow_momentum * (1.0f + vol_informed) + 0.5f * price_confirm;
+    /* Composite quality — spread-aware */
+    float quality = flow_momentum * (1.0f + vol_informed)
+                  + 0.5f * price_confirm
+                  - 0.1f * spread_at_entry / fmaxf(atr_abs, 1e-6f);  /* penalize wide spreads */
     reward = micro_reward_scale * tanhf(quality / 3.0f);
 }
 
@@ -85,11 +102,8 @@ if (market_dim >= 50) {
     const float* ofi_cur = batch_states + (long long)i * state_dim + 66;
     for (int k = 0; k < 8; k++) ps[30 + k] = ofi_cur[k];
 }
-ps[PREV_CLOSE_SLOT] = raw_close;  /* store close for next bar's price_confirm */
 ```
 
-**PREV_CLOSE_SLOT:** Use slot 6 (currently reserved). Define `#define PREV_CLOSE_SLOT 6`.
-
 **Note:** `batch_states` must be passed to `env_step_batch` as a new kernel parameter. Currently the kernel receives `features` (42-dim market features) but NOT the full `batch_states` (96-dim with OFI). Add `const float* __restrict__ batch_states` and `int state_dim` parameters.
 
 ### Phase D: Rank Normalization Fix
@@ -237,7 +251,7 @@ No code change needed — PopArt handles this correctly by design.
 
 | File | Change |
 |------|--------|
-| `experience_kernels.cu` | PORTFOLIO_STRIDE=38, micro-reward body, OFI delta storage, batch_states param |
+| `experience_kernels.cu` | PORTFOLIO_STRIDE=38, micro-reward body with tgt[4]/tgt[5], OFI delta storage, batch_states param |
 | `gpu_experience_collector.rs` | Wire ofi_gpu into state_gather, pass batch_states to env_step |
 | `gpu_dqn_trainer.rs` | Fix concat_ofi indices 42→66, OFI embed MLP, wider Mamba2 params, d_h_history GEMMs |
 | `mamba2_temporal_kernel.cu` | Update history kernel to copy SH2+8 dims with ofi_embed param |
@@ -248,6 +262,10 @@ No code change needed — PopArt handles this correctly by design.
 | `reward_shaping_kernel.cu` | Lower is_trade threshold to 1e-5 |
 | `dqn-production.toml` | micro_reward_scale = 0.1 |
 | `constructor.rs` | State layout docs, OFI delta indices 74-81 |
+| `fxcache.rs` | target_dim 4→6, add raw_open + mid_price_open to targets |
+| `data_loading.rs` | Populate tgt[4]=raw_open, tgt[5]=mid_price_open from OHLCV + MBP-10 |
+| `precompute_features` binary | Rebuild fxcache with 6-slot targets |
+| `train_baseline_rl.rs` / `evaluate_baseline.rs` | Update TARGET_DIM constant from 4 to 6 |
 
 ### Config Changes