fix(ml): OFI preload uses adapter's load_ofi_features, not trainer's empty field

Root cause: preload_data() called loader.ofi_features.take() on the internal
DQN trainer, but load_training_data() only loads OHLCV bars — it never
populates ofi_features. The OFI loading is done by the hyperopt adapter's
own load_ofi_features() method.

Fixes:
- preload_data() now calls self.load_ofi_features() directly
- load_ofi_features() uses self.mbp10_data_dir when set (was hardcoded ../mbp10)
- input_dim pre-computation uses OFI-aware size (51 when enabled, 43 otherwise)
- CI 'latest' package: delete-then-upload to avoid GitLab duplicate file issue
- Warn when mbp10_data_dir is set but no OFI features loaded

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

crates/ml/src/hyperopt/adapters/dqn.rs

@@ -1205,10 +1205,14 @@ impl DQNTrainer {
         }
         .map_err(|e| MLError::TrainingError(format!("Failed to preload data: {}", e)))?;
 
-        // Extract OFI features loaded by the trainer (from MBP-10 + trades data)
-        let ofi_features = loader.ofi_features.take();
+        // Load OFI features from MBP-10 data (separate from OHLCV loading).
+        // The internal trainer's load_training_data() only loads OHLCV — OFI must be loaded
+        // explicitly via the adapter's load_ofi_features() which uses self.mbp10_data_dir.
+        let ofi_features = self.load_ofi_features();
         if let Some(ref ofi) = ofi_features {
             info!("OFI features preloaded: {} bars x 8 dims (VPIN, Kyle's Lambda, OFI, trade imbalance)", ofi.len());
+        } else if self.mbp10_data_dir.is_some() {
+            warn!("MBP-10 data dir configured but no OFI features loaded — state_dim mismatch likely");
         }
 
         let elapsed = preload_start.elapsed();
@@ -1459,9 +1463,14 @@ impl DQNTrainer {
         use crate::features::mbp10_loader::load_mbp10_snapshots_sync;
         use crate::features::ofi_calculator::OFICalculator;
 
-        let mbp10_dir = std::path::Path::new(&self.dbn_data_dir)
-            .parent()
-            .map(|p| p.join("mbp10"))?;
+        // Use explicit mbp10_data_dir if set, otherwise fall back to ../mbp10 relative to OHLCV dir
+        let mbp10_dir = if let Some(ref dir) = self.mbp10_data_dir {
+            std::path::PathBuf::from(dir)
+        } else {
+            std::path::Path::new(&self.dbn_data_dir)
+                .parent()
+                .map(|p| p.join("mbp10"))?
+        };
 
         if !mbp10_dir.exists() {
             info!(

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

@@ -289,7 +289,8 @@ impl DQNTrainer {
         }
 
         // Pre-compute hidden dims to get accurate model size for batch sizing.
-        let input_dim: usize = 43;
+        let ofi_pre = hyperparams.mbp10_data_dir.is_some();
+        let input_dim: usize = if ofi_pre { 51 } else { 43 };
         let output_dim: usize = 5;
         let hidden_dims: Vec<usize> = match hyperparams.hidden_dim_base {
             Some(base) => {

infra/k8s/argo/ci-pipeline-template.yaml

@@ -523,8 +523,16 @@ spec:
                 "${GITLAB}/api/v4/projects/1/packages/generic/foxhunt-services/${TAG}/${BIN_NAME}"
             done
 
-            # Also publish under 'latest' so deployments always have a known version
+            # Delete old 'latest' package then re-upload (GitLab doesn't overwrite files)
             echo "=== Tagging as 'latest' ==="
+            OLD_PKG=$(curl -sf -H "PRIVATE-TOKEN: ${GITLAB_PAT}" \
+              "${GITLAB}/api/v4/projects/1/packages?package_name=foxhunt-services&package_version=latest" \
+              | grep -oP '"id":\K[0-9]+' | head -1)
+            if [ -n "$OLD_PKG" ]; then
+              curl -sf -X DELETE -H "PRIVATE-TOKEN: ${GITLAB_PAT}" \
+                "${GITLAB}/api/v4/projects/1/packages/${OLD_PKG}" && \
+                echo "Deleted old 'latest' package (ID ${OLD_PKG})" || true
+            fi
             for bin in "$WORKSPACE/bin/services/"*; do
               BIN_NAME=$(basename "$bin")
               curl -f --upload-file "$bin" \
@@ -652,8 +660,16 @@ spec:
                 "${GITLAB}/api/v4/projects/1/packages/generic/foxhunt-training/${TAG}/${BIN_NAME}"
             done
 
-            # Also publish under 'latest' so training workflows always have a known version
+            # Delete old 'latest' package then re-upload (GitLab doesn't overwrite files)
             echo "=== Tagging as 'latest' ==="
+            OLD_PKG=$(curl -sf -H "PRIVATE-TOKEN: ${GITLAB_PAT}" \
+              "${GITLAB}/api/v4/projects/1/packages?package_name=foxhunt-training&package_version=latest" \
+              | grep -oP '"id":\K[0-9]+' | head -1)
+            if [ -n "$OLD_PKG" ]; then
+              curl -sf -X DELETE -H "PRIVATE-TOKEN: ${GITLAB_PAT}" \
+                "${GITLAB}/api/v4/projects/1/packages/${OLD_PKG}" && \
+                echo "Deleted old 'latest' package (ID ${OLD_PKG})" || true
+            fi
             for bin in "$WORKSPACE/bin/training/"*; do
               BIN_NAME=$(basename "$bin")
               curl -f --upload-file "$bin" \