feat(batch27): JEPA ViT bridge, WebDataset shard reading, training loop

Gap 2 — rtx-vision ViT bridge (jepa_vision_bridge.rs, 8 tests):
- ViT::forward_features(): patch reps without classification head
- ViT::encode_patch_indices(): shape-correct placeholder for GPU dispatch
- RtxVisionJepaEncoder implementing JepaEncoder (vision-bridge feature)
- From<&ViTConfig> for JepaViTConfig config conversion
- rtx-vision added as optional dep; vision-bridge feature gate

Gap 3 — WebDataset tar-shard reading (jepa_data.rs, +12 tests, 47 total):
- parse_tar_bytes(): pure stdlib tar parser (512-byte block format)
- read_webdataset_shard(): file reader with ShardLoadStats timing
- WebDatasetRecord: key, image_bytes, label, extension
- ShuffleBuffer: fixed-capacity reservoir sampling via LCG PRNG
- JepaDataPipeline::from_filesystem(): validates paths, loads shards, builds pipeline

Gap 5 — Training loop runner (jepa_runner.rs + examples/jepa_train.rs, 15 tests):
- JepaRunConfig with TOML-style key=value parser
- run_jepa_training(): full training loop (JepaTrainerV2, cosine LR, checkpointing)
- JepaCheckpoint::save() writes JSON summary; load() stub
- examples/jepa_train.rs: --config/--size/--steps/--dry-run CLI flags

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

crates/models/rtx-vision/src/models/vit.rs

@@ -112,6 +112,42 @@ impl ViT {
             .map_err(VisionError::from)
     }
 
+    /// Returns all patch representations `[seq_len+1, embed_dim]` (CLS token first,
+    /// then patches). Does NOT apply the classification head. Suitable for feature
+    /// extraction (e.g. JEPA encoders).
+    pub fn forward_features(&self, x: &Tensor) -> Result<Tensor> {
+        let x = self.patch_embed.forward(x)?;
+        let x = self.patch_embed.add_class_token(&x)?;
+        let mut x = x;
+        for block in &self.blocks {
+            x = block.forward(&x)?;
+        }
+        self.norm.forward(&x)
+    }
+
+    /// Encode specific patch indices (0-indexed, skipping the CLS token at
+    /// position 0 in the feature sequence).
+    ///
+    /// Returns a flattened `[n_patches, embed_dim]` f32 vector.
+    ///
+    /// # Placeholder note
+    ///
+    /// The returned values are currently **zeros of the correct shape**. Extracting
+    /// arbitrary rows from a `Tensor` requires `narrow` + `to_vec()`, which depends
+    /// on backend dispatch not yet wired for all targets. The shape contract
+    /// (`n_patches * embed_dim` elements) is already correct and will be filled with
+    /// real values once GPU tensor row-extraction is available (Batch 27).
+    pub fn encode_patch_indices(&self, x: &Tensor, patch_indices: &[usize]) -> Result<Vec<f32>> {
+        // Run full forward pass to get [n+1, embed_dim] features.
+        let _features = self.forward_features(x)?;
+        // patch i lives at position i+1 (position 0 is the CLS token).
+        // Placeholder: return zeros with the correct shape until Tensor row
+        // extraction (narrow + to_vec) is wired to the GPU backend.
+        let n_patches = patch_indices.len();
+        let d = self.config.embed_dim;
+        Ok(vec![0.0f32; n_patches * d])
+    }
+
     fn extract_class_token(&self, x: &Tensor) -> Result<Tensor> {
         let shape = x.shape().dims();
         match shape.len() {