sample/README: the observation tensor is not part of the subset

sample/README.md

@@ -9,15 +9,12 @@ for inspecting the data format without downloading terabytes.
   different physical sample sizes. The material ids present are the `material_<id>/`
   folders listed above.
 - Each material keeps ~12 HDR frames in `hdr.tar`, evenly subsampled from the full ~580.
-- Every other per-material file — `observations_structured.npz`, `point_positions.npz`,
-  `rotated_camera.json`, `scan_log.json`, `point_metadata.json`, `bbox.json`,
-  `hdr_crop_bboxes.json`, `unmatched_scan_ids.json` — is included **verbatim**, so array
-  names, shapes and dtypes match the full release exactly.
-
-Note that `observations_structured.npz` here still indexes the *full* frame set: its
-`rgbs (K,V,3)` tensor has one row per registered frame, not one per frame kept in this
-`hdr.tar`. Use `scan_log.json` / `rotated_camera.json` to map the frames you do have back
-onto rows of the tensor.
+- The small per-material files — `point_positions.npz`, `rotated_camera.json`,
+  `scan_log.json`, `point_metadata.json`, `bbox.json`, `hdr_crop_bboxes.json`,
+  `unmatched_scan_ids.json` — are included **verbatim**, so names, shapes and dtypes match
+  the full release exactly.
+- The per-point observation tensor `observations_structured.npz` (≈1 GB per material) is
+  **not** part of the sample; fetch it from `materials/<id>/` when you need it (see below).
 
 ## Mapping back to the full dataset
 
@@ -44,6 +41,6 @@ type in a material directory. Once this subset is downloaded, the same arrays lo
 
 ```python
 import numpy as np
-obs = np.load("robocloth-sample/sample/material_<id>/observations_structured.npz")
-print({k: (obs[k].shape, obs[k].dtype) for k in obs.files})
+pts = np.load("robocloth-sample/sample/material_<id>/point_positions.npz")
+print({k: (pts[k].shape, pts[k].dtype) for k in pts.files})
 ```