SOURCES: clarify RTM timestamp-based frame picker

SOURCES.md

@@ -301,13 +301,21 @@ containing a list of 256 short video clips (`sensor_video[].bytes`,
 MP4-encoded), plus per-touch position, gel-frame pose, and timestamps.
 
 **How we processed it.**
-1. For each touch video clip, decoded with OpenCV and kept the **middle
-   frame** (near peak contact, where the imprint is most informative).
-2. No empty-frame filtering needed — every middle frame is by
-   construction at peak contact.
-3. No dedupe — each touch is a unique digit at a unique position.
-4. Per-row metadata: `digit_class` (0–9), `episode` (which of the 600
-   physical digit objects), `obj_name="digit_<N>"`.
+1. For each touch video clip, used the upstream **`touch_start_time_rel`
+   and `touch_end_time_rel` timestamps** to locate the actual contact
+   window (typically ~6 frames near the end of a 60–73-frame clip), then
+   picked the frame at the *middle of the contact window*. A naive
+   "middle-of-video" picker selects the gel *before* the touch starts —
+   we tried that first and the resulting samples were all blank gel, so
+   the timestamp-aware picker was necessary.
+2. Fallback to "frame with maximum central-region std" when timestamps
+   are missing or stale.
+3. No empty-frame filtering needed — every picked frame is at peak
+   contact by construction.
+4. No dedupe — each touch is a unique digit at a unique position.
+5. Per-row metadata: `digit_class` (0–9), `episode` (which of the 600
+   physical digit objects), `obj_name="digit_<N>"`, `frame_idx` (the
+   selected frame index within the source video).
 
 Kept rate: **153,600 raw → 153,600** (100%).