SciEdu_pauses_3exp_210epi_leafdir_ctx2

leapshared/SciEdu_pauses_3exp_210epi_leafdir_ctx with one boundary moved. In each standard electroscope episode, the hold (task 15 Do the electroscope experiment. hold the plate still at the electroscope) now starts one action chunk (50 frames) before the plate comes to rest over the electroscope, in the pause the hold ends in. Same frames, videos, states and actions, the same 20 task sentences and every other boundary; only the task_index of 6,344 frames changes, from 15 to 19.

• Why. The recording page started a hold's sentence one chunk before the arm came to rest after the Space that ended the step before it. In all 60 standard electroscope episodes (140-199) that Space was pressed while the demonstrator held the plate off to the right of the electroscope (right fingertip, median of 60, [24.4, -10.5, 38.8] cm; model frame, x forward, y left, z up) for about 2.2 s, before bringing it back over the electroscope ([20.0, -0.1, 40.0] cm) and holding it there for about 3.8 s. So the hold began with the pause off to the right, and the last action chunk of the step before it (task 19, Do the electroscope experiment. move the plate toward and away from the electroscope, then stop above it) ended there: off to the right in 59 of 60 episodes, 11.4 cm (median) from where the plate is then held. On the arm (2026-10-04) a policy trained with that boundary held the plate there, 0.4 cm from the recordings' pause off to the right.
• What moved. The start of task 15 after task 19, in episodes 140-199, by 78-217 frames (median 100): 6,344 frames in all, which go from 15 to 19 (57 training episodes, 6,033 frames; 311 in the validation episodes 147, 167 and 187). The last chunk of 19 now ends over the electroscope (0.9 cm from the hold's pause, median; none off to the right). The plate takes about 34 frames to come back over the electroscope, so in 59 episodes the hold now starts in the last 2-25 frames (median 16) of the pause off to the right, and in one after it. 19 lasts 9.1 s instead of 5.7 s (median), 15 lasts 3.8 s instead of 7.1 s. Where the hold ends has not moved.
• Everything else is that of the ctx dataset, frame for frame: the leaf episodes 0-59, the Redox episodes 60-139, the electroscope recovery clips 200-209, and every other segment and boundary. meta/info.json, meta/stats.json and meta/tasks.parquet are the same byte for byte.
• Spec config/labels/electroscope_pauses_ctx2.yaml reproduces every frame's sentence of episodes 140-209; the rule is MOVES in scripts/dataset/context_split_labels.py and its numbers come from scripts/diagnostics/hold_boundary.py; why and how plan/context-split-labels.md, all in the OpenArm SciEdu robot repository.