fix(pour_kettle_cup): drop the 2 unvalidated 787-era openarm episodes
Drops the two 787-era episodes from the openarm kettle bundle, leaving the five 775-era demos the kettle_pour thresholds were actually calibrated on.
Why these two. They were merged into the consolidated set later than the calibration cohort and never re-validated against the tightened directional checker. Two independent defects, either sufficient on its own:
- ep5 is not a pour. It fails the directional gate at
lean_toward_cup = -0.876against a>= +0.400requirement — the kettle leans away from the cup — plusspout_xy_dist = 0.158vs<= 0.130. Not marginal, opposite. It was accepted at collection time by the older direction-blind checker. - Neither can be state-replayed at all. Both use the pre-rename
_hand_fingerjoint naming, so replay dies withKeyError: "Invalid name 'openarm_wuji/left_hand_finger1_joint1'"against the current model.
The filter is by recorded _mujoco_state.buffer length (775 vs 787), not by episode index — the era boundary is the actual defect line, and ep6 happening to pass the checker does not make it replayable.
| before | after | |
|---|---|---|
parrot/pour_kettle_cup/openarm_wuji |
7 | 5 |
Verified: all five survivors pass the tightened kettle_pour checker. This makes the shipped bundle single-era, so it is no longer a mixed-era set.
Unblocks test_recorded_openarm_demos_pass_tightened_checker, which had been failing on ep5 since the data moved to a path the test could actually find (it was silently skipping before).