trajs v3: canonicalize parrot trajectory table representation

Content-only migration of trajs/parrot/*/tianji_wuji/trajectory.pkl.gz (30 datasets, ~2,600 episodes, ~770k states). Flat parrot layout unchanged — all existing loaders work as-is.

Why

Data audit found three generations of table-representation debt:

• 20/30 datasets carry NO front_table key at all → metasim IsaacSim _set_states hard-indexes every scene object → native-Isaac replay crashes KeyError: 'front_table' (MuJoCo's lenient partial-state semantics masked this for two months)
• Two slab eras encode the table as a rigid slab with directly-set z (z=0.38 early / 0.686 / 0.636) while the delivery world builds the ArtVIP lift desk (height in joint dofs) → Isaac writes the desk root to the slab pose (silent misplacement)
• 5 mixed-era files (e.g. grasp_any spans 4 collection eras) and 62k collection-era object names (ycb_apple, cuboid_model)

What changed

1. SLAB front_table entries converted with the runtime seat_scene_on_lift_table (byte-identical semantics to the task reset path, incl. the below-native-top extra-dz prop raise) — 168,224 states
2. ABSENT episodes get the lift-desk fragment injected at the task's measured rise (recovered from Stage-A replay lift dofs: 0.0904 std / 0.0404 pack_box / ~0 trash_bin; landmark prop rest-z validated per task, e.g. grasp_orange min prop z 0.706 vs implied top 0.705) — 595,890 states
3. Collection-era object keys canonicalized (ycb_apple→apple, cuboid_model→cuboid) — 62,307 renames
4. _mujoco_state byte buffers preserved untouched; metadata.format_note = "parrot_v3_canonical" + per-task provenance in trajs/parrot/migration_report_v3.json

Validation

• v3 re-audit: every episode uniform DESK(dof) front_table, zero legacy names
• MuJoCo PD-replay regression (10 tasks, orig-vs-v3): no regressions, and slab-era replays are FIXED (grasp_apple 0/3→3/3, grasp_bottle 0/2→2/2)
• Migration code: SomaStacks/WUJI_SimBench#48 (scripts/migrate_parrot_trajs_v3.py)