Sync root QA docs with release prefix

docs/SANITY_CHECKS_AND_LEAKAGE_GUARDS.md

@@ -1,40 +1,33 @@
-# Sanity Checks and Leakage Guards
-
-## Alignment checks
-
-- `T_rgb_side_right == T_rgb_side_left == T_rgb_wrist == T_seg == T_state == T_action == T_graph`.
-- `action[t]` corresponds to transition from state/frame `t` to `t+1`, following RoboCasa/LeRobot convention.
-- Episode `control_freq` is recorded.
-
-## Camera checks
-
-- All v1 episodes export `robot0_agentview_right`, `robot0_agentview_left`, and `robot0_eye_in_hand`.
-- `primary_side_camera` is `robot0_agentview_right`.
-- Report per-view node visibility; do not assume right always sees best.
-
-## Graph leakage checks
-
-- No instruction-derived pair edge appears in `rel_state`.
-- No task-goal edge appears in `rel_prior` unless marked as `goal_grounding.supervision_only` outside graph tensors.
-- No LLM/reasoning output is used for contact/support/held/current physical state.
-- Prior affordance edges are category-broadcast, not target-pair-only.
-- Node inventory is deterministic from sim metadata/config, not language-filtered.
-
-## Extraction quality checks
-
-- Projection/mask consistency: projected node body center should usually fall inside or near its segmentation mask when visible.
-- Contact aggregation stats logged before/after filtering.
-- Event-count sanity by task family:
-  - open/close tasks: articulation threshold event;
-  - pick-place tasks: grasp, release, support changes;
-  - switch/knob tasks: robot contact/control + state change;
-  - navigation task: base-motion flag and relation/visibility changes.
-- Quarantine episodes failing hard asserts; do not silently drop them.
-
-## Shape checks
-
-- `robot/state.npy`: `(T, 16)`.
-- `robot/action.npy`: `(T, 12)`.
-- `node_feats`: `(T, N_max, D_node)`.
-- `rel_state`: `(T, N_max, N_max, C_state)`.
-- `rel_prior`: `(N_max, N_max, C_prior)`.
+# Sanity checks and leakage guards
+
+This file is intentionally aligned with `docs/QA_AND_LEAKAGE.md` for GWAM_Data v1.2 sparse-v2.
+
+Normative current shapes:
+
+```text
+node_state: float32[T, 256, 32]
+view evidence + RLE masks: sparse visible node/view pairs
+dynamic_edges: sparse JSONL contact + visibility_change
+Phase 2: sparse visible-only feature files plus phase2_summary.json
+```
+
+Historical/stale shapes that must not be used for v1.2 sparse-v2 loaders:
+
+```text
+317-D / 323-D node_feats
+dense rel_state / rel_prior
+dense visual_views
+distractor family id 1
+state_on/state_off events.jsonl vocab
+```
+
+Minimum package gate:
+
+1. `_PHASE1_DONE.json` for every packaged episode.
+2. `audit/phase2_summary.json` for every packaged Phase 2 episode.
+3. No invisible Phase 2 features.
+4. Phase 2 pairs partition Phase 1 visible pairs.
+5. No NaN/Inf in `node_state`.
+6. Padded rows all-zero; loaders gate on active channel.
+7. Goal/instruction/object-role fields remain supervision-only.
+8. Known drop `target/OpenDrawer/episode_000459` documented and not packaged as a half episode.