fix(datasets): enforce one parquet row group per episode in v3 data writes

LeRobot v3 data shards must hold exactly one row group per episode so a
reader can fetch episode i with pq.ParquetFile(path).read_row_group(i)
(a byte-range read) instead of loading the whole shard. The recording
writer already does this (one write_table per episode); the aggregate
and lerobot-annotate re-write paths instead concatenated many episodes
and wrote them in one shot, collapsing the file to a single row group.

- io_utils: add write_table_one_row_group_per_episode (one ParquetWriter,
  one write_table per episode — same pattern as the recording writer);
  to_parquet_with_hf_images embeds images then writes per-episode row
  groups; to_parquet_one_row_group_per_episode wraps it for plain frames
- aggregate: route non-image data writes through the per-episode writer;
  leave the episodes-metadata parquet untouched (already one row/episode)
- annotate: rewrite shards via the per-episode writer instead of a single
  bulk pq.write_table
- tests: invariant coverage through the aggregate (image + video) and
  annotate paths

No change to on-disk schema, paths, naming, rollover thresholds, or
compression. Readers stay backward-compatible (old collapsed files load).

tests/annotations/test_writer.py

@@ -28,6 +28,7 @@ import pytest
 pytest.importorskip("datasets", reason="datasets is required (install lerobot[dataset])")
 pytest.importorskip("pandas", reason="pandas is required (install lerobot[dataset])")
 
+import pandas as pd  # noqa: E402
 import pyarrow.parquet as pq  # noqa: E402
 
 from lerobot.annotations.steerable_pipeline.reader import iter_episodes  # noqa: E402
@@ -344,6 +345,78 @@ def test_annotation_metadata_sync_allows_non_streaming_load(
     assert len(dataset) == 24
 
 
+def _build_packed_dataset(root: Path, episode_lengths: list[int], *, fps: int = 10) -> Path:
+    """Pack several episodes into a single shard (vs build_annotation_dataset's one-per-file),
+    so the writer's rewrite must re-emit one row group per episode instead of collapsing them."""
+    from lerobot.datasets.io_utils import write_tasks
+    from lerobot.utils.io_utils import write_json
+
+    data_dir = root / "data" / "chunk-000"
+    data_dir.mkdir(parents=True, exist_ok=True)
+
+    episode_index, frame_index, timestamp, task_index, subtask_index = [], [], [], [], []
+    for ep, length in enumerate(episode_lengths):
+        episode_index += [ep] * length
+        frame_index += list(range(length))
+        timestamp += [round(i / fps, 6) for i in range(length)]
+        task_index += [0] * length
+        subtask_index += [0] * length  # legacy column the writer must drop
+    pd.DataFrame(
+        {
+            "episode_index": episode_index,
+            "frame_index": frame_index,
+            "timestamp": timestamp,
+            "task_index": task_index,
+            "subtask_index": subtask_index,
+        }
+    ).to_parquet(data_dir / "file-000.parquet", index=False)
+
+    tasks_df = pd.DataFrame({"task_index": [0]}, index=pd.Index(["do the thing"], name="task"))
+    write_tasks(tasks_df, root)
+    write_json(
+        {"codebase_version": "v3.1", "fps": fps, "features": {}, "total_episodes": len(episode_lengths)},
+        root / "meta" / "info.json",
+    )
+    return root
+
+
+def test_writer_one_row_group_per_episode(tmp_path: Path) -> None:
+    """Rewriting a packed shard must keep one row group per episode, not collapse
+    every episode into a single giant row group."""
+    episode_lengths = [4, 6, 5]  # unequal lengths, all in one shard
+    root = _build_packed_dataset(tmp_path / "ds", episode_lengths)
+    shard = root / "data" / "chunk-000" / "file-000.parquet"
+    assert pq.ParquetFile(shard).metadata.num_row_groups == 1, "fixture should start collapsed"
+
+    staging_dir = tmp_path / "stage"
+    for ep in range(len(episode_lengths)):
+        _stage_episode(
+            staging_dir,
+            ep,
+            plan=[
+                {
+                    "role": "assistant",
+                    "content": f"subtask for ep {ep}",
+                    "style": "subtask",
+                    "timestamp": 0.0,
+                    "tool_calls": None,
+                }
+            ],
+        )
+
+    records = list(iter_episodes(root))
+    LanguageColumnsWriter().write_all(records, staging_dir, root)
+
+    # One row group per episode, with row counts matching the episode lengths.
+    md = pq.ParquetFile(shard).metadata
+    assert md.num_row_groups == len(episode_lengths)
+    assert [md.row_group(i).num_rows for i in range(md.num_row_groups)] == episode_lengths
+    # Language columns are still present after the per-episode rewrite.
+    table = pq.read_table(shard)
+    assert "language_persistent" in table.column_names
+    assert "language_events" in table.column_names
+
+
 def test_speech_atom_shape_matches_plan_spec() -> None:
     atom = speech_atom(2.5, "I'm cleaning up!")
     assert atom["role"] == "assistant"

tests/datasets/test_aggregate.py

@@ -32,6 +32,26 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from tests.fixtures.constants import DUMMY_REPO_ID
 
 
+def assert_data_shards_one_row_group_per_episode(root):
+    """Every aggregated DATA shard must have exactly one parquet row group per episode."""
+    import pyarrow.parquet as pq
+
+    shards = sorted((root / "data").rglob("*.parquet"))
+    assert shards, f"no data shards found under {root}/data"
+    n_episodes = 0
+    for shard in shards:
+        pf = pq.ParquetFile(shard)
+        episodes = pf.read(columns=["episode_index"]).column("episode_index").to_pylist()
+        assert pf.metadata.num_row_groups == len(set(episodes)), shard
+        for i in range(pf.metadata.num_row_groups):
+            rg_episodes = set(
+                pf.read_row_group(i, columns=["episode_index"]).column("episode_index").to_pylist()
+            )
+            assert len(rg_episodes) == 1, f"{shard} row group {i} spans episodes {rg_episodes}"
+        n_episodes += len(set(episodes))
+    return n_episodes
+
+
 def assert_episode_and_frame_counts(aggr_ds, expected_episodes, expected_frames):
     """Test that total number of episodes and frames are correctly aggregated."""
     assert aggr_ds.num_episodes == expected_episodes, (
@@ -566,6 +586,41 @@ def assert_image_frames_integrity(aggr_ds, ds_0, ds_1):
             )
 
 
+@pytest.mark.parametrize("use_videos", [True, False], ids=["video", "image"])
+def test_aggregate_one_row_group_per_episode(tmp_path, lerobot_dataset_factory, use_videos):
+    """Aggregated DATA shards keep one row group per episode (not one collapsed group).
+
+    Covers both the non-image (``df.to_parquet``) and image
+    (``to_parquet_with_hf_images``) write branches, including the merge-into-
+    existing-file branch via a low file-size threshold that forces packing.
+    """
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "rg_0",
+        repo_id=f"{DUMMY_REPO_ID}_rg_0",
+        total_episodes=3,
+        total_frames=60,
+        use_videos=use_videos,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "rg_1",
+        repo_id=f"{DUMMY_REPO_ID}_rg_1",
+        total_episodes=4,
+        total_frames=80,
+        use_videos=use_videos,
+    )
+
+    aggr_root = tmp_path / "rg_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_rg_aggr",
+        aggr_root=aggr_root,
+    )
+
+    n_episodes = assert_data_shards_one_row_group_per_episode(aggr_root)
+    assert n_episodes == ds_0.num_episodes + ds_1.num_episodes
+
+
 def test_aggregate_image_datasets(tmp_path, lerobot_dataset_factory):
     """Test aggregation of image-based datasets preserves HuggingFace Image schema.