feat: language annotation pipeline (#3471)

Steerable annotation pipeline (lerobot-annotate) that populates the language_persistent and language_events columns introduced in PR 1 (#3467) directly into data/chunk-*/file-*.parquet.

This is PR 2 of the three-PR plan:

PR 1 (Add extensive language support #3467): schema + DSL + rendering, base of this PR
PR 2 (this PR): annotation pipeline writing into PR 1's columns
PR 3: model with language prediction and runtime
A VLM (Qwen-VL family, served on vLLM) watches each episode's video and emits grounded language annotations: subtasks, plans, memory, task rephrasings, interjections + speech, and per-camera VQA. The pipeline is built for production annotation at scale — single-camera grounding, embedded-frame inputs, a describe-then-segment grounding flow, and a deterministic full-episode coverage guarantee — informed by Scale's dense-captioning findings (representation > sampling, rules > reasoning, model capacity is the biggest lever, two-pass systems compound errors)

tests/fixtures/dataset_factories.py

@@ -552,3 +552,64 @@ def lerobot_dataset_factory(
 @pytest.fixture(scope="session")
 def empty_lerobot_dataset_factory() -> LeRobotDatasetFactory:
     return partial(LeRobotDataset.create, repo_id=DUMMY_REPO_ID, fps=DEFAULT_FPS)
+
+
+def build_annotation_dataset(
+    root: Path,
+    episode_specs: list[tuple[int, int, str]],
+    *,
+    fps: int = 10,
+) -> Path:
+    """Build a minimal LeRobot-shaped dataset on disk for annotation tests.
+
+    ``episode_specs`` is a list of ``(episode_index, num_frames, task_text)``.
+    Each episode is written to its own
+    ``data/chunk-000/file-{ep:03d}.parquet`` so the writer's per-shard
+    rewrite path is exercised. The dataset carries the minimum
+    ``meta/tasks.parquet`` + ``meta/info.json`` the reader / executor need;
+    it has no videos, so the modules fall back to text-only prompts.
+
+    Shared by the annotation-pipeline pytest fixtures (``tests/annotations/
+    conftest.py``) and the opt-in E2E smoke run so the fixture shape lives
+    in exactly one place.
+    """
+    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)
+
+    tasks: dict[int, str] = {}
+    for episode_index, num_frames, task_text in episode_specs:
+        if task_text not in tasks.values():
+            tasks[len(tasks)] = task_text
+        task_index = next(k for k, v in tasks.items() if v == task_text)
+        frame = pd.DataFrame(
+            {
+                "episode_index": [episode_index] * num_frames,
+                "frame_index": list(range(num_frames)),
+                "timestamp": [round(i / fps, 6) for i in range(num_frames)],
+                "task_index": [task_index] * num_frames,
+                "subtask_index": [0] * num_frames,  # legacy column the writer must drop
+            }
+        )
+        frame.to_parquet(data_dir / f"file-{episode_index:03d}.parquet", index=False)
+
+    # Canonical tasks frame: indexed by task string with a ``task_index``
+    # column, matching what ``lerobot.datasets.io_utils.load_tasks`` expects.
+    tasks_df = pd.DataFrame(
+        {"task_index": list(tasks.keys())},
+        index=pd.Index(list(tasks.values()), name="task"),
+    )
+    write_tasks(tasks_df, root)
+
+    write_json(
+        {
+            "codebase_version": "v3.1",
+            "fps": fps,
+            "features": {},
+            "total_episodes": len(episode_specs),
+        },
+        root / "meta" / "info.json",
+    )
+    return root