feat(datasets): add lerobot-dataset-quality to flag outlier episodes

[maeste]

Summary / Motivation

Once a teleoperation dataset grows past a few dozen episodes, reviewing each one in lerobot-dataset-viz is impractical, so bad demonstrations (struggled long attempts, sharp corrections, hesitations, wrong end poses) silently stay in the training set. This PR adds lerobot-dataset-quality, a read-only CLI that computes deterministic per-episode metrics from the recorded actions and flags statistical outliers, turning "watch 200 episodes" into "watch these 8". The flagged list feeds directly into the existing lerobot-dataset-viz → lerobot-edit-dataset --operation.type delete_episodes workflow.

Related issues

• Closes: [feat] lerobot-dataset-quality: per-episode quality metrics and outlier flagging for recorded datasets #3760
• Related: [feat] lerobot-check-calibration: detect leader/follower calibration drift in recorded datasets #3758 (companion diagnostic for leader/follower calibration drift)

What changed

• src/lerobot/scripts/lerobot_dataset_quality.py: new analysis tool. Core logic is in pure, importable functions (compute_episode_metrics, detect_outliers, evaluate_dataset_quality) with a thin argparse main() following the lerobot-dataset-viz CLI conventions (--repo-id, --root). Only the action column is read — no image/video decoding, so it runs in seconds.
• Metrics per episode: n_frames/duration_s, median_jerk and p95_jerk (smoothness vs. isolated spikes), max_velocity, static_fraction (hesitations), and final-pose distance from the across-episode mean. Outliers flagged with the IQR rule (--k-iqr, default 1.5).
• pyproject.toml: new lerobot-dataset-quality entry point.
• tests/scripts/test_dataset_quality.py: 11 unit tests on synthetic trajectories with known properties (smooth vs. spiked vs. holding), outlier detection (uniform set → no flags; long episode → duration_high; divergent end pose → final_state_high), episode grouping/sorting, and the end-to-end report including error paths.
• docs/source/using_dataset_tools.mdx: user-facing documentation section.

No breaking changes; purely additive and read-only.

How was this tested (or how to run locally)

• Tests added: tests/scripts/test_dataset_quality.py — uv run pytest tests/scripts/test_dataset_quality.py -v (11 passed)
• pre-commit run passes on all touched files (ruff, mypy, bandit, typos, prettier).
• Manually verified on real SO-101 teleop datasets: flagged episodes matched the ones already identified as bad by manual review in lerobot-dataset-viz (over-long struggled attempts and episodes ending in the wrong pose).
• Quick reviewer repro: lerobot-dataset-quality --repo-id lerobot/pusht (any dataset with an action feature works) or add --output-format json for raw numbers.

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green
• Community Review: I have reviewed another contributor's open PR and linked it here: # (pending)

Reviewer notes

• The IQR rule was chosen over z-scores because episode metrics are typically skewed (e.g. a few very long episodes); it is also what flags nothing on a perfectly uniform dataset.
• median_jerk vs p95_jerk is deliberate: the median is robust to isolated corrective spikes, the p95 is there to catch exactly those spikes.
• The tool prints a suggested lerobot-edit-dataset delete command but never modifies anything itself; the docs stress reviewing flagged episodes visually before deleting.
• Happy to rename the command, tune default thresholds, or fold this elsewhere if maintainers prefer.

🤖 Generated with Claude Code

[maeste]

Superseded by #3762, which combines this tool and #3759 into a single dataset-diagnostics PR.