Parse LeRobot metadata into a typed internal archive model

[kstonekuan]

Current behavior

The LeRobot importer has typed top-level episode and video-window dictionaries, but _SourceArchive still contains untyped nested values at src/hflow/importers/lerobot.py:50-89:

info: dict
numeric_features: dict[str, dict]

_fetch_info_json() returns a raw dictionary at src/hflow/importers/lerobot.py:260-276. _ensure_source_archive() then parses fps, path templates, features, and feature kinds at several later points in src/hflow/importers/lerobot.py:300-469. Conversion continues to index the archive and inspect optional dictionaries at src/hflow/importers/lerobot.py:639-685.

Recent metadata bugs such as #299 and the boolean shape-dimension fix in fbfd553 both came from values remaining loosely typed after the external JSON boundary.

Verified against main at fbfd553.

Pattern to copy

parse_project_config() accepts an open external document, validates its supported fields once, and returns a ProjectConfig dataclass. DatasetSource in the importer already provides the same internal shape for repository identity.

What to build

Parse supported meta/info.json fields into immutable internal values before metadata discovery or conversion continues. Give the source archive, episode rows, video windows, and supported numeric feature specifications attribute-based domain types.

Keep the downloaded info.json cache as the original JSON-compatible document. Unknown upstream fields can remain ignored. Unsupported selected feature shapes should still fail before any episode or manifest is published.

This shares _ensure_source_archive() with #293 and #299. Preserve those fixes and rebase if either lands first.

Definition of done

1. A dedicated parser turns raw meta/info.json into a typed dataset-information value.
2. FPS, data and video path templates, robot type, video feature names, and supported numeric feature shapes are validated at that boundary.
3. _ensure_source_archive() returns an immutable archive type, with no TypedDict containing nested bare dictionaries.
4. Episode conversion consumes typed archive, episode, video-window, and numeric-feature values.
5. Code after the parser does not call .get() on raw LeRobot feature specifications.
6. Existing valid synthetic v3 metadata produces the same selected episodes, paths, schemas, provenance, and manifest.
7. Existing malformed FPS, dtype, shape, and required-feature refusals keep failing before dataset output.
8. Unknown metadata fields remain compatible.

Non-goals

• Modeling every Hugging Face API or DuckDB row
• Supporting new LeRobot dtypes or feature ranks
• Changing the prepared manifest or canonical MCAP format
• Adding LeRobot, Pydantic, or another dependency

Validation

uv run ruff check --fix
uv run ruff format
uv run ty check
uv run pytest -q tests/test_lerobot_converter.py
uv run pytest -q

[rakesh0x]

this issue should get closed, as the fix already been pushed and merged @kstonekuan

[Sagar-024]

Thanks for flagging it @rakesh0x, but I think #313 closed #299 (non-finite
fps), which is a different issue. Real fix, just not this one.

Checked the DoD against current main to be sure: _SourceArchive is still
a TypedDict holding info: dict and numeric_features: dict[str, dict],
_fetch_info_json() still returns a raw dict, and episode conversion still
reads feature specs with .get() on raw dicts (lerobot.py:516, 553). The
finite-fps guard from #313 covers part of DoD item 7, but the typed archive
itself is not there yet.

@kstonekuan could you confirm this one is still open? If so I'd like to
take it.