[rakesh0x]

Summary

Fixes #299.

The LeRobot metadata boundary in _ensure_source_archive validated that fps from meta/info.json is numeric and positive, but not finite — so float("nan") and float("inf") passed validation and only failed later with misleading errors (no meta/episodes parquet files found, or a raw int() conversion failure in episode conversion).

This change mirrors the existing TransformConfig.gop_seconds pattern (math.isfinite) at the metadata boundary:

• fps must now be finite as well as numeric and positive.
• The contextual ValueError names the offending value and states the complete rule: FPS must be finite and positive.
• The refusal happens before episode metadata discovery, so no episode metadata, source video, Parquet, or canonical output is written after an invalid info document is read.

Tests

Parametrized metadata-boundary regressions in tests/test_lerobot_converter.py:

• NaN, +inf, and -inf are refused as invalid FPS (new behavior).
• Boolean, missing, nonnumeric, zero, and negative FPS keep their existing refusal.
• A normal positive FPS (30) is still accepted through discovery.
• Each refusal case asserts the error names the bad value, mentions "finite and positive", and that _hf_tree episode discovery never runs and no meta/episodes output is created.
• The regressions were verified to fail against unpatched main and pass with the fix.

No new dependency and no converter-version bump.

Validation

• uv run ruff check --fix — clean
• uv run ruff format — clean
• uv run ty check — clean
• uv run pytest -q tests/test_lerobot_converter.py — 18 passed
• Full suite from repo root (tests + packages/hflow-server/tests per testpaths): 1255 passed, 11 skipped, 0 failed

[kstonekuan]

LGTM, merging.

Validated on your branch: ruff check, ruff format --check, ty check clean, 1260 passed / 6 skipped.

Also checked the tests can detect the fix being removed. Dropped the math.isfinite clause and the nan and positive-infinity cases failed, the other six still passed. That's the right split: -inf was already caught by <= 0, so nan and +inf are the actual new coverage and your parametrization makes that visible instead of hiding it behind one green test.

"finite and positive" matches what _camera_motion.py:59 and _frame_statistics.py:98 already say, so the message reads like the rest of the repo.

One thing, and it's an invitation not a rule: the good first issue pool is where newcomers land, and you're three PRs in now, so it's worth leaving those for people who need somewhere to start. #320 (parse LeRobot metadata into a typed archive model) is the natural next step from this one, and it's where the _SourceArchive TypedDict you just validated a field of gets turned into something that can't hold a bad fps in the first place.

Or the higher-leverage version: you already filed #287 off Egocentric-10K, and that report was more useful than any of these fixes. More of that, pointing HFlow at real corpora and telling us what breaks, is worth more to us than the starter queue.