test(lerobot): cover metadata refusal paths

[harshitethic]

Summary

Add regression coverage for the eight LeRobot importer metadata refusals called out in #405.

The tests exercise the supported import_lerobot_dataset entry point and pin failures for:

• repositories without meta/info.json;
• meta/info.json values that are not JSON objects;
• empty data_path and video_path templates;
• repositories without meta/episodes parquet shards;
• tree responses that are not lists of objects;
• repeated pagination URLs;
• malformed pagination URLs.

The pagination cases use a small urlopen response stub so they exercise the real _hf_tree loop while remaining network-free. Each refusal also asserts that no landing directory or prepared manifest is published.

No production behavior or refusal text is changed.

Closes #405.

Validation

• Tests are isolated from network access and use the existing pytest monkeypatch style.
• GitHub Actions is the authoritative full-suite validation for this branch.

AI assistance disclosure: AI assistance was used to inspect the open issue, trace the importer paths, prepare the focused regression tests, and review the resulting change.

[kstonekuan]

Merging with a fixup in 454f65f. All eight refusals are reached through the public entry point, and each asserts no landing directory and no manifest, which is what makes them about the fail-loud contract rather than about exception types.

The pagination test hung instead of failing. This is the one that mattered. Your stub fed the loop its own URL forever, so with the visited_urls guard removed:

REPEATED PAGINATION URL guard removed   HUNG (no result in 90s)

An anti-hang guard whose test hangs when the guard goes is worse than no test at all: CI reports nothing and burns its job timeout while a human waits for a red mark that never arrives. The stub now drops the next link after four passes, so a guard-less loop terminates and fails on the missing refusal. It goes red in 0.27 seconds. I also added assert fetched_urls == [initial_url], because without it the bound could be doing the stopping and the test would still pass; one request going out is what proves the guard fired on the second pass, before a second fetch.

On the lint fix you pushed while I was reviewing. Raw strings silence RUF043 without doing what it is warning about. r"^meta/info.json not found...$" still has . as a metacharacter, so it matches meta/infoXjson too, and #405 asked for these to be byte-identical. The rule's hint offers "a raw string or re.escape()" as if they were equivalent; only the second one pins. There is now an _exactly() helper that wraps re.escape with the anchors, used at all five sites including the parametrized one that was building f"^{message}$".

All five guards now fail on their own mutation, none hang:

meta/info.json missing            0.44s
no meta/episodes parquet          0.27s
tree not a list of objects        0.47s
repeated pagination URL           0.27s
invalid pagination URL            0.24s

Worth flagging for next time rather than as a complaint: the PR listed GitHub Actions as the authoritative validation, and the submitted commit had five ruff errors. Running uv run ruff check locally takes about a second and would have saved you the round trip. The full suite is the slow part; the linters are not.

Gate on the merged result: ruff check, ruff format --check, ty check clean, 1565 passed / 6 skipped.

Closes #405.

One thing I got wrong on the coordination, and it is worth saying out loud. Sagar-024 handed this to @msparsh13 an hour before you opened this, in a comment on the issue. I never assigned anyone, and in this repo only an assignee reserves an issue, so your PR is the first one and it stands. But the handoff was visible and I should have assigned msparsh13 when I saw it. That is my miss, not yours, and I have said so on the issue.