feat: benchmaxxing datasets stats manifest summary + sanity check (closes #113)

[felipeocampoos]

What

Adds benchmaxxing datasets stats <manifest>, a sanity-check summary for a manifest before it's used in an expensive run.

• Loads the manifest via benchmaxxing.data.load_cases (which already enforces the hard invariants -- empty manifest, malformed rows, out-of-range answer_index -- raising before this command sees the data).
• Prints: row count and per-modality counts, an MCQ options-per-case distribution, an optional on-disk image_ref check via --image-root, the label distribution, and how many rows carry meta.
• benchmaxxing datasets (no subcommand) is unchanged -- it still lists adapters; list is the explicit form of the same default, so nothing that already depends on the bare datasets command breaks.
• On a missing/malformed manifest, prints a clear error: ... to stderr and exits 1 instead of a stack trace.

A bug I caught testing this

My first pass computed the answer_index range check as not (0 <= (c.answer_index or -1) < len(c.options)). Since 0 is falsy in Python, a perfectly valid answer_index=0 got treated as "missing" and flagged as out of range. Fixed to c.answer_index if c.answer_index is not None else -1, and added test_datasets_stats_accepts_answer_index_zero as a regression test so it can't come back silently.

Testing

tests/test_cli.py:

• MCQ manifest summary (row/modality counts, options distribution, label distribution)
• image manifest with and without --image-root
• missing-manifest and empty-manifest non-zero exit
• the answer_index=0 regression above
• datasets and datasets list produce identical output

Also manually verified against a synthetic MCQ manifest and a synthetic imaging manifest (with a real file + a missing one) outside the test suite.

Suite: 402 passed, 6 skipped, ruff check . clean. The one failure I see locally (test_checksum_is_stable_and_matches_hashlib) is pre-existing on main (Windows CRLF vs a literal-string hash fixture), unrelated to this change.

CONTRIBUTING.md updated with the new subcommand.

Closes #113.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[sebasmos]

Heads up @felipeocampoos, this now conflicts with main after #145 and #148 landed (both touched cli.py's subcommand registration). A quick rebase onto main should clear it; the datasets stats command itself was clean and passing (495 tests, ruff clean) when I tested the merge locally. Re-request review after the rebase and I'll land it.

[sebasmos]

Thanks @felipeocampoos, datasets stats is a genuinely useful addition and the list + stats split is the right shape. Nice touch that a bare benchmaxxing datasets still defaults to list, so existing callers and the smoke pipeline keep working, and moving statsmodels into the datasets extra reads fine.

One fix before it can land: the branch looks based on a main from before #145, so the version command here has lost its --verbose flag (git SHA + installed extras). Could you rebase onto current main so version --verbose survives the merge? After that it is good to go, and it will slot in next to #142 (MedMCQA) once that rebases too.

[felipeocampoos]

Rebased onto current main -- version --verbose (#145) survives the merge, cli.py's subcommand registration is unified (version --verbose + datasets list/stats coexist), and the conflict is resolved. 588 passed, 6 skipped, ruff clean. Ready for another look.

[sebasmos]

Confirmed, thanks @felipeocampoos. Rebased cleanly onto current main (diff is purely additive: cli.py + tests/test_cli.py + a CONTRIBUTING.md note, no deletions), version --verbose survives, datasets list/datasets stats coexist and a bare datasets still defaults to list. Verified independently: 587 passed, 7 skipped, ruff clean. Approving and merging.