check_file: an oversize .parquet file passes validation because the size rejection is overwritten

[toolshedlabs-hash]

Summary

check_file returns is_check_passed: True for a parquet file it has already rejected as over
the size limit, so files.upload(..., check=True) goes ahead and uploads it.

lib/utils/files.py, the size branch does not return:

if file_size > MAX_FILE_SIZE_GB * NUM_BYTES_IN_GB:
    report_dict["message"] = f"Maximum supported file size is ..."
    report_dict["is_check_passed"] = False     # falls through
elif file_size == 0:
    ...
    return report_dict                         # the empty case does return
else:
    report_dict["file_size"] = file_size

and further down it merges the format result over the top:

report_dict.update(data_report_dict)

_check_parquet ends its success path with an unconditional report_dict["is_check_passed"] = True
at line 364, so the merge puts the rejection back to True. _check_jsonl and _check_csv never
set the flag to True, so they are not affected. Parquet is the only one.

That flag is the upload guard, in three places: resources/files.py:161, resources/files.py:347,
and lib/cli/api/files/upload.py:34.

Reproduction

together 2.28.0, pyarrow 25.0.0, Python 3.14.0, fresh venv.

MAX_FILE_SIZE_GB is 50.1 and I did not want to generate a 50 GB file, so for the rows marked
"over the limit" I set together.lib.utils.files.MAX_FILE_SIZE_GB to 1e-9 for that one call and
restored it afterwards. Nothing else is patched. The under-the-limit rows run on the stock
constant. The other four rows are there so a reader can tell the patched threshold is not simply
forcing everything to True.

file	over the limit	is_check_passed	file_size	message
valid parquet	yes	True	None	Maximum supported file size is ...
valid parquet	no	True	1956	Checks passed
valid jsonl	yes	False	None	Maximum supported file size is ...
valid jsonl	no	True	50	Checks passed
malformed jsonl	yes	False	None	Error parsing json payload. Unexpected format on line 2.
csv, purpose fine-tune	yes	False	None	CSV files are not supported for fine-tune.
empty file	yes	False	0	File is empty

Full report dict for the first row:

{"is_check_passed": true,
 "message": "Maximum supported file size is 1e-09 GB. Found file with size of 0.0 GB.",
 "found": true, "file_size": null, "utf8": null, "line_type": null, "text_field": null,
 "key_value": null, "has_min_samples": null, "num_samples": 2, "load_json": null,
 "load_csv": null, "filetype": "parquet"}

file_size is null in that report as well, since the size branch never assigns it.

Suggested fix

Return from the oversize branch the way the empty branch already does:

if file_size > MAX_FILE_SIZE_GB * NUM_BYTES_IN_GB:
    report_dict["message"] = f"Maximum supported file size is ..."
    report_dict["file_size"] = file_size
    report_dict["is_check_passed"] = False
    return report_dict

Making _check_parquet set is_check_passed only when it is not already False would also work,
though the early return keeps the whole function honest for any check added later.