Fix MetadataConfigs dropping parquet shards for non-consecutive configs

[rodriguescarson]

Fixes #8269

Problem

MetadataConfigs._from_exported_parquet_files_and_dataset_infos() groups the exported parquet files with itertools.groupby(exported_parquet_files, itemgetter("config")). groupby only groups consecutive equal keys, but the result is built as a dict keyed by config name:

metadata_configs = {
    config_name: { ... }
    for config_name, parquet_files_for_config in groupby(exported_parquet_files, itemgetter("config"))
}

So if the same config appears again after a different config (e.g. default, other, default), groupby yields default as two separate groups and the second one overwrites the first in the dict — the earlier shard URLs are silently lost. The same applies to the inner groupby over splits.

Fix

Sort the exported files by (config, split) before grouping, so all rows for a given config/split are consecutive. The sort is stable, so shard order within a split (0000.parquet, 0001.parquet, …) is preserved.

Tests

Added test_from_exported_parquet_files_keeps_all_shards_when_configs_non_consecutive — it feeds a non-consecutive default config and asserts both shard URLs survive (the test fails on main and passes with this change). The existing test_split_order_... test still passes (final ordering is still driven by dataset_infos). pytest tests/test_metadata_util.py → 9 passed; ruff check clean.

🤖 Generated with Claude Code