fix: replace fill-value sentinels in HDF5 compound datasets

[Aman-Cool]

Related Issues / Pull Requests

Follow-up to #61, which closed #60.

Description

#61 fixed fill-value replacement for flat numeric HDF5 datasets by checking data.dtype.kind in ("f", "i", "u"). That guard intentionally skipped compound/structured datasets (dtype.kind == "V") — so they never go through _collect_fill_values, and their sentinel values (like -9999.0) land in the DataFrame as real data.

This is the exact same bug from #60, just for compound-typed datasets. The symptom is identical: completeness reports 100% on fields that have missing entries, and outlier detection spikes because the sentinel value is hundreds of standard deviations from the actual data.

The fix runs after pd.DataFrame(data) is built for a compound dataset; checks dtype.kind == "V", calls the existing _collect_fill_values method, and replaces matched sentinels with NaN on each numeric column. No new logic, just extends what #61 already set up to cover this case too.

Quick sanity check with a compound dataset (depth, temperature), _FillValue = -9999.0, one sentinel in temperature across 4 rows:

• before the fix: isna().sum() -> 0, completeness -> 100%
• after: isna().sum() -> 1, completeness -> 75%

What changes are proposed in this pull request?

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality not to work as expected; for instance, examples in this repository must be updated too)
• This change requires a documentation update

Checklist:

• My code modifies existing public API, or introduces new public API, and I updated or wrote documentation
• I have commented my code
• My code requires documentation updates, and I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[Aman-Cool]

@jeanbez @sbyna @kaveenh, Noticed this while stress-testing #61 with some real-world scientific HDF5 files, compound datasets were still coming back 100% complete even when they clearly had missing entries.
Took a bit to track down but the guard in #61 only covered flat numeric arrays, compound types just slipped right through. Reused _collect_fill_values since the logic was already there, just needed to hook it in after the DataFrame gets built.