Custom loader silently omits a declared-but-absent column

[jepegit]

Summary

Found while adding the deliberately-broken fixture set for #655.

When a raw file is missing a column that the instrument definition's normal_headers_renaming_dict declares (e.g. voltage), the custom loader loads successfully and silently drops that column rather than raising or warning. The failure only surfaces much later, downstream, as a KeyError when something accesses the (now absent) column.

Reproduce

testdata/bad/custom_missing_column.csv (voltage column removed) loaded via the custom instrument with custom_instrument_001.yml:

• loads without error, 60 rows
• voltage (legacy) / potential (native) is absent from data.raw
• accessing it raises KeyError

Pinned as current behavior in tests/test_bad_fixtures.py::test_missing_required_column_omits_it_silently.

Desired

Loading a file that lacks a declared column should surface a clear error or warning naming the missing column at load time, not drop it silently. The desired contract is asserted (currently xfail) in tests/test_bad_fixtures.py::test_missing_column_should_surface_a_clear_error — when this is fixed, remove the xfail.

Notes

Low-risk, well-scoped: the fix is a validation step in the custom loader's header-renaming path. Part of the #655 Phase-2 robustness follow-ups.