fix: DateTimeParts propagates null component rows instead of throwing (#235)

[dfa1]

What

vortex.datetimeparts splits a date/time into component children (days + seconds + subseconds). DateTimePartsArrays.readLong assembled them per row and threw VortexException: DateTimeParts: null cell at index 0 whenever a component decoded to a null.

That crash surfaced after PR #225 taught RunEndEncodingDecoder to propagate nulls: a null run now returns a MaskedArray (previously it decoded as a filler value like 0). A datetimeparts column whose day-component is a nullable RunEnd child then hit the throw, blocking export of the Raincloud slugs bi-yalelanguages and bi-euro2016.

Fix

In DateTimePartsEncodingDecoder.decode, mirror the AlpEncodingDecoder (#210) unwrap+re-wrap pattern:

1. After decoding each component, if it is a MaskedArray, capture its validity and unwrap to .inner().
2. Intersect the component validities (a row is valid only when every part is valid — matching the Rust reference: each part must be present to reassemble the epoch count).
3. Build the LazyDateTimePartsLongArray over the unwrapped raw values.
4. If any component contributed nulls, wrap the reassembled array in a MaskedArray carrying the combined mask.

DateTimePartsArrays.readLong no longer throws on a masked cell — it recurses to the raw value; null rows are now tracked by the reassembled array's own outer mask (downstream ExtensionStorage/decodeAll already gate on that).

Tests

• New decode_nullDaysComponent_propagatesNullRowInsteadOfThrowing: a stub days-child returning a MaskedArray with row 0 null yields a masked result (row 0 invalid, row 1 valid and reassembled) instead of throwing.
• Updated readLong_nullMaskedCell_* to assert the new non-throwing read-through behavior.
• Full reader suite green (1190 tests).

Matrix updates for the two corpus slugs are handled separately — this PR just removes the crash.

Fixes #235.

🤖 Generated with Claude Code