count() prunes the pushed projection to empty, which rendered a bare
SELECT * -- returning the raw, uncast schema. A non-Spark-native column
(e.g. Timestamp(NANOSECOND), unsigned) then reached the reader and failed
with UNSUPPORTED_ARROWTYPE. Only the empty-projection case leaked; the
all-columns (null projection) path already injected casts.

When the projection prunes to empty and the table has any cast column,
emit a single readable probe column instead of SELECT *
(SchemaConverter.probeColumn: prefer a castless column, else the first
column with its cast applied). A count/column-less scan only needs the row
count, and the emitted stream stays Spark-native. All-native tables keep
plain SELECT *.

Tests: SchemaConverterTest.probeColumn* (unit) and
AdbcSourceTest.typesCountWorks* (E2E, reproduces the failure); the example
pyspark harness gains a types-table count() assertion.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Spark's ArrowColumnVector backs ArrayType only from a variable ListVector,
never a FixedSizeListVector, so a fixed-size list reached the reader and
failed with UNSUPPORTED_ARROWTYPE FixedSizeList. SchemaConverter mapped it
to ArrayType (schema probe passed) but needsCast did not flag it, so no
arrow_cast was emitted; renderArrowType would also have produced an
identity FixedSizeList(N) target.

- needsCast: flag FixedSizeList unconditionally (the fixed layout is
  unreadable regardless of element type).
- renderArrowType: render FixedSizeList as a variable List(<elem>) so
  arrow_cast converts fixed->variable, with the element rendered cast-aware
  (e.g. FixedSizeList<Float16> -> List(Float32)).

Tests: SchemaConverterTest.fixedSizeListCastsToVariableList (unit); the
example `types` table gains a FixedSizeList<UInt16,2> `vec` column and
AdbcSourceTest / the pyspark harness assert it round-trips to Array<Integer>
with the element widened.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…ne authority

Enumerating "maps to a Spark type" and "readable by ArrowColumnVector"
by hand let them drift: a type could map to a Spark DataType yet have no
vectorized accessor, crashing at task time (FixedSizeList, and latent for
LargeList / FixedSizeBinary / Date64). Replace the three hand-kept switches
with a single authority.

- sparkConsumable(ArrowType): the read gate, mirroring Spark 4.0's
  ArrowColumnVector.initAccessor accessor set (verified against the class),
  deliberately the narrower gate.
- sparkTarget(Field): the nearest consumable field, recursing into children
  (identity when already consumable). One place normalizes unsigned->signed,
  Float16->Float32, non-us timestamp->us, Date64->Date32, Time->int,
  FixedSizeBinary->Binary, FixedSizeList/LargeList->List.

toSparkType, needsCast, and castTargetString now all derive from
sparkTarget, so the gates cannot disagree and adding a type is one case.
toSparkSchema asserts at plan time that every target is consumable, so an
unsupported type (e.g. Interval) fails in planning naming the column,
instead of an opaque executor UNSUPPORTED_ARROWTYPE.

Closes three latent bugs alongside FixedSizeList: LargeList, FixedSizeBinary,
and Date64 previously mapped to a Spark type without a cast.

Tests: SchemaConverterTest gains cases for LargeList/FixedSizeBinary/Date64,
a "sparkTarget is always consumable" invariant, and a fail-fast case; the
example `types` table gains labels (LargeList<Utf8>), digest
(FixedSizeBinary), and day (Date64) columns, asserted end-to-end in
AdbcSourceTest and the pyspark harness. The E2E session enables the Java 8
date API so DateType decodes without needing sun.util.calendar opened.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>