isthmus: SqlExpressionToSubstrait builds its base schema with the names-discarding toSubstrait, so nested field names are dropped

[nielspardon]

SqlExpressionToSubstrait has its own private toNamedStruct, and it builds the struct with the single-argument TypeConverter.toSubstrait(RelDataType):

for (Map.Entry<String, RelDataType> entry : nameToTypeMap.entrySet()) {
  names.add(entry.getKey());
  types.add(converterProvider.getTypeConverter().toSubstrait(entry.getValue()));
}
return NamedStruct.of(names, Type.Struct.builder().fields(types).nullable(false).build());

That overload allocates a throwaway list for the depth-first field names and never hands it back, so only the top-level column names reach names while the struct keeps every nested field. NamedStruct.of does not check the two against each other, so the malformed schema is written out:

new SqlExpressionToSubstrait().convert("n + 1", List.of("CREATE TABLE t (s ROW(x INT, y VARCHAR), n BIGINT)"))

// baseSchema.names = [S, N]          -- 2 names
// baseSchema.struct = struct{ struct{ i32, string }, i64 }   -- 4 depth-first name slots

TypeConverter.toNamedStruct on the same row type produces [s, x, y, n], which is what the plan path emits, so plans and extended expressions disagree on the same DDL. Reading the extended expression back and applying the names — toCalcite(factory, struct, names) — then indexes past the end of the list.

The fix is to call the existing TypeConverter.toNamedStruct rather than re-deriving the schema, which also removes one of the duplicate depth-first walkers noted in #1178.

Distinct from #1200 (the same class binds column indexes within their own table), #1201 and #1202.

Measured on main at 944b921.