A cidr or varchar column prunes no chunk groups, because its key is treated as cross-type against inet or text #1433

Found reviewing #1426, whose cd (cidr) range arm removes 0 chunk groups where the same values as inet remove 5. A sequel to #477: that fix added a cross-type path, and binary-coercible types fall through it.

The measurement

On PG17 at main, 12 row groups of contiguous values:

Both columns have complete zone maps (12 of 12, min and max). EXPLAIN still counts the filter in Columnar Pushed-Down Filters, which is #479's point. Answers are correct, because the executor rechecks the qual; the cost is a full read of the column.

Cause

A comparison on a cidr column resolves to inet >= inet. The planner side strips the RelabelType, so the scan key reaches pgcolumnar_make_predicates (src/columnar_reader.c) with a column of type cidr and an argument subtype of inet.

• getBaseType() doesn't unify binary-coercible types, so crossType is true (columnar_reader.c:885).
• GetDefaultOpClass(cidr, btree) gives inet_ops, family network_ops.
• get_opfamily_proc(network_ops, cidr, inet, BTORDER_PROC) finds nothing, because network_ops registers only (inet, inet). The code hits continue and drops the key.

varchar against text is the same case: text_ops registers (text, text), not (varchar, text).

Suggested fix

When the column type or the argument type is binary-coercible to the opclass's input type (get_opclass_input_type, IsBinaryCoercible), compare both as that type: treat the key as same-type and use the opclass's ordinary order proc. A test should cover cidr, varchar and a domain over each, both pruning and the answers against heap.

🤖 Generated with Claude Code