Preserve PostgreSQL substring offset semantics

[fallintoplace]

Summary

pg_clickhouse translates PostgreSQL substring functions to ClickHouse for remote execution:

PostgreSQL overload	ClickHouse function
substring(text, ...) / substr(text, ...)	substringUTF8
substring(bytea, ...) / substr(bytea, ...)	substring

Those mappings are only semantically safe for a subset of PostgreSQL bounds. This change retains pushdown for safe constant bounds and evaluates every other case locally in PostgreSQL.

PostgreSQL and ClickHouse differ

The difference affects both mappings, not just text values:

Type	PostgreSQL expression	PostgreSQL result	ClickHouse behavior before this change
text	substring('val1' FROM -1)	val1	A negative start counts from the end
text	substring('val1' FROM 0 FOR 2)	v	Empty result
text	substring('val1' FROM 2 FOR -1)	Error: negative substring length not allowed	Returns a shortened value
bytea	substring('val1'::bytea FROM -1 FOR 3)	v	A negative start counts from the end
bytea	substring('val1'::bytea FROM 0 FOR 2)	v	Empty result
bytea	substring('val1'::bytea FROM 2 FOR -1)	Error: negative substring length not allowed	Returns a shortened value

Unconditionally shipping these calls changes results and can suppress a PostgreSQL error.

Approach

Push down substring and substr only when their bounds are known-safe constants at planning time:

Function form	Pushed down when	Evaluated locally when
substring(value, start) / substr(value, start)	start is a non-NULL constant >= 1	start is NULL, non-constant, zero, or negative
substring(value, start, length) / substr(value, start, length)	start >= 1 and length >= 0, both non-NULL constants	Either bound is NULL, non-constant, or outside that range

The guard is applied consistently to text and bytea overloads. Safe cases such as substring(val, 1, 3) and substr(val, 2) continue to use ClickHouse. A parameterized offset is evaluated locally because its value is unavailable when the remote expression is planned.

Tests

Regression coverage includes:

• text values with negative and zero starts
• bytea values with a negative start
• PostgreSQL's negative-length error
• a generic prepared plan with a parameterized offset
• EXPLAIN output confirming unsafe calls stay out of the remote SQL

Verified with:

make TESTS=test/sql/functions.sql PGUSER=postgres installcheck

on PostgreSQL 19 against ClickHouse 26.3.17.56 and 23.3.22.3.

[serprex]

can we push down conditional logic to match postgres instead?

cc @JoshDreamland who recently did this sort of thing to get correct NULL handling with IN in #317