fix: traverse missing expression and query children in visitor adapters

[AndreasLanz01]

What

ExpressionVisitorAdapter and SelectVisitorAdapter skipped some child expressions and sub-selects. Every visitor built on them, including StatementFeatureVisitor, therefore never saw a function call or subquery in these positions:

Node	Child that was not visited
Column	array subscript and slice: a[f()], t.a[1:f()]
Function	named parameters: substring(x FROM 2 FOR 3), position(x IN y), overlay(x PLACING y FROM 2); an attribute expression
TrimFunction	the FROM expression: trim(LEADING 'x' FROM f())
AnyComparisonExpression	the sub-select: = ANY (...), > ALL (...), = SOME (...)
LikeExpression	ESCAPE expression
MemberOfExpression	left expression
TimezoneExpression	the time zone: d AT TIME ZONE f()
JsonExpression	operands after ->, ->>, #>, #>>
JsonAggregateFunction	key and value of json_objectagg(k : v), its ORDER BY, OVER (PARTITION BY .. ORDER BY ..)
Values (select path)	all rows: top-level VALUES (...), IN (VALUES (...))
PlainSelect	LIMIT, LIMIT .. BY, START WITH .. CONNECT BY
SetOperationList	WITH, ORDER BY, LIMIT, OFFSET, FETCH
ParenthesedSelect	OFFSET / FETCH got a null context instead of the caller's

Why

StatementFeatureVisitor.analyse(statement, pureFunctions) is meant for read-only guards: a function that is not proven pure leaves MODIFIES_DATA / MODIFIES_SCHEMA possible and is reported in getUnresolvedReferences(). With pg_read_file not declared pure, each of these reported nothing before this change:

SELECT substring(pg_read_file('x') FROM 2 FOR 3)
SELECT trim(LEADING 'x' FROM pg_read_file('x'))
SELECT a FROM t WHERE a = ANY (SELECT pg_read_file('x'))
VALUES (pg_read_file('x'))
SELECT a FROM t LIMIT length(pg_read_file('x'))
SELECT d AT TIME ZONE pg_read_file('x') FROM t
SELECT j -> pg_read_file('x') FROM t
SELECT a[length(pg_read_file('x'))] FROM t
SELECT 1 UNION SELECT 2 LIMIT length(pg_read_file('x'))

The same gap hid row locks and data-modifying CTEs in these positions, e.g. WHERE a = ANY (SELECT b FROM u FOR UPDATE) reported no MODIFIES_TRANSACTION, and VALUES ((WITH c AS (DELETE FROM u RETURNING b) SELECT b FROM c)) reported no MODIFIES_DATA.

The fix is in the adapters, not in StatementFeatureVisitor, so every subclass of the adapters sees these children too.

Testing

• StatementFeatureVisitorTest: new nested class ExpressionPositions:

  • 29 positions where a function must be reported unresolved (and nothing is reported when it is declared pure);
  • 5 positions where a nested FOR UPDATE must give MODIFIES_TRANSACTION;
  • 2 positions where a nested data-modifying CTE must give MODIFIES_DATA;
  • a control that top-level VALUES (1, 2) still reads nothing and returns rows.

  On base 26d3e02a all 36 parameterized cases fail and the control passes.

• ./gradlew check on Windows 11 with the JDK 17 toolchain: 9,579 test cases, 0 failures, 0 errors, 25 skipped. Grammar ambiguity, Spotless, Checkstyle, PMD, SpotBugs and JaCoCo verification pass.

• Not run: mvn verify, JMH (no grammar or parser change), Linux/macOS.

Not changed

• FromItemVisitorAdapter.visit(TableFunction) is still empty; table functions in FROM are handled in fix: inspect every table function in feature analysis and validation #2760.
• PlainSelect FOR UPDATE OF .. / WAIT .. and lateral views are still not visited (existing @todos).

Origin

First observed with com.manticore-projects.jsqlformatter:jsqlparser:5.4.104 and Dialect.POSTGRESQL. Reproduced on upstream master 26d3e02a78e061f5fecf2eb2e783bf0482e10840 with the default parser. No existing issue is linked.

AI disclosure: this change was prepared with an AI coding agent (Claude Code) and reviewed by me.

[manticore-projects]

Thank you very much for improving this!