[tanjiroK-coder]

ALTER TABLE t TRUNCATE PARTITION ALL passes the mayModifyData() read-only guard

StatementFeatureVisitor.visit(Alter) returned schema-only for every ALTER TABLE, whatever its actions. TRUNCATE and DROP carry MODIFIES_DATA so that the guard in the docs, if (connection.isReadOnly() && features.mayModifyData()), still stops them (see destructiveDdlAlsoModifiesData), but the ALTER TABLE actions that destroy rows did not: ALTER TABLE t TRUNCATE PARTITION ALL empties a partitioned MySQL table and reported certain=[MODIFIES_SCHEMA], and so did DROP PARTITION, EXCHANGE PARTITION (it swaps the rows of the partition and the other table) and DISCARD / IMPORT TABLESPACE (they delete or replace the table's data file). An action the grammar does not model is kept as raw tokens under AlterOperation.UNSPECIFIC, so ClickHouse's ALTER TABLE t DELETE WHERE 1 = 1 and ALTER TABLE t UPDATE c = 0 WHERE 1 = 1 also gave mayModifyData() == false.

Those partition and tablespace actions now report MODIFIES_DATA as certain, REPAIR PARTITION reports it as possible like REPAIR TABLE does, and an unmodelled action reports it as possible with the unsupported marker, the way an UnsupportedStatement or a CREATE FUNCTION body is treated. Of the 571 ALTER TABLE literals in the test suite, 17 now report MODIFIES_DATA as certain, 39 as possible (35 of them unmodelled actions such as PostgreSQL OWNER TO, SET LOGGED or ENABLE TRIGGER, an over-report on statements that already modify the schema) and 515 are unchanged, so is(..) changes for those 17 only. The 12 new StatementFeatureVisitorTest cases fail before and pass after; ./gradlew check (12439 tests) and mvn verify (12421 tests) pass.