fix: return the constant instead of the first child when simplifying conjunctions

[sayanx64]

What problem were solved in this pull request?

Issue Number: close #637

Problem:

ConjunctionSimplificationRule folds a conjunction that contains a boolean constant: x AND false is always false and x OR true is always true. But instead of replacing the conjunction with that constant, it replaced it with the first child (child_exprs.front()), which is only correct when the constant happens to be the first child.

This is reachable from plain SQL. ComparisonSimplificationRule folds 1 = 2 into false, and because the grammar rule condition AND condition_list appends the first condition last, the constant ends up after id = 1:

create table t(id int, v int);
insert into t values(1, 10);
insert into t values(2, 20);
select * from t where 1 = 2 and id = 1;

Before (returns a row, and the scan is kept):

miniob > select * from t where 1 = 2 and id = 1;
id | v
1 | 10
miniob > explain select * from t where 1 = 2 and id = 1;
Query Plan
OPERATOR(NAME)
PROJECT
└─TABLE_SCAN(t)

After (empty, and the predicate subtree is pruned):

miniob > select * from t where 1 = 2 and id = 1;
id | v
miniob > explain select * from t where 1 = 2 and id = 1;
Query Plan
OPERATOR(NAME)
PROJECT

where id = 1 and 1 = 2 was already correct before this change, because there the constant happens to be the first child.

What is changed and how it works?

In ConjunctionSimplificationRule::rewrite:

• x AND false / x OR true: replace the conjunction with false / true instead of its first child.
• When every child is removed (true AND true, false OR false), replace the conjunction with true (AND) or false (OR). Before, an empty conjunction was left behind, and ConjunctionExpr::get_value returns true for an empty conjunction, so false OR false evaluated to true.
• Set change_made whenever the expression is modified (also pointed out in the issue).
• Continue from the iterator returned by vector::erase instead of reusing the invalidated one.

New unit test unittest/observer/conjunction_simplification_rule_test.cpp covers both short-circuit cases, the all-removed case, removing true from AND / false from OR, a single remaining child, and the no-op case. The first four tests fail on main.

Other information

• Thanks @what-is-me for the report and the test case. The change to the two short-circuit branches matches the fix linked in the issue; this PR also handles the empty-conjunction case, change_made and the iterator. The test from the issue description passes with this change.
• Why not return false/true when one of the children's expressions returns false/true in AND/OR directly in with the conjunction simplification rule #440 asks about the same code.
• Verified locally (Ubuntu, GCC 15.2, debug build with ASAN): the new unit test, the SQL above, and python3 test/case/miniob_test.py --test-cases=basic. BplusTreeLog.concurrency in bplus_tree_log_test aborts on my machine, but it aborts the same way on unmodified main (looks like [BUG] bplus_tree_log_test failed occasionally. #520), so it is unrelated to this change.

[sayanx64]

Hi @hnwyllmm, CI is green on this PR. Could you take a look when you have time?

While checking its first CI run, which failed in mvcc_trx_log_test (the flaky test from #442), I found the cause of that flake and opened #651 with a fix and a deterministic test.