fix(tlp): avoid false positives in TLP

[kumarUjjawal]

Summary

This PR fixes false positives in the TLP oracles.

The main change is that TLP WHERE and HAVING predicates now use a safe boolean generator. That keeps the predicates well-typed and valid for TLP checks.

The NoCrash path still keeps broader invalid-expression coverage, so this does not reduce the general crash-fuzzing goal.

Why

Before this change, the fuzzer could generate invalid predicates for TLP queries.

That caused cases where:

• the base query succeeded
• the partitioned TLP query failed during planning or type coercion

Those were reported like oracle failures, but they were fuzzer-side false positives rather than real DataFusion correctness bugs.

What changed

• restrict TLP predicate generation to safe boolean expressions
• keep the old broader expression generation for non-TLP fuzzing
• add regression coverage for valid boolean predicate execution
• update deterministic integration snapshots

[2010YOUY01]

Thank you for working on this!

The current expression generation strategy is to try to produce an expression of the target type on a best-effort basis. For example, generate_random_expr(BOOLEAN) returns a valid boolean expression about 80% of the time; in the remaining cases, it produces either an invalid expression or one with a different type.

This design is simpler and introduces more randomness, which makes it more likely to detect panics or crashes from invalid queries.

The tradeoff introduced by this PR is that we now need to implement a dedicated generator for valid boolean expressions. This also requires updating it whenever new boolean-returning expressions are added. In return, the valid query rate for TLP oracle testing improves from ~80% to 100%.

I’d like to think about this a bit more. For now, I prefer to keep things simple. However:

• If the query success rate becomes too low, this feature would be necessary.
• Alternatively, we could try improving the current “best-effort” generator to increase the valid type rate.

[kumarUjjawal]

Sounds good! I will explore this further.