Summary

#26337 routes definite == and != results through the shared equality evaluator before falling back to normal expression inference. When that evaluator returns ambiguous truthiness, however, we still run the full comparison dispatcher, repeating literal, tuple, and constraint-set work just to distinguish a builtin bool result from an arbitrary custom comparison return type.

This splits the top-level dispatch between equality and the other comparison operators. Equality evaluation now distinguishes definite truthiness, ambiguous builtin bool, and comparisons whose result type still needs to be inferred. Known builtin comparisons can stop after the shared evaluator, while custom comparisons continue through structural dispatch and the existing dunder fallback. The remaining literal dispatch uses a narrower non-equality operator enum, so it no longer duplicates the == and != cases.

To make the shared path complete, this also moves constraint-set equality and builtin tuple equality into the equality evaluator. Fixed-length tuples compare their elements using equality semantics even for an outer !=, while uncertain element results still fall back to expression inference so custom return types and bool-conversion diagnostics are preserved. Regression coverage verifies that equality and inequality over union arms retain each custom dunder return type.