Incorrect fold across CFG join block argument after inlining

[nkanazawa1989]

A fold after inlining appears to risk using a pre-branch initializer instead of the CFG join block argument. The reduced pattern is:

x = default
if flag:
    x = dynamic_value
return use(x)

After inlining, the IR correctly represents x as a join block argument. Any fold of use(x) after the join must use that block argument, not the initializer from before the branch.
Here is a pure-Kirin regression test:

from kirin import passes
from kirin.passes.inline import InlinePass
from kirin.prelude import basic


@basic
def helper(flag: bool, dynamic_value: float) -> float:
    x = 10.0
    if flag:
        x = dynamic_value + 1.0
    return x - 2.0


@basic
def main(flag: bool, dynamic_value: float) -> float:
    return helper(flag, dynamic_value)


def test_inline_fold_preserves_cfg_join_argument():
    assert main(True, 3.0) == 2.0
    assert main(False, 3.0) == 8.0

    InlinePass(main.dialects)(main)
    passes.Fold(main.dialects)(main)

    assert main(True, 3.0) == 2.0
    assert main(False, 3.0) == 8.0

The key invariant is that a fold after a CFG join must respect the join block argument. It must not replace the joined value with the pre-branch initializer unless the branch value is proven identical on all incoming edges.

#322 is possible related