HLIL drops zx when folding byte-width arithmetic, producing an incorrect expression

Version and Platform (required):

• Binary Ninja Version: Version 5.0.7648
• Edition: Non-Commercial
• OS: Ubuntu
• OS Version: 24.04
• CPU Architecture: x64

Bug Description:
HLIL discards zero- and sign-extensions when folding expressions, producing arithmetic that is numerically incorrect.

MLIL correctly models a narrowing step as its own statement, which justifies reducing a constant modulo the narrower width. HLIL folds the chain into a single expression and drops the extension, but keeps the reduced constant — so the reduction is no longer valid.

Expected Behavior:
target_zx should be

uint64_t target_zx(int32_t arg1, int32_t arg2) {
  return (uint8_t)(arg1 - 0x2402) + arg2 - 0x2402;
}

Example where the behaviour is wrong:
If arg1=0x2405 and arg2=0x2405, arg1 - 2 + arg2 - 0x2402 = 0x2406 whereas the expected returned value is 0x6.

How to reproduce:
Open x86_zx_drop or recompile x86_zx_drop.c and look at the functions target_*.

Screenshots/Video Recording:

Binary:
x86_zx_drop.zip

Additional Information:
Reproduces with zx and sx, and whether the extension comes from a byte-width instruction (movzx/movsx, MSP430 .b forms) or from a mask that MLIL canonicalises into an extension (and #0xff). Multiple extensions in one folded chain are all dropped.

Also reproduces on the current Binary Ninja build via Decompiler Explorer (dogbolt.org), so this is not specific to 5.0.7648.

Possibly related to #6371 (also HLIL and extension handling), though the symptom differs.

Open or recompile x86_zx_drop.c and inspect the target_* functions in both MLIL and HLIL, focusing on the zero- and sign-extension steps described in the report. Compare the generated HLIL with the expected target_zx expression and verify that zx and sx remain effective after folding, including the supplied argument values.