[skyon0522-ai]

Fixes #490.

When result = result and (first() and second()) short-circuits because first() returns false, the compiler can leave the previous value in result. The issue's example returns true from the helper-call version while the direct-comparison version returns false. This change preserves the short-circuited value in the assignment destination, as required by Lua 5.1's logical-operator semantics.

Changes proposed in this pull request:

• Use the existing TESTSET path for a computed operand when an and branch exits the expression, so its false or nil value reaches the result register before the jump.
• Add a regression test for helper-call comparisons below, at, inside, and above the bounds. It fails before the fix and passes afterward.

Validation on revision 75f497656b1c6864139dd2a7d88cf96d09550814 with this patch:

• Focused regression: failing before, passing after; the original example and five boundary inputs were also checked through an external API host.
• make build, go test -mod=readonly ./..., and git diff --check passed.
• go vet -mod=readonly ./... reports six diagnostics (self-assignment and unreachable code). The unmodified revision produces the same six diagnostics with the same toolchain.

Checks used Go 1.27.2 on Linux/amd64. This is outside the repository's Go 1.24/1.25 CI matrix; those versions and other platforms were not run locally.

Prepared and checked with OpenAI Codex assistance.