nx_self_host_load_codegen_repro.nx

nx_self_host_load_codegen_repro.nx -- MINIMAL REPRO for the self-host codegen bug surfaced by the bootstrap-stability audit + fixed-point audit on 2026-05-21. Bug class: the native x86_64 NishiLang binary (built via the bootstrap loop from the qemu-RV64 substrate) emits INCORRECT asm for two specific patterns: A. return <local-variable>: Expected: movq <slot>(%rbp), %rax ; ret Got: movabsq $0, %rax ; ret <-- WRONG: returns 0 B. call <named-function>: Expected: call sys_mmap (or other name) Got: call <garbage-bytes-as-string> <-- WRONG: name corrupted Both fail patterns involve reading from memory that was written earlier (the stack slot for the local; the name-string buffer for the function name). Common factor: memory-load codegen emitted by the substrate-compiled native binary is broken somehow. The qemu-RV64 substrate emits these patterns CORRECTLY for user code, so the bug is baked into the native binary itself by the RV64-qemu compiler when it emits nx_compile_x86.nx + nx_opt.nx + nx_x86_64_ctx.nx asm. This source is itself a TEST PROGRAM (returns expected 18 = 7+11) AND a self-host-bootstrap regression KAT. When the substrate codegen bug is fixed: 1. The bash test runs this binary natively, expects exit 18. 2. The bootstrap-stability audit (bench/nx_bootstrap_stability_audit.sh) passes for nx_u8_load_test.nx etc. 3. The fixed-point audit (bench/nx_self_host_fixed_point_audit.sh) returns PASS (stage_2.s == stage_3.s). Until that bug is fixed, this KAT is the SMALLEST repro: 7-line function with two locals + return-sum. expect_exit: 18 license_tier: ORIGINAL