Native differential + memory-safety harness: run the guest's Core code without a zkVM

[defenwycke]

docs/FUZZING.md names this "the highest-value next step" and says the project's central claim — that the guest's VerifyScript/sighash/libsecp and the C++ consensus math match real Core — is the one thing fuzzing does not cover. It also names the missing prerequisite: standing up $HAZYNC_BASE and compiling the guest's C++ for the host.

That prerequisite now exists.

What it does

fuzz-native/build.sh compiles 23 Core TUs plus secp256k1 for x86-64, mirroring prover/methods/guest/build.rs exactly minus the rv32im/ilp32 flags, and links verify_input.cpp — the guest's own shim — against it.

harness	result
differential.cpp	27 checks, 0 failures — subsidy vs an independent halving schedule, the 21M supply cap, CVE-2012-2459, retarget clamps saturating both ways, cumulative work
memsafety.cpp	661 cases, 0 sanitizer findings under ASan+UBSan — 352 trapped by the bounds check, 309 threw, both fail-closed

No GPU, no prover, no zkVM.

⛔ A correction to FUZZING.md's premise

It says "Core consensus code is portable". Core is; the provisioned tree is not, and both base patches are why:

patch	why a host build fails
0001	adds a Serialize(int) overload — on LP64 int is int32_t, so it is a redefinition
0002	routes CSHA256 through RISC0's accelerator — sys_sha_compress does not exist off-guest

Neither is optional; provision-vps.sh applies both to every tree. The harness overlays pristine copies of exactly those two files and leaves everything else as the guest sees it.

Limits stated, not buried

• ECMULT_WINDOW_SIZE is deliberately unset: the guest builds window 21 and regenerates the table, while the checked-in one hard-errors on a mismatch. The window is a speed/size trade that does not change what verification decides — but that stops being true if this is extended to compare timing.
• The independent references are written from the protocol rule. Where one is derived from the same reasoning as the implementation it is marked, because such a check catches transcription errors and not logic errors.
• Return values in the memory-safety pass are not asserted. A malformed tx may legitimately be rejected or parsed into nonsense; the only claim is that no wild read occurs.

It caught its own author three times

The initial CVE-2012-2459 vector was the wrong shape; SIGILL read as a crash when it is MiniReader failing closed as designed; and 156 SIGABRTs counted as findings when every one was an uncaught exception — a real ASan report would have been invisible among them.