lez-fuzzing/docs/fuzzing.md

# Fuzzing Guide

This document covers how to run fuzz targets, add new targets, minimise failures,
and convert findings into regression tests.

The fuzzing infrastructure lives in a **separate repository** (`lez-fuzzing/`) which
reads the Logos Execution Zone (LEZ) codebase from `../logos-execution-zone/` (a sibling
directory that must be cloned separately).

---

## Architecture

The fuzz workspace (`fuzz/`) is a single Cargo workspace that covers **both** fuzzing
engines via Cargo features.  No separate Cargo manifest is needed.

| | libFuzzer lane | AFL++ lane |
|---|---|---|
| **Build command** | `cargo fuzz build <TARGET>` | `cd fuzz && cargo afl build --no-default-features --features fuzzer-afl --release --bin <TARGET>` |
| **Run command** | `cargo fuzz run <TARGET>` | `afl-fuzz -i fuzz/corpus/<TARGET> -o afl-output/<TARGET> -- fuzz/target/release/<TARGET>` |
| **Cargo feature** | `fuzzer-libfuzzer` (default) | `fuzzer-afl` |
| **Harness entry** | `::libfuzzer_sys::fuzz_target!(…)` | `fn main() { ::afl::fuzz!(…) }` |
| **`main()` presence** | Suppressed via `#![no_main]` | Required; provided by `afl::fuzz!` |
| **`fuzz/Cargo.toml`** | ✅ Source of truth | ✅ Same file — covers both lanes |

The engine is selected at the call site via the `fuzz_props::fuzz_entry!` macro:

```rust
#![cfg_attr(feature = "fuzzer-libfuzzer", no_main)]

fuzz_props::fuzz_entry!(|data: &[u8]| {
    // … harness body …
});
```

The `cfg` attributes in the macro expansion resolve against the **calling crate's** features
(`fuzz/`), not `fuzz_props`'s features.

---

## Prerequisites

```bash
# libFuzzer lane
rustup install nightly
rustup component add llvm-tools-preview --toolchain nightly
cargo install cargo-fuzz

# AFL++ lane (additional)
# macOS:
brew install afl-fuzz

# Linux — build from source (apt packages are several major versions behind):
git clone https://github.com/AFLplusplus/AFLplusplus
cd AFLplusplus && make distrib && sudo make install
cd ..

# Rust wrapper (all platforms):
cargo install cargo-afl
```

---

## Repository Setup

`lez-fuzzing` is a **standalone repository** — it does **not** use git submodules.
It expects the LEZ codebase to be cloned at `../logos-execution-zone` relative to itself.

```bash
# Clone both repositories side-by-side into the same parent directory:
git clone <LEZ_REPO_URL>           logos-execution-zone
git clone <LEZ_FUZZING_REPO_URL>   lez-fuzzing

# The directory layout must be:
#   <parent>/
#   ├── logos-execution-zone/
#   └── lez-fuzzing/
```

---

## How to Run

All fuzz targets must be run with `RISC0_DEV_MODE=1` to disable expensive ZK
proof generation. The `just` recipes handle this automatically.

```bash
# From lez-fuzzing/

# Run all targets for 30 s each (libFuzzer)
just fuzz

# Run a specific target for 120 s (libFuzzer)
RISC0_DEV_MODE=1 cargo fuzz run fuzz_state_transition -- -max_total_time=120

# Run the saved corpus (regression mode, no mutations)
just fuzz-regression
```

---

## Available Fuzz Targets

| Target | What it fuzzes | Entry point |
|--------|---------------|-------------|
| `fuzz_transaction_decoding` | Borsh decoding of `NSSATransaction`, `Block`, and `HashableBlockData`; roundtrip re-encoding of successfully decoded transactions | `fuzz/fuzz_targets/fuzz_transaction_decoding.rs` |
| `fuzz_stateless_verification` | `transaction_stateless_check()` no-panic on arbitrary bytes; idempotency — a transaction that passes the check must pass it again | `fuzz/fuzz_targets/fuzz_stateless_verification.rs` |
| `fuzz_state_transition` | `execute_check_on_state()` across up to 8 transactions with fuzz-driven initial state and monotonically-advancing block context; asserts **StateIsolationOnFailure** (balances unchanged on rejection), **BalanceConservation** (total balance unchanged on success), and **ReplayRejection** (nonce consumed on first acceptance) | `fuzz/fuzz_targets/fuzz_state_transition.rs` |
| `fuzz_block_verification` | Three block-hash invariants: **HashRoundTrip** (`HashableBlockData::from(Block)` is lossless), **HashPreimage** (block_id, prev_block_hash, timestamp each individually affect the hash), **TxOrderCommitment** (reversing the transaction list changes the hash) | `fuzz/fuzz_targets/fuzz_block_verification.rs` |
| `fuzz_encoding_roundtrip` | `decode(encode(tx)) == Ok(tx)` and `encode(decode(encode(tx))) == encode(tx)` for `PublicTransaction` and `ProgramDeploymentTransaction`; raw bytes that decode successfully must re-encode identically (canonical encoding) | `fuzz/fuzz_targets/fuzz_encoding_roundtrip.rs` |
| `fuzz_signature_verification` | Signature correctness (sign→verify), no-panic on random bytes, cross-key soundness | `fuzz/fuzz_targets/fuzz_signature_verification.rs` |
| `fuzz_replay_prevention` | A tx accepted in block N must be rejected when replayed in block N+1 (nonce consumed); fuzz-driven initial state exposes nonce edge cases (nonce 0, `u128::MAX`, zero-balance sender) | `fuzz/fuzz_targets/fuzz_replay_prevention.rs` |
| `fuzz_state_diff_computation` | **Forward containment**: `ValidatedStateDiff` only modifies accounts declared in `affected_public_account_ids()`; **Reverse completeness**: every declared account actually modified by `execute_check_on_state` appears in the diff | `fuzz/fuzz_targets/fuzz_state_diff_computation.rs` |
| `fuzz_validate_execute_consistency` | `validate_on_state` and `execute_check_on_state` must agree on success/failure; diff accuracy (forward + reverse); **BalanceConservation** on success | `fuzz/fuzz_targets/fuzz_validate_execute_consistency.rs` |
| `fuzz_state_serialization` | `V03State` Borsh no-panic (**NoPanic**) + **StateSerializationRoundtrip** (`encode(decode(encode(decode(data)))) == encode(decode(data))`) + **NullifierDeduplication** (hand-written `NullifierSet` deserializer returns `Err`, not panic, on duplicate nullifiers) | `fuzz/fuzz_targets/fuzz_state_serialization.rs` |
| `fuzz_witness_set_verification` | `WitnessSet::is_valid_for` no-panic on adversarial input (**NoPanic**); **CorrectVerification** (`WitnessSet::for_message` always passes `is_valid_for` on the same message); **MessageIsolation** (witness set built for message A fails `is_valid_for` on any Borsh-distinct message B) | `fuzz/fuzz_targets/fuzz_witness_set_verification.rs` |
| `fuzz_program_deployment_lifecycle` | `V03State::transition_from_program_deployment_transaction` no-panic on arbitrary WASM bytecode (**NoPanic**); **BalanceIsolation** (successful deployment must not move tokens); **StateIsolationOnFailure** (failed deployment must not change any genesis account balance or nonce) | `fuzz/fuzz_targets/fuzz_program_deployment_lifecycle.rs` |
| `fuzz_apply_state_diff_split_path` | **SplitPathEquivalence**: for every known account, `validate_on_state` + `apply_state_diff` must produce exactly the same balance, nonce, data, and program_owner as `execute_check_on_state`; **NonceIncrementCorrectness**: nonce after the split path equals nonce after the direct path for all signer accounts (catches bugs in the two-step `apply_state_diff` nonce-increment logic) | `fuzz/fuzz_targets/fuzz_apply_state_diff_split_path.rs` |
| `fuzz_multi_block_state_sequence` | **LongRangeBalanceConservation**: total genesis-account balance identical before and after all N (≤ 16) blocks; **FailedTxNonceStability**: every genesis-account nonce unchanged after a rejected transaction; **PerBlockReplayRejection**: every transaction accepted in block B is rejected in block B+1 (cumulative nonce-interaction coverage) | `fuzz/fuzz_targets/fuzz_multi_block_state_sequence.rs` |
| `fuzz_sequencer_vs_replayer` | **SequencerReplayerEquivalence**: for every known account (genesis ∪ diff-declared), the sequencer path (`validate_on_state` → `apply_state_diff`) and the replayer path (`execute_check_on_state`) must produce identical balance, nonce, data, and program_owner after applying a full block of up to 8 transactions plus the mandatory clock invocation; **ReplayerAcceptsAllSequencerTxs**: every transaction accepted by `validate_on_state` must also be accepted by `execute_check_on_state`; **ClockConsistency**: the mandatory clock invocation must succeed on both paths and leave both states identical | `fuzz/fuzz_targets/fuzz_sequencer_vs_replayer.rs` |

---

## How to Add a New Fuzz Target

### Step 1 — Scaffold with `just new-target`

```bash
just new-target my_feature
```

This single command does four things automatically:

| What | Where |
|---|---|
| Creates the corpus directory | `fuzz/corpus/fuzz_my_feature/` |
| Writes a typed fuzz target template | `fuzz/fuzz_targets/fuzz_my_feature.rs` |
| Appends `[[bin]]` entry to `fuzz/Cargo.toml` | Covers **both** the libFuzzer and AFL++ lanes |
| Inserts target into every CI matrix + perf loop | `.github/workflows/fuzz.yml` |

The generated template uses `fuzz_props::fuzz_entry!` and works with both engines
without modification.

### Step 2 — Implement the target

Edit `fuzz/fuzz_targets/fuzz_my_feature.rs`.  Replace the placeholder with the
function under test and any invariant assertions.  Use the typed wrappers from
[`fuzz_props::arbitrary_types`](../fuzz_props/src/arbitrary_types.rs) for
structured input, or the proptest generators from
[`fuzz_props::generators`](../fuzz_props/src/generators.rs) for richer strategies.

### Step 3 — Automated registration (cargo-fuzz + CI)

`just new-target` calls [`scripts/add_fuzz_target.py`](../scripts/add_fuzz_target.py)
which:
- Appends the `[[bin]]` entry to [`fuzz/Cargo.toml`](../fuzz/Cargo.toml).
  This **single entry** covers both the libFuzzer lane (`cargo fuzz build`) and
  the AFL++ lane (`cargo afl build --no-default-features --features fuzzer-afl`).
- Inserts the target name into every strategy matrix and the perf-baseline shell
  loop in [`.github/workflows/fuzz.yml`](../.github/workflows/fuzz.yml).

> **Manual fallback:** if you create a target without `just new-target`, add the
> entry yourself:
>
> ```toml
> [[bin]]
> name = "fuzz_my_feature"
> path = "fuzz_targets/fuzz_my_feature.rs"
> test = false
> bench = false
> ```

### Step 4 — Verify

```bash
# Verify the libFuzzer build
RISC0_DEV_MODE=1 cargo fuzz build fuzz_my_feature
just fuzz-regression   # runs the new target against its (empty) corpus

# Verify the AFL++ build (same fuzz/Cargo.toml — no separate manifest needed)
cd fuzz && cargo afl build \
  --no-default-features \
  --features fuzzer-afl \
  --release \
  --bin fuzz_my_feature
```

### Quick reference: what to touch

| File | Action | Automated? |
|---|---|---|
| `fuzz/fuzz_targets/fuzz_<name>.rs` | Create | ✅ `just new-target` |
| `fuzz/corpus/fuzz_<name>/` | Create | ✅ `just new-target` |
| `fuzz/Cargo.toml` | Add `[[bin]]` (covers both lanes) | ✅ `just new-target` |
| `Justfile` | Nothing — auto-discovers | ✅ automatic |
| `.github/workflows/fuzz.yml` | Add to 3 matrix lists | ✅ `just new-target` |

---

## AFL++ Parallel Fuzzing Lane

### Prerequisites

Install AFL++ natively on your machine.

> **Note on Linux package versions**: The `afl++` package in Debian stable (Bookworm)
> and Ubuntu LTS is several major versions behind the current AFL++ 4.x series and may
> be incompatible with `cargo-afl`. **Build from source** for a current version.

```bash
# macOS — Homebrew keeps the formula up to date
brew install afl-fuzz

# Linux — build from source (~5 min)
git clone https://github.com/AFLplusplus/AFLplusplus
cd AFLplusplus
make distrib        # builds all components: afl-fuzz, afl-cc, afl-clang-fast, …
sudo make install
cd ..

# Rust build wrapper (all platforms)
cargo install cargo-afl
```

> **macOS: crash reporter must be disabled** — AFL++ detects the macOS `ReportCrash`
> daemon and aborts if it is active (it delays crash notifications and causes AFL++ to
> mis-classify crashes as timeouts).  The `just fuzz-afl` and `just fuzz-afl-parallel`
> recipes disable it automatically for the duration of the run and re-enable it on exit
> (via a shell `trap`).  You can also manage it manually:
>
> ```bash
> # Disable (run once before a long session)
> just afl-macos-setup
>
> # Re-enable afterward
> just afl-macos-teardown
> ```
>
> Or use the raw `launchctl` commands shown in the AFL++ error message:
>
> ```bash
> SL=/System/Library; PL=com.apple.ReportCrash
> launchctl unload -w ${SL}/LaunchAgents/${PL}.plist
> sudo launchctl unload -w ${SL}/LaunchDaemons/${PL}.Root.plist
> ```

### Build

```bash
# All targets
just afl-build

# Single target
just afl-build-target fuzz_state_transition
```

Both commands compile `fuzz/` with `--no-default-features --features fuzzer-afl`.
Output binaries land in `fuzz/target/release/`.

### Run (single instance)

```bash
# 120-second smoke run
just fuzz-afl fuzz_state_transition

# Custom duration
just fuzz-afl fuzz_state_transition 600
```

### Run (parallel)

```bash
# 1 main + 3 secondary instances for 5 minutes
just fuzz-afl-parallel fuzz_state_transition 4 300

# AFL++ rule: always start the main instance first;
# secondary instances are started with -S flags automatically.
```

### Monitor

```bash
just afl-status fuzz_state_transition
# … calls afl-whatsup afl-output/fuzz_state_transition
```

### Triage

```bash
# Minimise a crash artifact to the smallest reproducing input
just afl-tmin fuzz_state_transition afl-output/fuzz_state_transition/default/crashes/id:000000,...

# Pretty-print as Rust byte literal (for pasting into a unit test)
just afl-fmt afl-output/fuzz_state_transition/default/crashes/id:000000,...
```

### Sync queue to shared corpus

```bash
# Copies afl-output/*/queue/id:* into fuzz/corpus/<target>/
# Run this after any AFL++ session to share findings with cargo-fuzz
just afl-corpus-sync
```

### How the shared harness works

| Mechanism | libFuzzer | AFL++ |
|---|---|---|
| **Entry macro** | `::libfuzzer_sys::fuzz_target!(…)` | `::afl::fuzz!(…)` inside `fn main()` |
| **`no_main` suppression** | `#![cfg_attr(feature = "fuzzer-libfuzzer", no_main)]` | Not applied (AFL++ needs a real `main`) |
| **Feature gate** | `cfg(feature = "fuzzer-libfuzzer")` | `cfg(feature = "fuzzer-afl")` |
| **Feature resolution** | Resolved at `fuzz/` (calling crate), not at `fuzz_props/` | Same |
| **`libfuzzer-sys` dep** | Optional, active under `fuzzer-libfuzzer` | Not compiled — avoids `main()` conflict |
| **`afl` dep** | Not compiled | Optional, active under `fuzzer-afl` |
| **Default build** | `default = ["fuzzer-libfuzzer"]` → `cargo fuzz` just works | Requires `--no-default-features --features fuzzer-afl` |

The `fuzz_props::fuzz_entry!` macro defined in [`fuzz_props/src/lib.rs`](../fuzz_props/src/lib.rs)
expands to the right entry point based on the active feature:

```rust
#[macro_export]
macro_rules! fuzz_entry {
    (|$data:ident: &[u8]| $body:block) => {
        #[cfg(feature = "fuzzer-libfuzzer")]
        ::libfuzzer_sys::fuzz_target!(|$data: &[u8]| $body);

        #[cfg(feature = "fuzzer-afl")]
        fn main() {
            ::afl::fuzz!(|$data: &[u8]| $body);
        }
    };
}
```

### CI (`.github/workflows/fuzz-afl.yml`)

The nightly AFL++ CI workflow has two jobs:

| Job | Triggers | Matrix |
|-----|----------|--------|
| `afl-smoke` | nightly + `workflow_dispatch` | 7 priority targets, 120 s each |
| `afl-coverage` | nightly, `needs: afl-smoke` | 3 key targets; LLVM HTML report |

The smoke job:
1. Builds the target with `cargo afl build --no-default-features --features fuzzer-afl`
2. Runs `afl-fuzz` for 120 s in `aflplusplus/aflplusplus:v4.40c` container
3. Syncs new queue entries into `fuzz/corpus/<target>/` and opens a corpus PR
4. Uploads crashes/hangs as a workflow artifact

The coverage job:
1. Downloads the smoke findings
2. Rebuilds with `RUSTFLAGS="-C instrument-coverage"`
3. Runs all corpus + queue inputs through the binary
4. Merges `.profraw` → `.profdata` → HTML report via `llvm-cov show`

---

## Updating the LEZ Dependency

`lez-fuzzing` reads LEZ source directly from `../logos-execution-zone`. To pick up LEZ
changes, simply update that repo:

```bash
cd ../logos-execution-zone
git pull --ff-only
cd ../lez-fuzzing

# Rebuild to confirm compatibility:
cargo build -p fuzz_props
RISC0_DEV_MODE=1 cargo fuzz build
```

The `just update-lez` recipe automates the pull:

```bash
just update-lez
```

---

## Minimising & Reproducing Failures

When `cargo fuzz` finds a crash it writes an artifact to
`fuzz/artifacts/fuzz_<target>/crash-<hash>`.

### Minimise (libFuzzer)

```bash
# Produces a smaller input that still triggers the same crash
just fuzz-tmin fuzz_state_transition fuzz/artifacts/fuzz_state_transition/crash-abc123
```

### Minimise (AFL++)

```bash
just afl-tmin fuzz_state_transition afl-output/fuzz_state_transition/default/crashes/id:000000,...
```

### Convert to a regression test

```bash
# libFuzzer: print bytes as a Rust byte-literal
cargo fuzz fmt fuzz_state_transition fuzz/artifacts/fuzz_state_transition/crash-abc123

# AFL++: print bytes as a Rust byte-literal
just afl-fmt afl-output/fuzz_state_transition/default/crashes/id:000000,...
```

Add the minimised file to the corpus so CI always reproduces it:

```bash
cp fuzz/artifacts/fuzz_state_transition/crash-abc123-minimised \
   fuzz/corpus/fuzz_state_transition/regression_001
```

Open a PR. The `regression` CI job will permanently block re-introduction of this bug.

---

## Coverage Reports

### Step 1 — libFuzzer coverage (via `cargo fuzz coverage`)

```bash
# Generates coverage for a single target
cargo fuzz coverage fuzz_state_transition

# Generates coverage for all targets
just coverage-all
```

Reports land in `fuzz/coverage/<target>/`.

### Step 2 — AFL++ LLVM coverage

Run after a successful AFL++ session (queue data in `afl-output/<target>/`):

```bash
# Combines libFuzzer + AFL++ corpus into a single LLVM HTML report
just coverage fuzz_state_transition
```

This:
1. Runs `cargo fuzz coverage` (step 1)
2. Detects `afl-output/fuzz_state_transition/` and builds the target with
   `RUSTFLAGS="-C instrument-coverage" cargo build --manifest-path fuzz/Cargo.toml --no-default-features --features fuzzer-afl --release`
3. Runs all AFL++ queue entries through the binary, collects `.profraw` files
4. Merges profiles with `llvm-profdata merge` and generates an HTML report with `llvm-cov show`
5. Writes the report to `coverage/afl/fuzz_state_transition/html/index.html`

The AFL++ CI coverage job (`afl-coverage` in [`.github/workflows/fuzz-afl.yml`](../.github/workflows/fuzz-afl.yml))
automates steps 2–5 and uploads the report as a workflow artifact.

---

## Invariant Framework

Shared invariants live in `fuzz_props/src/invariants.rs`. Each invariant implements
`ProtocolInvariant` and is automatically run by `assert_invariants()`.

Concrete invariants currently registered in `assert_invariants()`:

| Invariant | Description | Implementation status |
|-----------|-------------|----------------------|
| `StateIsolationOnFailure` | Per-account balance must not change for any tracked account when a transaction is rejected | ✅ Fully implemented |
| `BalanceConservation` | Total balance of all known accounts must be conserved when a transaction succeeds | ✅ Fully implemented |
| `FailedTxNonceStability` | Every account's nonce must remain unchanged when a transaction is rejected | ✅ Fully implemented |
| `ReplayRejection` | An accepted transaction must be rejected when replayed | ⚠️ Registry stub — always returns `None` from `InvariantCtx`; use `assert_replay_rejection()` directly (see note below) |
| `NonceIncrementCorrectness` | Every signer account's nonce must be incremented by exactly one after a successful transaction | ⚠️ Registry stub — always returns `None` from `InvariantCtx`; use `assert_nonce_increment_correctness()` directly (see note below) |

> **Note on stub invariants:** `ReplayRejection` and `NonceIncrementCorrectness` cannot be
> fully exercised through `InvariantCtx` alone.  Each requires information that is consumed
> before `InvariantCtx` is built:
>
> - **`ReplayRejection`**: `execute_check_on_state` returns the `NSSATransaction` on `Ok`,
>   consuming `self`.  Replaying it requires re-applying the returned transaction to the
>   post-execution state — not possible via a shared `&InvariantCtx`.  Use the standalone
>   `assert_replay_rejection(applied_tx, state, next_block_id, next_timestamp)` helper
>   immediately after each successful execution.  The proptest suite `replay_rejection_proptest`
>   in `fuzz_props/src/invariants.rs` provides reproducible structured coverage of this
>   invariant.
>
> - **`NonceIncrementCorrectness`**: `apply_state_diff` consumes the `ValidatedStateDiff`
>   whose signer-account list is private to the `nssa` crate.  The caller must derive signer
>   IDs from the transaction's witness set before consuming the diff, then call the standalone
>   `assert_nonce_increment_correctness(signer_ids, nonces_before, state_after)` helper.
>   The `signer_account_ids()` helper in `fuzz_props::generators` extracts signer `AccountId`s
>   from an `NSSATransaction`'s witness set.

Additional invariants enforced **inline** in specific targets (not via `ProtocolInvariant`):

| Invariant | Targets |
|-----------|---------|
| `HashRoundTrip` / `HashPreimage` / `TxOrderCommitment` | `fuzz_block_verification` |
| Diff forward containment / reverse completeness | `fuzz_state_diff_computation` |

To add a new invariant:

1. Add a zero-size struct implementing `ProtocolInvariant`.
2. Register it in the `invariants` slice inside `assert_invariants()`.
3. Write a `#[test]` in `fuzz_props` that triggers and detects a synthetic violation.

---

## Input Generators

The `fuzz_props` crate provides two layers of input generation:

### `fuzz_props::arbitrary_types` (libFuzzer / `Arbitrary`)

Typed wrappers that implement `Arbitrary` for LEZ structs.  Use them directly as
fuzz target parameters for zero-boilerplate structured fuzzing.

| Wrapper | Wraps |
|---------|-------|
| `ArbAccountId` | `AccountId` (any 32-byte array) |
| `ArbNonce` | `Nonce` (any `u128`) |
| `ArbPrivateKey` | `PrivateKey` (valid scalar; known-good fallback for the negligible invalid range) |
| `ArbPublicKey` | `PublicKey` (50 % derived from a valid private key; 50 % raw bytes with fallback) |
| `ArbSignature` | `Signature` (random 64-byte value; may be cryptographically invalid) |
| `ArbPubTxMessage` | `Message` for `PublicTransaction` (0–7 accounts, arbitrary instruction data) |
| `ArbWitnessSet` | `WitnessSet` (0–3 `(Signature, PublicKey)` pairs; mixes valid and invalid) |
| `ArbPublicTransaction` | `PublicTransaction` (composed from `ArbPubTxMessage` + `ArbWitnessSet`) |
| `ArbProgramDeploymentTransaction` | `ProgramDeploymentTransaction` (arbitrary bytecode) |
| `ArbHashableBlockData` | `HashableBlockData` (0–7 `ArbNSSATransaction` entries, random header fields) |
| `ArbNSSATransaction` | `NSSATransaction` (`Public` or `ProgramDeployment` variant; `PrivacyPreserving` excluded) |

### `fuzz_props::generators` (libFuzzer helpers + proptest strategies)

| Generator | Covers |
|-----------|--------|
| `arbitrary_fuzz_state()` | 1–8 fuzz-driven accounts with arbitrary IDs, balances, and private keys; used by `fuzz_state_transition`, `fuzz_replay_prevention`, `fuzz_validate_execute_consistency`, `fuzz_state_diff_computation` |
| `arb_fuzz_native_transfer()` | Correctly-signed native-transfer `NSSATransaction` referencing accounts from an `arbitrary_fuzz_state()` result; gives the fuzzer a path to successful state transitions |
| `arbitrary_transaction()` | Structured `NSSATransaction` (`Public` or `ProgramDeployment`) from unstructured bytes via `ArbNSSATransaction` |
| `arb_borsh_transaction_bytes()` | Raw Borsh bytes including invalid encodings |
| `signer_account_ids()` | Extracts `AccountId`s of all signers from an `NSSATransaction`'s witness set; used to derive signer IDs before `apply_state_diff` consumes the diff |
| `arb_native_transfer_tx()` | Valid native-transfer `NSSATransaction` between known testnet genesis accounts (proptest strategy) |
| `test_accounts()` | Returns `(AccountId, PrivateKey)` pairs from `testnet_initial_state` |
| `arb_hashable_block_data()` | `HashableBlockData` with 0–8 valid native transfers (proptest strategy) |
| `arb_invalid_account_state_tx()` | Phantom accounts + overflow amounts — expected to be rejected (IS-3) |
| `arb_duplicate_tx_sequence()` | Duplicated + re-ordered transaction sequences (IS-4) |
| `arb_pathological_sequence()` | Zero-value, self-transfer, max-nonce inputs (IS-5) |

---

## Performance Baseline

Measured on a 4-core x86_64 Linux runner with `RISC0_DEV_MODE=1`:

| Target | Throughput |
|--------|-----------|
| `fuzz_transaction_decoding` | ~200 000 exec/sec |
| `fuzz_stateless_verification` | ~30 000 exec/sec |
| `fuzz_state_transition` | ~5 000 exec/sec |
| `fuzz_block_verification` | ~50 000 exec/sec |
| `fuzz_encoding_roundtrip` | ~150 000 exec/sec |
| `fuzz_signature_verification` | ~20 000 exec/sec |
| `fuzz_replay_prevention` | ~5 000 exec/sec |
| `fuzz_state_diff_computation` | ~10 000 exec/sec |
| `fuzz_validate_execute_consistency` | ~3 000 exec/sec |
| `fuzz_state_serialization` | ~100 000 exec/sec *(estimate)* |
| `fuzz_witness_set_verification` | ~15 000 exec/sec *(estimate)* |
| `fuzz_program_deployment_lifecycle` | ~4 000 exec/sec *(estimate)* |
| `fuzz_apply_state_diff_split_path` | ~5 000 exec/sec *(estimate)* |
| `fuzz_multi_block_state_sequence` | ~1 000 exec/sec *(estimate)* |
| `fuzz_sequencer_vs_replayer` | ~2 000 exec/sec *(estimate)* |

> Throughput figures for the five new targets are rough estimates; run `just perf-baseline`
> locally or check the `perf-baseline` CI artifact for up-to-date measurements.

Recommended local settings for longer runs:

```bash
# libFuzzer — use all available cores
RISC0_DEV_MODE=1 cargo fuzz run fuzz_state_transition \
  -- -max_total_time=3600 -jobs=$(nproc) -workers=$(nproc)

# AFL++ — parallel (1 main + N-1 secondary)
just fuzz-afl-parallel fuzz_state_transition $(nproc) 3600
```

---

## ZK-Proof Cost Warning

`PrivacyPreservingTransaction` uses `risc0-zkvm` (seconds per proof).
All fuzz targets **must** set `RISC0_DEV_MODE=1` in the environment and the `just`
recipes handle this automatically via:

```just
export RISC0_DEV_MODE := "1"
```

Do **not** invoke full proof generation inside any fuzz target. The `RISC0_DEV_MODE=1`
flag stubs out ZK proof generation and replaces it with a fast mock implementation.

---

## Known Limitations & Future Work

| Item | Notes |
|------|-------|
| `PrivacyPreservingTransaction` coverage | Excluded from `fuzz_encoding_roundtrip` because its ZK receipt cannot be reconstructed in a fuzzing loop. A dedicated slow target with `RISC0_DEV_MODE=1` and `proptest` should be added after the current targets are stable |
| `fuzz_validate_execute_consistency` new-account detection | If `execute_check_on_state` creates a brand-new account absent from both the genesis set and the diff, that state-widening will not be detected — full detection requires iterating all accounts in `V03State`, which the API does not currently expose |
| Differential testing (sequencer vs replayer) | ✅ Implemented — `fuzz_sequencer_vs_replayer` feeds the same block through the sequencer path (`validate_on_state` → `apply_state_diff`) and the replayer path (`execute_check_on_state`) and asserts identical state for all known accounts |
| AFL++ integration | ✅ Implemented — `just afl-build`, `just fuzz-afl`, `just fuzz-afl-parallel`; nightly CI in `.github/workflows/fuzz-afl.yml`; single `fuzz/Cargo.toml` covers both engines via feature flags |
| LEZ version tracking | There is no submodule pin — `lez-fuzzing` reads `../logos-execution-zone` as checked out. Update that repo to a release tag or a tested commit, then run `just update-lez` (which does `git pull --ff-only`) and open a PR to bump it |