chore: repoint URLs at git.swaits.com and bump for release

Rewrites remaining github.com / git.sr.ht references to git.swaits.com,
translating GitHub URL shapes to Gitea's (/blob/<ref>/ -> /src/branch/<ref>/,
/tree/<tag> -> /src/tag/<tag>, /actions/workflows/X -> /actions?workflow=X,
/edit/ -> /_edit/).

Left pointing at GitHub on purpose: pull-request links (PRs were not migrated)
and GitHub-only surfaces Gitea lacks (security advisories, discussions).

.cargo/mutants.toml

@@ -0,0 +1,18 @@
+# Use nextest for faster test execution
+test_tool = "nextest"
+
+# Skip slow proptests during mutation testing — the 60+ unit/interop tests
+# are sufficient to catch mutants. Full proptests run separately via just test-unit.
+additional_cargo_test_args = ["-E", "not test(proptests)"]
+
+# False positives excluded:
+#
+# - `> 0` → `>= 0` or `< 0` on usize: tautological (usize always >= 0)
+# - `< 254` → `<= 254` in max_run: both branches yield 254 at the boundary
+#
+# The legacy-msrv const_copy is excluded via #[mutants::skip] in source.
+exclude_re = [
+    "replace > with < in encode",
+    "replace > with >= in encode",
+    "replace < with <= in encode",
+]

.github/workflows/ci.yml

@@ -0,0 +1,69 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  test:
+    name: Test (${{ matrix.rust }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        rust: [stable, "1.93"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: ${{ matrix.rust }}
+      - run: cargo test
+
+  test-legacy-msrv:
+    name: Test (1.83, legacy-msrv)
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@master
+        with:
+          toolchain: "1.83"
+      - run: cargo test --features legacy-msrv
+
+  clippy:
+    name: Clippy
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: clippy
+      - run: cargo clippy -- -D warnings
+
+  fmt:
+    name: Format
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: rustfmt
+      - run: cargo fmt -- --check
+
+  doc:
+    name: Doc
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+      - run: cargo doc --no-deps
+
+  deny:
+    name: Cargo Deny
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: EmbarkStudios/cargo-deny-action@v2

.gitignore

@@ -2,6 +2,8 @@
 /fuzz/target
 /fuzz/corpus
 /fuzz/artifacts
+/mutants.out
+/mutants.out.old
 Cargo.lock
 __pycache__/
 *.pyc

CHANGELOG.md

@@ -0,0 +1,122 @@
+# Changelog
+
+## 0.3.4 — 2026-04-01
+
+### Changed
+
+- **Encode shrunk by 41 bytes** — simplified branch structure in encoder exit
+  path. 388 B encode, 392 B decode, 780 B total (was 821 B). Zero benchmark
+  regressions.
+- **Clarified decode invariant** — added inline comment explaining the subtle
+  trailing-0x01 COBS rule (end-of-data marker vs zero byte).
+- **Clippy cleanup** — `saturating_sub` in proptest (cosmetic, test-only).
+
+### Added
+
+- **GitHub Actions CI** — tests on stable + MSRV 1.93 + 1.83/legacy-msrv,
+  clippy, fmt, doc, and cargo-deny.
+- **`cargo-deny` integration** — `deny.toml` with permissive-only license
+  policy, advisory checks, source auditing. Added `just test-deny` gate.
+- **Encode-only fuzz target** — `fuzz_encode` exercises buffer-too-small
+  paths and validates output properties independently of decode.
+- **Feature coverage** — `just test-unit` and `just test-clippy` now test
+  both default and `legacy-msrv` features. `just test-doc` builds with
+  `--all-features`.
+
+### Fixed
+
+- **README accuracy** — overhead description changed from "exactly" to
+  "at most ... plus 1", added miri to test mentions, updated code size
+  numbers.
+- **CLAUDE.md rewrite** — updated stale API signatures, LOC count, roadmap,
+  dev-deps list, and development commands.
+
+## 0.3.3 — 2026-04-01
+
+### Fixed
+
+- **Fix compilation for downstream users** — `#[mutants::skip]` attribute was
+  not gated behind `cfg(test)`, causing unresolved module error when ucobs is
+  used as a dependency.
+
+## 0.3.2 — 2026-04-01 (yanked)
+
+### Testing
+
+- **Mutation testing: 100% kill rate** — 89 mutants tested, 59 caught, 30
+  unviable (type system rejects them), 0 missed. Uses `cargo-mutants` v27
+  with nextest. Added `mutants` crate for `#[mutants::skip]` annotations.
+- **111 tests** (was 106): added block-boundary interop verification against
+  corncobs, exact-fit and too-small decode buffer sizing for the batch zero
+  path.
+
+## 0.3.1 — 2026-03-31
+
+### Performance
+
+- **Decode: rewritten with shrinking-slice pattern** — decode zeros is 14%
+  faster (25,018 vs 29,114 instructions at 4 KB), decode nonzero flipped from
+  5% behind corncobs to 2% ahead, decode mixed closed from 8% behind to tied.
+  ucobs now wins or ties 16 of 18 benchmarks.
+- **Encode zeros: 6% faster** — zero fast path moved before `split_at` calls,
+  eliminating per-byte overhead. Now tied with cobs (previously 6% behind).
+- **Code size reduced** — 821 B total (was 869 B). Encode 429 B, decode 392 B.
+
+### Changed
+
+- **Replaced criterion with iai-callgrind** — benchmarks now measure
+  deterministic instruction counts via Valgrind instead of wall-clock time.
+  Zero run-to-run variance. Requires `valgrind` installed.
+- **Expanded benchmark coverage** — 9 payload sizes (0–4096 B) × 3 patterns
+  × 3 crates × encode/decode = 162 benchmarks, all running in ~2 minutes.
+- **Code size measurement** added to `just bench` via `just bench-size`.
+- **Mutation testing** added via `cargo-mutants` (`just test-mutants`).
+- Added `mise.toml` for dev tool management.
+
+## 0.3.0 — 2026-03-31
+
+### Performance
+
+- **Encode nonzero/mixed: ~2x faster** — uses const `copy_from_slice` (memcpy at
+  runtime) via Rust 1.93's stabilization of const slice methods. Sub-slice scan
+  pattern helps LLVM eliminate bounds checks.
+- **Encode zeros: ~50% faster** — fast path skips scan loop overhead for leading
+  zero bytes.
+- **Decode zeros: ~4x faster** — batch consecutive `0x01` codes and fill zeros
+  via `fill(0)` (memset) instead of per-byte iteration.
+- Encode now wins or ties corncobs on nonzero/mixed workloads.
+- Decode zeros now leads all competitors by 3x.
+
+### Added
+
+- `legacy-msrv` feature flag — enables compilation on Rust 1.83+ (e.g. Xtensa
+  toolchain) by replacing const `copy_from_slice` with a manual byte loop.
+  All other optimizations (sub-slice scan, zero fast path, decode batch fill)
+  work without this feature on any supported Rust version.
+
+### Changed
+
+- **MSRV bumped from 1.83 to 1.93** — required for const `copy_from_slice`.
+  Use the `legacy-msrv` feature for older toolchains.
+- Dev-dependencies updated: `cobs` 0.3 to 0.5, `criterion` 0.5 to 0.8,
+  `rand` 0.8 to 0.10.
+
+## 0.2.0 — 2026-03-30
+
+### Added
+
+- `const fn` encode — encode data at compile time with zero runtime cost.
+- Comprehensive documentation with 7 usage examples (basic, buffer sizing,
+  framing, stream parsing, compile-time encoding, error handling).
+- Initial encode/decode throughput optimizations.
+
+### Changed
+
+- MSRV set to 1.83.
+
+## 0.1.0 — 2026-03-29
+
+- Initial release.
+- `no_std`, zero-alloc, `unsafe`-free COBS encoder and decoder.
+- 106 tests: canonical vectors, property-based (proptest), fuzz targets,
+  cross-crate interop with `corncobs` and `cobs`.

CLAUDE.md

@@ -10,9 +10,9 @@
 
 1. **Correct** — 100% COBS spec-compliant (Cheshire & Baker 1999). No shortcuts, no deviations.
 2. **Interoperable** — proven byte-for-byte compatibility with `corncobs`, `cobs` crates and Python `cobs` packages.
-3. **Insanely tested** — canonical vectors, property-based tests (proptest), fuzz targets (cargo-fuzz), cross-crate interop, randomized payloads. If you think testing is done, you're 10% done.
+3. **Insanely tested** — canonical vectors, property-based tests (proptest), fuzz targets (cargo-fuzz), cross-crate interop, miri, mutation testing, randomized payloads. If you think testing is done, you're 10% done.
-4. **Fast** — benchmark against every competing implementation. Must be at least as fast, ideally fastest. Benchmarks are CI-enforced.
+4. **Fast** — benchmark against every competing implementation. Must be at least as fast, ideally fastest. Benchmarks run via `just bench`.
-5. **Tiny** — `no_std`, zero-alloc, zero runtime dependencies. ~120 lines of implementation. Runs on 8-bit MCUs and servers alike.
+5. **Tiny** — `no_std`, zero-alloc, zero runtime dependencies. ~140 lines of implementation. Runs on 8-bit MCUs and servers alike.
 6. **Trustworthy** — small enough to audit by hand. No unsafe. No panics in release. Fuzz-hardened.
 
 ## Architecture
@@ -22,7 +22,7 @@ Single-file library: `src/lib.rs`
 ### Public API (3 functions)
 
 ```rust
-pub fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize>
+pub const fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize>
 pub fn decode(src: &[u8], dest: &mut [u8]) -> Option<usize>
 pub const fn max_encoded_len(src_len: usize) -> usize
 ```
@@ -32,26 +32,29 @@ pub const fn max_encoded_len(src_len: usize) -> usize
 ## Development
 
 ```sh
-# Run all tests
+# Run all quality gates (tests, clippy, fmt, doc, fuzz, miri, mutants)
-cargo test
+just test
+
+# Unit + property + interop tests only
+just test-unit
+
+# Benchmarks (iai-callgrind, requires valgrind)
+just bench
 
 # Fuzz (requires nightly + cargo-fuzz)
-cd fuzz
+just test-fuzz
-cargo +nightly fuzz run fuzz_decode -- -max_total_time=60
-cargo +nightly fuzz run fuzz_roundtrip -- -max_total_time=60
 ```
 
 ## Dev Dependencies (test-only)
 
+- `cobs` — cross-validation against the `cobs` 0.5 crate
 - `corncobs` — cross-validation against another COBS implementation
+- `iai-callgrind` — deterministic instruction-count benchmarks
+- `mutants` — mutation testing annotations
 - `proptest` — property-based testing
 - `rand` — randomized test payloads
 
 ## Roadmap
 
-- More interop targets: `cobs` crate, Python `cobs` package
-- Benchmarks (criterion) vs corncobs, cobs, and others
 - quickcheck in addition to proptest
 - Even more fuzz targets and edge-case generators
-- MSRV policy and CI
-- `#[deny(unsafe_code)]` badge

Cargo.toml

@@ -1,13 +1,13 @@
 [package]
 name = "ucobs"
-version = "0.1.0"
+version = "0.3.5"
 edition = "2021"
-rust-version = "1.56"
+rust-version = "1.93"
 license = "MIT"
 description = "μCOBS — Consistent Overhead Byte Stuffing. no_std, zero-alloc, fastest, most-tested COBS implementation."
 authors = ["Stephen Waits <steve@waits.net>"]
-repository = "https://github.com/swaits/ucobs"
+repository = "https://git.swaits.com/swaits/ucobs"
-homepage = "https://github.com/swaits/ucobs"
+homepage = "https://git.swaits.com/swaits/ucobs"
 documentation = "https://docs.rs/ucobs"
 readme = "README.md"
 keywords = ["cobs", "framing", "no-std", "embedded", "codec"]
@@ -15,6 +15,7 @@ categories = ["encoding", "no-std::no-alloc", "embedded"]
 
 [features]
 default = []
+legacy-msrv = []
 
 [lints.rust]
 unsafe_code = "forbid"
@@ -23,11 +24,12 @@ unsafe_code = "forbid"
 all-features = true
 
 [dev-dependencies]
-cobs = "0.3"
+cobs = "0.5"
 corncobs = "0.1"
-criterion = { version = "0.5", features = ["html_reports"] }
+iai-callgrind = "0.16"
+mutants = "0.0.4"
 proptest = "1"
-rand = "0.8"
+rand = "0.10"
 
 [[bench]]
 name = "bench"

README.md

@@ -8,8 +8,8 @@
 
 [Consistent Overhead Byte Stuffing](https://en.wikipedia.org/wiki/Consistent_Overhead_Byte_Stuffing)
 (COBS) encodes arbitrary bytes so that `0x00` never appears in the output,
-allowing `0x00` to serve as an unambiguous frame delimiter. Overhead is exactly
+allowing `0x00` to serve as an unambiguous frame delimiter. Overhead is at most
-1 byte per 254 input bytes (worst case).
+1 byte per 254 input bytes, plus 1.
 
 ## Features
 
@@ -18,111 +18,299 @@ allowing `0x00` to serve as an unambiguous frame delimiter. Overhead is exactly
 - **100% spec-compliant** — Cheshire & Baker, IEEE/ACM Transactions on Networking, 1999
 - **Proven interoperability** — cross-validated against `corncobs`, `cobs`, and Python `cobs`
 - **Insanely tested** — canonical vectors, property-based tests, fuzz targets,
-  cross-crate interop, randomized payloads, and more coming
+  cross-crate interop, miri, mutation testing, randomized payloads
 
 ## Performance
 
 Benchmarked against the two other `no_std` COBS crates using
-[Criterion](https://github.com/bheisler/criterion.rs) across three payload
+[iai-callgrind](https://github.com/iai-callgrind/iai-callgrind) (instruction
-patterns: **all zeros** (worst case — every byte triggers a code emit),
+counts via Valgrind — deterministic, zero run-to-run variance). Three payload
-**no zeros**, and **mixed** (0x00–0xFF cycling). Throughput in MB/s, higher
+patterns: **all zeros** (worst case), **no zeros**, and **mixed** (0x00–0xFF
-is better.
+cycling). Instruction count per call, lower is better.
+Legend: 🥇 winner, 🥉 last, 🤝 tie (<3%).
 
 ### Encode
 
-| Payload | Pattern | μCOBS | `cobs` 0.3 | `corncobs` 0.1 |
+| Payload | Pattern | μCOBS | `cobs` 0.5 | `corncobs` 0.1 |
 |---------|---------|------:|-----------:|---------------:|
-| 64 B    | zeros   |   595 |        748 |            180 |
+| 64 B    | zeros   | 🤝 1,171 | 🤝 1,164 |    🥉 4,098   |
-| 64 B    | mixed   | 1,853 |      1,011 |          1,734 |
+| 64 B    | nonzero | 🤝 512 | 🥉 1,100  |       🤝 514   |
-| 64 B    | nonzero | 1,884 |      1,012 |          2,110 |
+| 64 B    | mixed   | 🥇 523 |  🥉 1,101 |          570   |
-| 256 B   | zeros   |   654 |        817 |            184 |
+| 256 B   | zeros   | 🤝 4,243 | 🤝 4,236 |   🥉 15,810   |
-| 256 B   | mixed   | 1,888 |      1,002 |          1,887 |
+| 256 B   | nonzero | 🤝 1,552 | 🥉 3,992 |   🤝 1,531    |
-| 256 B   | nonzero | 1,964 |      1,004 |          1,968 |
+| 256 B   | mixed   | 🤝 1,561 | 🥉 3,993 |   🤝 1,585    |
-| 4096 B  | zeros   |   675 |        845 |            188 |
+| 4096 B  | zeros   | 🤝 65,904 | 🤝 65,897 | 🥉 250,271  |
-| 4096 B  | mixed   | 1,734 |      1,016 |          1,746 |
+| 4096 B  | nonzero | 🤝 22,077 | 🥉 61,993 | 🤝 21,711   |
-| 4096 B  | nonzero | 1,775 |      1,020 |          2,245 |
+| 4096 B  | mixed   | 🤝 22,932 | 🥉 62,009 | 🤝 22,656   |
 
 ### Decode
 
-| Payload | Pattern | μCOBS | `cobs` 0.3 | `corncobs` 0.1 |
+| Payload | Pattern | μCOBS | `cobs` 0.5 | `corncobs` 0.1 |
 |---------|---------|------:|-----------:|---------------:|
-| 64 B    | zeros   |   595 |        440 |            732 |
+| 64 B    | zeros   | 🥇 499 | 🥉 2,354  |        1,507   |
-| 64 B    | mixed   | 7,041 |        504 |          7,991 |
+| 64 B    | nonzero | 🥇 179 | 🥉 1,778  |          185   |
-| 64 B    | nonzero | 9,900 |        502 |          9,807 |
+| 64 B    | mixed   |    226 | 🥉 1,787  |       🥇 206   |
-| 256 B   | zeros   |   656 |        472 |            742 |
+| 256 B   | zeros   | 🥇 1,473 | 🥉 8,882 |       5,539   |
-| 256 B   | mixed   |17,024 |        645 |         16,207 |
+| 256 B   | nonzero | 🤝 240 | 🥉 6,607  |       🤝 247   |
-| 256 B   | nonzero |18,456 |        647 |         19,549 |
+| 256 B   | mixed   |    285 | 🥉 6,613  |       🥇 266   |
-| 4096 B  | zeros   |   675 |        483 |            768 |
+| 4096 B  | zeros   | 🥇 25,018 | 🥉 139,698 |   86,435   |
-| 4096 B  | mixed   |17,522 |        656 |         22,959 |
+| 4096 B  | nonzero | 🤝 1,315 | 🥉 103,253 |  🤝 1,337   |
-| 4096 B  | nonzero |37,987 |        623 |         40,005 |
+| 4096 B  | mixed   | 🤝 2,086 | 🥉 103,394 |  🤝 2,067   |
 
 ### Code size
 
 Measured from `.text` section of release-optimized symbols (`encode` + `decode`
-combined):
+combined). Run `just bench-size` to reproduce.
 
-| Crate          | Code size |
+| Crate          | `encode` | `decode` | Total |
-|----------------|----------:|
+|----------------|--------:|---------:|------:|
-| μCOBS          |   541 B   |
+| μCOBS          |   388 B |    392 B | 780 B |
-| `cobs` 0.3     |   537 B   |
+| `cobs` 0.5     |   282 B |    494 B | 776 B |
-| `corncobs` 0.1 |   637 B   |
+| `corncobs` 0.1 |   375 B |    262 B | 637 B |
 
-All three crates are tiny. The difference is negligible for any target.
+All three crates are under 1 KB. μCOBS and `cobs` are nearly tied;
+`corncobs` is the smallest. μCOBS's safe `const fn` encoding with
+`copy_from_slice`/memcpy requires `split_at` bounds checks that
+generate panic paths — the cost of combining compile-time safety with
+runtime speed. Embedded users targeting size over speed can build with
+`opt-level = "s"` to shrink further.
 
 ### Crate properties
 
-| Property       | μCOBS  | `cobs` 0.3       | `corncobs` 0.1 |
+| Property       | μCOBS  | `cobs` 0.5       | `corncobs` 0.1 |
 |----------------|:------:|:-----------------:|:--------------:|
 | `no_std`       | always | opt-in[^1]        | always         |
 | Zero-alloc     | yes    | opt-in[^2]        | yes            |
 | `unsafe`-free  | yes    | yes               | yes            |
-| Implementation | ~120 LOC | ~790 LOC        | ~645 LOC       |
+| `const fn` encode | yes | no                | no             |
+| Implementation | ~140 LOC | ~790 LOC        | ~645 LOC       |
 
 [^1]: Requires disabling the default `std` feature.
 [^2]: `alloc` is enabled by default via the `std` feature.
 
-> Measured on Intel N100, Rust 1.93.1, Linux 6.12. Run `just bench` to
+> Benchmarked with iai-callgrind (Valgrind instruction counting) for
-> reproduce. Full Criterion reports are generated in `target/criterion/report/`.
+> deterministic, reproducible results. Run `just bench` to reproduce.
+> Requires `valgrind` installed.
 
-## Usage
+## When to use μCOBS
+
+| You need… | μCOBS | `cobs` 0.5 | `corncobs` 0.1 |
+|---|:---:|:---:|:---:|
+| Minimal code to audit | **~140 LOC** | ~790 LOC | ~645 LOC |
+| Compile-time (`const fn`) encode | **yes** | no | no |
+| Fewest encode instructions | **ties cobs/corncobs** | zeros only | nonzero only |
+| Fewest decode instructions (zero-heavy) | **yes (3–5×)** | no | 2nd |
+| Fewest decode instructions (nonzero) | **yes/tied** | no | tied |
+| Fewest decode instructions (mixed) | tied at 4 KB | no | **slight edge at small sizes** |
+| Dead-simple API (3 functions) | **yes** | yes | more surface area |
+| In-place / streaming encode | no | no | **yes** |
+| `no_std` + zero-alloc by default | **yes** | opt-in | **yes** |
+| Thorough test suite | **111 tests, fuzz, proptest, miri, mutants** | basic | basic |
+
+**Choose μCOBS** if you want the smallest, most auditable COBS implementation
+with a `const fn` encoder, a dead-simple 3-function API, and leading or tied
+instruction efficiency across most workloads — plus dominant decode performance
+on zero-heavy data (3–5× fewer instructions than alternatives). Wins or ties
+16 of 18 benchmarks.
+
+**Choose `corncobs`** if you need in-place encoding, an iterator-based API,
+or your workload is dominated by small mixed-pattern decoding where it uses
+~8% fewer instructions at 64–256 B.
+
+**Choose `cobs`** if you need `std` convenience features or a streaming state
+machine. Note: `cobs` uses 50–75× more instructions than μCOBS/corncobs for
+decoding non-zero data.
+
+## Examples
+
+### Basic encode and decode
 
 ```rust
-// Encode
+// Encode: [0x11, 0x00, 0x33] → [0x02, 0x11, 0x02, 0x33]
 let mut buf = [0u8; 16];
 let n = ucobs::encode(&[0x11, 0x00, 0x33], &mut buf).unwrap();
 assert_eq!(&buf[..n], &[0x02, 0x11, 0x02, 0x33]);
 
-// Decode
+// Decode reverses it
 let mut out = [0u8; 16];
-let n = ucobs::decode(&[0x02, 0x11, 0x02, 0x33], &mut out).unwrap();
+let m = ucobs::decode(&buf[..n], &mut out).unwrap();
+assert_eq!(&out[..m], &[0x11, 0x00, 0x33]);
+```
+
+### Buffer sizing
+
+Use `max_encoded_len` to determine the required destination buffer size:
+
+```rust
+let data = [0x01, 0x02, 0x03];
+let max = ucobs::max_encoded_len(data.len()); // 4 bytes
+
+let mut buf = [0u8; 4];
+let n = ucobs::encode(&data, &mut buf).unwrap();
+assert_eq!(n, 4); // fits exactly
+```
+
+If the destination is too small, `encode` returns `None` rather than panicking:
+
+```rust
+let mut tiny = [0u8; 1];
+assert_eq!(ucobs::encode(&[0x01, 0x02], &mut tiny), None);
+```
+
+### Framing for transport
+
+Append a `0x00` sentinel after encoding to delimit frames on a wire. Strip
+it before decoding:
+
+```rust
+let data = [0x11, 0x00, 0x33];
+
+// Encode and append sentinel
+let mut frame = [0u8; 16];
+let n = ucobs::encode(&data, &mut frame).unwrap();
+frame[n] = 0x00;
+let wire = &frame[..n + 1];
+
+// Receive: strip sentinel, then decode
+let cobs_data = &wire[..wire.len() - 1];
+let mut out = [0u8; 16];
+let m = ucobs::decode(cobs_data, &mut out).unwrap();
+assert_eq!(&out[..m], &data);
+```
+
+### Parsing a stream of frames
+
+Split a byte stream on `0x00` to extract individual COBS frames:
+
+```rust
+// Two frames separated by 0x00 sentinels
+let stream = [
+    0x02, 0x11, 0x02, 0x33, 0x00,  // frame 1: [0x11, 0x00, 0x33]
+    0x03, 0xAA, 0xBB, 0x00,        // frame 2: [0xAA, 0xBB]
+];
+
+let mut out = [0u8; 16];
+let frames: Vec<&[u8]> = stream.split(|&b| b == 0x00)
+    .filter(|f| !f.is_empty())
+    .collect();
+
+let n = ucobs::decode(frames[0], &mut out).unwrap();
 assert_eq!(&out[..n], &[0x11, 0x00, 0x33]);
 
-// Buffer sizing
+let n = ucobs::decode(frames[1], &mut out).unwrap();
-let max = ucobs::max_encoded_len(256); // 258
+assert_eq!(&out[..n], &[0xAA, 0xBB]);
+```
+
+### Compile-time encoding
+
+`encode` is `const fn`, so you can build COBS-encoded tables at compile
+time with zero runtime cost:
+
+```rust
+const PING: [u8; 2] = {
+    let mut buf = [0u8; 2];
+    match ucobs::encode(&[0x01], &mut buf) {
+        Some(_) => buf,
+        None => panic!("buffer too small"),
+    }
+};
+
+const ACK: [u8; 3] = {
+    let mut buf = [0u8; 3];
+    match ucobs::encode(&[0x06, 0x00], &mut buf) {
+        Some(_) => buf,
+        None => panic!("buffer too small"),
+    }
+};
+
+// Baked into the binary — no runtime encoding needed
+assert_eq!(PING, [0x02, 0x01]);
+assert_eq!(ACK, [0x02, 0x06, 0x01]);
+```
+
+### Error handling
+
+Both `encode` and `decode` return `Option<usize>` — `None` on failure,
+never panic:
+
+```rust
+let mut buf = [0u8; 8];
+
+// Destination too small
+assert_eq!(ucobs::encode(&[1, 2, 3], &mut buf[..1]), None);
+
+// Malformed input (zero byte in encoded stream)
+assert_eq!(ucobs::decode(&[0x00], &mut buf), None);
+
+// Truncated frame (code byte promises more data than exists)
+assert_eq!(ucobs::decode(&[0x05, 0x11], &mut buf), None);
+
+// Empty input is valid
+assert_eq!(ucobs::encode(&[], &mut buf), Some(1)); // encodes to [0x01]
+assert_eq!(ucobs::decode(&[], &mut buf), Some(0)); // decodes to []
 ```
 
 ## API
 
 | Function | Signature | Description |
 |---|---|---|
-| `encode` | `(src: &[u8], dest: &mut [u8]) -> Option<usize>` | COBS-encode `src` into `dest`. Returns byte count or `None` if buffer too small. |
+| `encode` | `const fn(src: &[u8], dest: &mut [u8]) -> Option<usize>` | COBS-encode `src` into `dest`. Returns byte count or `None` if buffer too small. |
 | `decode` | `(src: &[u8], dest: &mut [u8]) -> Option<usize>` | Decode COBS data. Returns byte count or `None` if input is malformed. |
 | `max_encoded_len` | `const fn(src_len: usize) -> usize` | Maximum encoded size for a given input length. |
 
 **Note:** The encoder does *not* append a trailing `0x00` sentinel. Append it
 yourself when framing for transport.
 
+## Minimum Supported Rust Version
+
+The default build requires **Rust 1.93+** (for const `copy_from_slice`).
+
+### `legacy-msrv` feature
+
+If you're targeting a toolchain older than 1.93 (e.g. Xtensa), enable the
+`legacy-msrv` feature to fall back to a manual byte copy in the encoder:
+
+```toml
+[dependencies]
+ucobs = { version = "0.3", features = ["legacy-msrv"] }
+```
+
+All other optimizations (sub-slice scan, zero fast path, decode batch fill)
+are available regardless of this feature. Only the encode copy phase is
+affected — it uses a byte loop instead of `copy_from_slice`/memcpy.
+
 ## Testing
 
+μCOBS has one of the most thorough test suites of any COBS implementation —
+111 tests across 8 categories:
+
+- **Canonical vectors** — every example from Cheshire & Baker 1999 (the original
+  COBS paper), plus the full Wikipedia vector set
+- **External corpora** — vectors drawn from `cobs-c`, `nanocobs`, `cobs2-rs`,
+  Jacques Fortier's C implementation, and the Python `cobs` package
+- **Cross-crate interop** — byte-for-byte validation against both `corncobs` and
+  `cobs` crates on randomized payloads up to 4 KB
+- **Property-based tests** — 6 proptest suites covering round-trip correctness,
+  no-zero-in-output invariants, encoding properties, decoding properties,
+  structural/algebraic properties, and boundary-region behavior
+- **Fuzz targets** — 3 `cargo-fuzz` harnesses (decode, round-trip, small-dest)
+  for continuous coverage of edge cases
+- **Mutation testing** — `cargo-mutants` with 100% kill rate (59/59 viable
+  mutants caught; 30 more rejected by the type system)
+- **Boundary tests** — exhaustive coverage of 254/255-byte block boundaries,
+  multi-block splits, buffer-exact fits, and off-by-one dest sizes
+- **Compile-time verification** — `const` assertions that validate encode
+  correctness at compile time
+
 ```sh
 # Unit + property + interop tests
 just test-unit
 
-# All quality gates (tests, clippy, fmt, doc, fuzz, miri)
+# All quality gates (tests, clippy, fmt, doc, fuzz, miri, mutants)
 just test
 
-# Fuzz testing only (requires nightly + cargo-fuzz)
+# Mutation testing only
-just test-fuzz
+just test-mutants
 ```
 
 ## License

benches/bench.rs

@@ -1,149 +1,271 @@
-use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
+use iai_callgrind::{library_benchmark, library_benchmark_group, main};
+use std::hint::black_box;
 
 // ── Test payloads ───────────────────────────────────────────────────
 
-fn payload_zeros(len: usize) -> Vec<u8> {
+fn zeros(n: usize) -> Vec<u8> {
-    vec![0u8; len]
+    vec![0u8; n]
 }
 
-fn payload_nonzero(len: usize) -> Vec<u8> {
+fn nonzero(n: usize) -> Vec<u8> {
-    (0..len).map(|i| (i % 255 + 1) as u8).collect()
+    (0..n).map(|i| (i % 255 + 1) as u8).collect()
 }
 
-fn payload_mixed(len: usize) -> Vec<u8> {
+fn mixed(n: usize) -> Vec<u8> {
-    (0..len).map(|i| (i % 256) as u8).collect()
+    (0..n).map(|i| (i % 256) as u8).collect()
 }
 
-// ── ucobs ───────────────────────────────────────────────────────────
+// ── Setup helpers ───────────────────────────────────────────────────
 
-fn ucobs_encode(data: &[u8]) -> Vec<u8> {
+fn setup_encode(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-    let mut buf = vec![0u8; ucobs::max_encoded_len(data.len()) + 1];
+    let buf = vec![0u8; ucobs::max_encoded_len(data.len()) + 1];
-    let n = ucobs::encode(data, &mut buf).unwrap();
+    (data, buf)
-    buf.truncate(n);
-    buf
 }
 
-fn ucobs_decode(encoded: &[u8], out: &mut [u8]) -> usize {
+fn setup_decode_ucobs(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-    ucobs::decode(encoded, out).unwrap()
+    let mut enc = vec![0u8; ucobs::max_encoded_len(data.len()) + 1];
+    let n = ucobs::encode(&data, &mut enc).unwrap();
+    enc.truncate(n);
+    let dec = vec![0u8; data.len() + 1];
+    (enc, dec)
 }
 
-// ── corncobs ────────────────────────────────────────────────────────
+fn setup_encode_corncobs(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-
+    let buf = vec![0u8; corncobs::max_encoded_len(data.len())];
-fn corncobs_encode(data: &[u8]) -> Vec<u8> {
+    (data, buf)
-    let mut buf = vec![0u8; corncobs::max_encoded_len(data.len())];
-    let n = corncobs::encode_buf(data, &mut buf);
-    buf.truncate(n);
-    buf
 }
 
-fn corncobs_decode(encoded: &[u8], out: &mut [u8]) -> usize {
+fn setup_decode_corncobs(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-    // corncobs expects trailing sentinel
+    let mut enc = vec![0u8; corncobs::max_encoded_len(data.len())];
-    corncobs::decode_buf(encoded, out).unwrap()
+    let n = corncobs::encode_buf(&data, &mut enc);
+    enc.truncate(n);
+    let dec = vec![0u8; data.len() + 1];
+    (enc, dec)
 }
 
-// ── cobs crate ──────────────────────────────────────────────────────
+fn setup_encode_cobs(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-
+    let buf = vec![0u8; cobs::max_encoding_length(data.len())];
-fn cobs_encode(data: &[u8]) -> Vec<u8> {
+    (data, buf)
-    let mut buf = vec![0u8; cobs::max_encoding_length(data.len())];
-    let n = cobs::encode(data, &mut buf);
-    buf.truncate(n);
-    buf
 }
 
-fn cobs_decode(encoded: &[u8], out: &mut [u8]) -> usize {
+fn setup_decode_cobs(data: Vec<u8>) -> (Vec<u8>, Vec<u8>) {
-    // unwrap_or(0): the cobs crate returns Err(EmptyFrame) for empty input,
+    let mut enc = vec![0u8; cobs::max_encoding_length(data.len())];
-    // which is valid COBS. We treat it as 0 decoded bytes for benchmarking.
+    let n = cobs::encode(&data, &mut enc);
-    cobs::decode(encoded, out).unwrap_or(0)
+    enc.truncate(n);
+    let dec = vec![0u8; data.len() + 1];
+    (enc, dec)
 }
 
-// ── Benchmarks ──────────────────────────────────────────────────────
+// ── Encode benchmarks ──────────────────────────────────────────────
 
-const SIZES: &[usize] = &[0, 1, 10, 64, 254, 255, 256, 1024, 4096];
+#[library_benchmark]
-
+#[bench::zeros_0(setup_encode(zeros(0)))]
-fn bench_encode(c: &mut Criterion) {
+#[bench::zeros_1(setup_encode(zeros(1)))]
-    let payloads: Vec<(&str, Box<dyn Fn(usize) -> Vec<u8>>)> = vec![
+#[bench::zeros_10(setup_encode(zeros(10)))]
-        ("zeros", Box::new(payload_zeros)),
+#[bench::zeros_64(setup_encode(zeros(64)))]
-        ("nonzero", Box::new(payload_nonzero)),
+#[bench::zeros_254(setup_encode(zeros(254)))]
-        ("mixed", Box::new(payload_mixed)),
+#[bench::zeros_255(setup_encode(zeros(255)))]
-    ];
+#[bench::zeros_256(setup_encode(zeros(256)))]
-
+#[bench::zeros_1024(setup_encode(zeros(1024)))]
-    for (name, make) in &payloads {
+#[bench::zeros_4096(setup_encode(zeros(4096)))]
-        let mut group = c.benchmark_group(format!("encode/{name}"));
+#[bench::nonzero_0(setup_encode(nonzero(0)))]
-        for &size in SIZES {
+#[bench::nonzero_1(setup_encode(nonzero(1)))]
-            let data = make(size);
+#[bench::nonzero_10(setup_encode(nonzero(10)))]
-            group.throughput(Throughput::Bytes(size as u64));
+#[bench::nonzero_64(setup_encode(nonzero(64)))]
-
+#[bench::nonzero_254(setup_encode(nonzero(254)))]
-            group.bench_with_input(BenchmarkId::new("ucobs", size), &data, |b, d| {
+#[bench::nonzero_255(setup_encode(nonzero(255)))]
-                let mut buf = vec![0u8; ucobs::max_encoded_len(d.len()) + 1];
+#[bench::nonzero_256(setup_encode(nonzero(256)))]
-                b.iter(|| {
+#[bench::nonzero_1024(setup_encode(nonzero(1024)))]
-                    black_box(ucobs::encode(black_box(d), &mut buf).unwrap());
+#[bench::nonzero_4096(setup_encode(nonzero(4096)))]
-                });
+#[bench::mixed_0(setup_encode(mixed(0)))]
-            });
+#[bench::mixed_1(setup_encode(mixed(1)))]
-
+#[bench::mixed_10(setup_encode(mixed(10)))]
-            group.bench_with_input(BenchmarkId::new("corncobs", size), &data, |b, d| {
+#[bench::mixed_64(setup_encode(mixed(64)))]
-                let mut buf = vec![0u8; corncobs::max_encoded_len(d.len())];
+#[bench::mixed_254(setup_encode(mixed(254)))]
-                b.iter(|| {
+#[bench::mixed_255(setup_encode(mixed(255)))]
-                    black_box(corncobs::encode_buf(black_box(d), &mut buf));
+#[bench::mixed_256(setup_encode(mixed(256)))]
-                });
+#[bench::mixed_1024(setup_encode(mixed(1024)))]
-            });
+#[bench::mixed_4096(setup_encode(mixed(4096)))]
-
+fn encode_ucobs((data, mut buf): (Vec<u8>, Vec<u8>)) {
-            group.bench_with_input(BenchmarkId::new("cobs", size), &data, |b, d| {
+    black_box(ucobs::encode(black_box(&data), &mut buf));
-                let mut buf = vec![0u8; cobs::max_encoding_length(d.len())];
-                b.iter(|| {
-                    black_box(cobs::encode(black_box(d), &mut buf));
-                });
-            });
-        }
-        group.finish();
-    }
 }
 
-fn bench_decode(c: &mut Criterion) {
+#[library_benchmark]
-    let payloads: Vec<(&str, Box<dyn Fn(usize) -> Vec<u8>>)> = vec![
+#[bench::zeros_0(setup_encode_corncobs(zeros(0)))]
-        ("zeros", Box::new(payload_zeros)),
+#[bench::zeros_1(setup_encode_corncobs(zeros(1)))]
-        ("nonzero", Box::new(payload_nonzero)),
+#[bench::zeros_10(setup_encode_corncobs(zeros(10)))]
-        ("mixed", Box::new(payload_mixed)),
+#[bench::zeros_64(setup_encode_corncobs(zeros(64)))]
-    ];
+#[bench::zeros_254(setup_encode_corncobs(zeros(254)))]
-
+#[bench::zeros_255(setup_encode_corncobs(zeros(255)))]
-    for (name, make) in &payloads {
+#[bench::zeros_256(setup_encode_corncobs(zeros(256)))]
-        let mut group = c.benchmark_group(format!("decode/{name}"));
+#[bench::zeros_1024(setup_encode_corncobs(zeros(1024)))]
-        for &size in SIZES {
+#[bench::zeros_4096(setup_encode_corncobs(zeros(4096)))]
-            let data = make(size);
+#[bench::nonzero_0(setup_encode_corncobs(nonzero(0)))]
-            let ucobs_enc = ucobs_encode(&data);
+#[bench::nonzero_1(setup_encode_corncobs(nonzero(1)))]
-            // corncobs expects sentinel in encoded form
+#[bench::nonzero_10(setup_encode_corncobs(nonzero(10)))]
-            let corncobs_enc = corncobs_encode(&data);
+#[bench::nonzero_64(setup_encode_corncobs(nonzero(64)))]
-            let cobs_enc = cobs_encode(&data);
+#[bench::nonzero_254(setup_encode_corncobs(nonzero(254)))]
-
+#[bench::nonzero_255(setup_encode_corncobs(nonzero(255)))]
-            group.throughput(Throughput::Bytes(size as u64));
+#[bench::nonzero_256(setup_encode_corncobs(nonzero(256)))]
-
+#[bench::nonzero_1024(setup_encode_corncobs(nonzero(1024)))]
-            group.bench_with_input(BenchmarkId::new("ucobs", size), &ucobs_enc, |b, enc| {
+#[bench::nonzero_4096(setup_encode_corncobs(nonzero(4096)))]
-                let mut buf = vec![0u8; size + 1];
+#[bench::mixed_0(setup_encode_corncobs(mixed(0)))]
-                b.iter(|| {
+#[bench::mixed_1(setup_encode_corncobs(mixed(1)))]
-                    black_box(ucobs_decode(black_box(enc), &mut buf));
+#[bench::mixed_10(setup_encode_corncobs(mixed(10)))]
-                });
+#[bench::mixed_64(setup_encode_corncobs(mixed(64)))]
-            });
+#[bench::mixed_254(setup_encode_corncobs(mixed(254)))]
-
+#[bench::mixed_255(setup_encode_corncobs(mixed(255)))]
-            group.bench_with_input(
+#[bench::mixed_256(setup_encode_corncobs(mixed(256)))]
-                BenchmarkId::new("corncobs", size),
+#[bench::mixed_1024(setup_encode_corncobs(mixed(1024)))]
-                &corncobs_enc,
+#[bench::mixed_4096(setup_encode_corncobs(mixed(4096)))]
-                |b, enc| {
+fn encode_corncobs((data, mut buf): (Vec<u8>, Vec<u8>)) {
-                    let mut buf = vec![0u8; size + 1];
+    black_box(corncobs::encode_buf(black_box(&data), &mut buf));
-                    b.iter(|| {
-                        black_box(corncobs_decode(black_box(enc), &mut buf));
-                    });
-                },
-            );
-
-            group.bench_with_input(BenchmarkId::new("cobs", size), &cobs_enc, |b, enc| {
-                let mut buf = vec![0u8; size + 1];
-                b.iter(|| {
-                    black_box(cobs_decode(black_box(enc), &mut buf));
-                });
-            });
-        }
-        group.finish();
-    }
 }
 
-criterion_group!(benches, bench_encode, bench_decode);
+#[library_benchmark]
-criterion_main!(benches);
+#[bench::zeros_0(setup_encode_cobs(zeros(0)))]
+#[bench::zeros_1(setup_encode_cobs(zeros(1)))]
+#[bench::zeros_10(setup_encode_cobs(zeros(10)))]
+#[bench::zeros_64(setup_encode_cobs(zeros(64)))]
+#[bench::zeros_254(setup_encode_cobs(zeros(254)))]
+#[bench::zeros_255(setup_encode_cobs(zeros(255)))]
+#[bench::zeros_256(setup_encode_cobs(zeros(256)))]
+#[bench::zeros_1024(setup_encode_cobs(zeros(1024)))]
+#[bench::zeros_4096(setup_encode_cobs(zeros(4096)))]
+#[bench::nonzero_0(setup_encode_cobs(nonzero(0)))]
+#[bench::nonzero_1(setup_encode_cobs(nonzero(1)))]
+#[bench::nonzero_10(setup_encode_cobs(nonzero(10)))]
+#[bench::nonzero_64(setup_encode_cobs(nonzero(64)))]
+#[bench::nonzero_254(setup_encode_cobs(nonzero(254)))]
+#[bench::nonzero_255(setup_encode_cobs(nonzero(255)))]
+#[bench::nonzero_256(setup_encode_cobs(nonzero(256)))]
+#[bench::nonzero_1024(setup_encode_cobs(nonzero(1024)))]
+#[bench::nonzero_4096(setup_encode_cobs(nonzero(4096)))]
+#[bench::mixed_0(setup_encode_cobs(mixed(0)))]
+#[bench::mixed_1(setup_encode_cobs(mixed(1)))]
+#[bench::mixed_10(setup_encode_cobs(mixed(10)))]
+#[bench::mixed_64(setup_encode_cobs(mixed(64)))]
+#[bench::mixed_254(setup_encode_cobs(mixed(254)))]
+#[bench::mixed_255(setup_encode_cobs(mixed(255)))]
+#[bench::mixed_256(setup_encode_cobs(mixed(256)))]
+#[bench::mixed_1024(setup_encode_cobs(mixed(1024)))]
+#[bench::mixed_4096(setup_encode_cobs(mixed(4096)))]
+fn encode_cobs((data, mut buf): (Vec<u8>, Vec<u8>)) {
+    black_box(cobs::encode(black_box(&data), &mut buf));
+}
+
+// ── Decode benchmarks ──────────────────────────────────────────────
+
+#[library_benchmark]
+#[bench::zeros_0(setup_decode_ucobs(zeros(0)))]
+#[bench::zeros_1(setup_decode_ucobs(zeros(1)))]
+#[bench::zeros_10(setup_decode_ucobs(zeros(10)))]
+#[bench::zeros_64(setup_decode_ucobs(zeros(64)))]
+#[bench::zeros_254(setup_decode_ucobs(zeros(254)))]
+#[bench::zeros_255(setup_decode_ucobs(zeros(255)))]
+#[bench::zeros_256(setup_decode_ucobs(zeros(256)))]
+#[bench::zeros_1024(setup_decode_ucobs(zeros(1024)))]
+#[bench::zeros_4096(setup_decode_ucobs(zeros(4096)))]
+#[bench::nonzero_0(setup_decode_ucobs(nonzero(0)))]
+#[bench::nonzero_1(setup_decode_ucobs(nonzero(1)))]
+#[bench::nonzero_10(setup_decode_ucobs(nonzero(10)))]
+#[bench::nonzero_64(setup_decode_ucobs(nonzero(64)))]
+#[bench::nonzero_254(setup_decode_ucobs(nonzero(254)))]
+#[bench::nonzero_255(setup_decode_ucobs(nonzero(255)))]
+#[bench::nonzero_256(setup_decode_ucobs(nonzero(256)))]
+#[bench::nonzero_1024(setup_decode_ucobs(nonzero(1024)))]
+#[bench::nonzero_4096(setup_decode_ucobs(nonzero(4096)))]
+#[bench::mixed_0(setup_decode_ucobs(mixed(0)))]
+#[bench::mixed_1(setup_decode_ucobs(mixed(1)))]
+#[bench::mixed_10(setup_decode_ucobs(mixed(10)))]
+#[bench::mixed_64(setup_decode_ucobs(mixed(64)))]
+#[bench::mixed_254(setup_decode_ucobs(mixed(254)))]
+#[bench::mixed_255(setup_decode_ucobs(mixed(255)))]
+#[bench::mixed_256(setup_decode_ucobs(mixed(256)))]
+#[bench::mixed_1024(setup_decode_ucobs(mixed(1024)))]
+#[bench::mixed_4096(setup_decode_ucobs(mixed(4096)))]
+fn decode_ucobs((enc, mut buf): (Vec<u8>, Vec<u8>)) {
+    black_box(ucobs::decode(black_box(&enc), &mut buf));
+}
+
+#[library_benchmark]
+#[bench::zeros_0(setup_decode_corncobs(zeros(0)))]
+#[bench::zeros_1(setup_decode_corncobs(zeros(1)))]
+#[bench::zeros_10(setup_decode_corncobs(zeros(10)))]
+#[bench::zeros_64(setup_decode_corncobs(zeros(64)))]
+#[bench::zeros_254(setup_decode_corncobs(zeros(254)))]
+#[bench::zeros_255(setup_decode_corncobs(zeros(255)))]
+#[bench::zeros_256(setup_decode_corncobs(zeros(256)))]
+#[bench::zeros_1024(setup_decode_corncobs(zeros(1024)))]
+#[bench::zeros_4096(setup_decode_corncobs(zeros(4096)))]
+#[bench::nonzero_0(setup_decode_corncobs(nonzero(0)))]
+#[bench::nonzero_1(setup_decode_corncobs(nonzero(1)))]
+#[bench::nonzero_10(setup_decode_corncobs(nonzero(10)))]
+#[bench::nonzero_64(setup_decode_corncobs(nonzero(64)))]
+#[bench::nonzero_254(setup_decode_corncobs(nonzero(254)))]
+#[bench::nonzero_255(setup_decode_corncobs(nonzero(255)))]
+#[bench::nonzero_256(setup_decode_corncobs(nonzero(256)))]
+#[bench::nonzero_1024(setup_decode_corncobs(nonzero(1024)))]
+#[bench::nonzero_4096(setup_decode_corncobs(nonzero(4096)))]
+#[bench::mixed_0(setup_decode_corncobs(mixed(0)))]
+#[bench::mixed_1(setup_decode_corncobs(mixed(1)))]
+#[bench::mixed_10(setup_decode_corncobs(mixed(10)))]
+#[bench::mixed_64(setup_decode_corncobs(mixed(64)))]
+#[bench::mixed_254(setup_decode_corncobs(mixed(254)))]
+#[bench::mixed_255(setup_decode_corncobs(mixed(255)))]
+#[bench::mixed_256(setup_decode_corncobs(mixed(256)))]
+#[bench::mixed_1024(setup_decode_corncobs(mixed(1024)))]
+#[bench::mixed_4096(setup_decode_corncobs(mixed(4096)))]
+fn decode_corncobs((enc, mut buf): (Vec<u8>, Vec<u8>)) {
+    let _ = black_box(corncobs::decode_buf(black_box(&enc), &mut buf));
+}
+
+#[library_benchmark]
+#[bench::zeros_0(setup_decode_cobs(zeros(0)))]
+#[bench::zeros_1(setup_decode_cobs(zeros(1)))]
+#[bench::zeros_10(setup_decode_cobs(zeros(10)))]
+#[bench::zeros_64(setup_decode_cobs(zeros(64)))]
+#[bench::zeros_254(setup_decode_cobs(zeros(254)))]
+#[bench::zeros_255(setup_decode_cobs(zeros(255)))]
+#[bench::zeros_256(setup_decode_cobs(zeros(256)))]
+#[bench::zeros_1024(setup_decode_cobs(zeros(1024)))]
+#[bench::zeros_4096(setup_decode_cobs(zeros(4096)))]
+#[bench::nonzero_0(setup_decode_cobs(nonzero(0)))]
+#[bench::nonzero_1(setup_decode_cobs(nonzero(1)))]
+#[bench::nonzero_10(setup_decode_cobs(nonzero(10)))]
+#[bench::nonzero_64(setup_decode_cobs(nonzero(64)))]
+#[bench::nonzero_254(setup_decode_cobs(nonzero(254)))]
+#[bench::nonzero_255(setup_decode_cobs(nonzero(255)))]
+#[bench::nonzero_256(setup_decode_cobs(nonzero(256)))]
+#[bench::nonzero_1024(setup_decode_cobs(nonzero(1024)))]
+#[bench::nonzero_4096(setup_decode_cobs(nonzero(4096)))]
+#[bench::mixed_0(setup_decode_cobs(mixed(0)))]
+#[bench::mixed_1(setup_decode_cobs(mixed(1)))]
+#[bench::mixed_10(setup_decode_cobs(mixed(10)))]
+#[bench::mixed_64(setup_decode_cobs(mixed(64)))]
+#[bench::mixed_254(setup_decode_cobs(mixed(254)))]
+#[bench::mixed_255(setup_decode_cobs(mixed(255)))]
+#[bench::mixed_256(setup_decode_cobs(mixed(256)))]
+#[bench::mixed_1024(setup_decode_cobs(mixed(1024)))]
+#[bench::mixed_4096(setup_decode_cobs(mixed(4096)))]
+fn decode_cobs((enc, mut buf): (Vec<u8>, Vec<u8>)) {
+    black_box(
+        cobs::decode(black_box(&enc), &mut buf)
+            .map(|r| r.parsed_size())
+            .unwrap_or(0),
+    );
+}
+
+// ── Groups ─────────────────────────────────────────────────────────
+
+library_benchmark_group!(
+    name = encode;
+    benchmarks = encode_ucobs, encode_corncobs, encode_cobs
+);
+
+library_benchmark_group!(
+    name = decode;
+    benchmarks = decode_ucobs, decode_corncobs, decode_cobs
+);
+
+main!(library_benchmark_groups = encode, decode);

deny.toml

@@ -0,0 +1,37 @@
+[graph]
+all-features = true
+
+[advisories]
+yanked = "deny"
+ignore = [
+    # bincode 1.3.3 — transitive dev-dep via iai-callgrind (latest 0.16.1).
+    # Unmaintained but stable and complete; no security issue. Dev-only, never
+    # ships in ucobs. Remove this ignore when iai-callgrind drops bincode.
+    "RUSTSEC-2025-0141",
+]
+
+[licenses]
+# Permissive licenses only. All copyleft (GPL, LGPL, AGPL, MPL, etc.) is denied by omission.
+allow = [
+    "MIT",
+    "Apache-2.0",
+    "BSD-2-Clause",
+    "BSD-3-Clause",
+    "ISC",
+    "Zlib",
+    "Unlicense",
+    "CC0-1.0",
+    "Unicode-3.0",
+    "Unicode-DFS-2016",
+]
+unused-allowed-license = "allow"
+
+[bans]
+multiple-versions = "warn"
+wildcards = "deny"
+
+[sources]
+unknown-registry = "deny"
+unknown-git = "deny"
+allow-registry = ["https://github.com/rust-lang/crates.io-index"]
+allow-git = []

fuzz/Cargo.toml

@@ -30,3 +30,9 @@ name = "fuzz_small_dest"
 path = "fuzz_targets/fuzz_small_dest.rs"
 test = false
 doc = false
+
+[[bin]]
+name = "fuzz_encode"
+path = "fuzz_targets/fuzz_encode.rs"
+test = false
+doc = false

fuzz/fuzz_targets/fuzz_encode.rs

@@ -0,0 +1,33 @@
+//! Fuzz target: encode arbitrary input into buffers of varying sizes.
+//! Exercises encode error paths (buffer too small) and validates output properties.
+#![no_main]
+use libfuzzer_sys::fuzz_target;
+
+fuzz_target!(|data: &[u8]| {
+    if data.len() > 1024 {
+        return;
+    }
+
+    let max = ucobs::max_encoded_len(data.len());
+
+    // Encode into a correctly-sized buffer — must always succeed.
+    let mut buf = vec![0u8; max];
+    let n = ucobs::encode(data, &mut buf)
+        .expect("encode failed with max_encoded_len buffer");
+    let encoded = &buf[..n];
+
+    // Output must never contain 0x00.
+    assert!(!encoded.contains(&0x00), "encoded output contains 0x00");
+
+    // Output length must be bounded by max_encoded_len.
+    assert!(n <= max, "encoded length exceeds max_encoded_len");
+
+    // Output must be at least 1 byte (even empty input encodes to [0x01]).
+    assert!(n >= 1, "encoded length is zero");
+
+    // Encode into every possible undersized buffer — must return None, never panic.
+    for size in 0..n {
+        let mut small = vec![0u8; size];
+        assert_eq!(ucobs::encode(data, &mut small), None);
+    }
+});

justfile

@@ -4,8 +4,8 @@
 default:
     @just --list
 
-# run ALL quality gates (tests, clippy, fmt check, doc, fuzz, miri)
+# run ALL quality gates (tests, clippy, fmt check, doc, deny, fuzz, miri, mutants)
-test: test-unit test-clippy test-fmt test-doc test-fuzz test-miri
+test: test-unit test-clippy test-fmt test-doc test-deny test-fuzz test-miri test-mutants
 
 # quick compile check (no tests)
 check:
@@ -14,11 +14,13 @@ check:
 # unit + integration + proptest tests (via nextest for parallelism)
 test-unit:
     cargo nextest run
+    cargo nextest run --features legacy-msrv
     cargo test --doc
 
 # clippy with all warnings as errors
 test-clippy:
     cargo clippy -- -D warnings
+    cargo clippy --features legacy-msrv -- -D warnings
 
 # check formatting
 test-fmt:
@@ -26,11 +28,16 @@ test-fmt:
 
 # doc build (catches broken doc links)
 test-doc:
-    cargo doc --no-deps
+    cargo doc --no-deps --all-features
+
+# license and advisory audit (requires cargo-deny)
+test-deny:
+    cargo deny check
 
 # fuzz all targets (60s each, requires nightly + cargo-fuzz)
 test-fuzz:
     cd fuzz && cargo +nightly fuzz run fuzz_decode -- -max_total_time=60
+    cd fuzz && cargo +nightly fuzz run fuzz_encode -- -max_total_time=60
     cd fuzz && cargo +nightly fuzz run fuzz_roundtrip -- -max_total_time=60
     cd fuzz && cargo +nightly fuzz run fuzz_small_dest -- -max_total_time=60
 
@@ -39,10 +46,26 @@ test-fuzz:
 test-miri:
     cargo +nightly miri test -- --skip proptests
 
-# run criterion benchmarks (ucobs vs cobs vs corncobs)
+# mutation testing (requires cargo-mutants)
-bench:
+test-mutants:
+    cargo mutants --no-shuffle
+
+# run iai-callgrind benchmarks (deterministic instruction counts, requires valgrind)
+bench: bench-iai bench-size
+
+# run iai-callgrind benchmarks (ucobs vs cobs vs corncobs)
+bench-iai:
     cargo bench
 
+# measure compiled code size (encode + decode .text bytes)
+bench-size:
+    @cargo build --release 2>/dev/null
+    @echo ""
+    @echo "Code size (release, encode + decode .text):"
+    @nm -S target/release/libucobs.rlib 2>/dev/null \
+        | grep 'T _ZN5ucobs' \
+        | awk '{size=strtonum("0x"$2); name=$4; sub(/.*ucobs[0-9]*/, "", name); sub(/17h.*/, "", name); printf "  ucobs %-7s %4d B\n", name, size}'
+
 # run python cobs benchmarks (requires: pip install cobs)
 bench-python:
     python3 bench_python.py

mise.toml

@@ -0,0 +1,7 @@
+[tools]
+rust = "1.93"
+"cargo:cargo-deny" = "latest"
+"cargo:cargo-fuzz" = "latest"
+"cargo:cargo-mutants" = "latest"
+"cargo:cargo-nextest" = "latest"
+"cargo:iai-callgrind-runner" = "latest"

proptest-regressions/lib.txt

@@ -0,0 +1,10 @@
+# Seeds for failure cases proptest has generated in the past. It is
+# automatically read and these particular cases re-run before any
+# novel cases are generated.
+#
+# It is recommended to check this file in to source control so that
+# everyone who runs the test benefits from these saved cases.
+cc 00411472de4d6cdb296c04bc52fd533354bc25ba1765807da8b7994868fcb987 # shrinks to len = 1
+cc 81ebf688f1dd5ea2a735756a7fb703b92c2a2033057ae91886577c6a63763cf5 # shrinks to data = [0]
+cc 32d0d2d26e2be6ec66b432f1e9baee378777b1a00b959922cff7a5784292ec4b # shrinks to data = [0]
+cc fe447e0e58f4c0b21f6e1b38a90bd33578c968967c956964a06013a4df977d68 # shrinks to data = [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]

src/lib.rs

@@ -9,54 +9,234 @@
 //! Overhead is at most 1 byte per 254 input bytes plus 1; see
 //! [`max_encoded_len`] for the exact formula.
 //!
-//! This crate is `no_std` and zero-alloc — suitable for embedded firmware.
+//! This crate is `no_std`, zero-alloc, and `unsafe`-free — it runs
+//! everywhere from 8-bit MCUs to servers.
 //!
-//! # Sentinel convention
+//! # Feature flags
+//!
+//! - **`legacy-msrv`** — Enables compilation on Rust 1.83+ by replacing
+//!   the const `copy_from_slice` (requires Rust 1.93) with a manual byte
+//!   loop in the encoder. All other optimizations are unaffected.
+//!
+//! # Quick start
+//!
+//! ```
+//! // Encode
+//! let mut buf = [0u8; 16];
+//! let n = ucobs::encode(&[0x11, 0x00, 0x33], &mut buf).unwrap();
+//! assert_eq!(&buf[..n], &[0x02, 0x11, 0x02, 0x33]);
+//!
+//! // Decode
+//! let mut out = [0u8; 16];
+//! let m = ucobs::decode(&buf[..n], &mut out).unwrap();
+//! assert_eq!(&out[..m], &[0x11, 0x00, 0x33]);
+//! ```
+//!
+//! # Buffer sizing
+//!
+//! Use [`max_encoded_len`] to determine the required destination buffer
+//! size before encoding:
+//!
+//! ```
+//! let data = [0x01, 0x02, 0x03];
+//! let max = ucobs::max_encoded_len(data.len()); // 4
+//!
+//! let mut buf = [0u8; 4];
+//! let n = ucobs::encode(&data, &mut buf).unwrap();
+//! assert_eq!(n, 4); // fits exactly
+//! ```
+//!
+//! If the destination is too small, `encode` returns `None` rather than
+//! panicking:
+//!
+//! ```
+//! let mut tiny = [0u8; 1];
+//! assert_eq!(ucobs::encode(&[0x01, 0x02], &mut tiny), None);
+//! ```
+//!
+//! # Framing for transport
 //!
 //! The encoder does **not** append a trailing `0x00` sentinel. The decoder
-//! expects input **without** a trailing sentinel. Callers should append /
+//! expects input **without** a trailing sentinel. Append and strip the
-//! strip the sentinel themselves when framing for transport:
+//! sentinel yourself when framing for transport:
 //!
 //! ```
 //! let data = [0x11, 0x00, 0x33];
 //!
-//! // Encode — leave room for the sentinel byte
+//! // Encode into a wire frame: [COBS bytes...] [0x00 sentinel]
 //! let mut frame = [0u8; 16];
 //! let n = ucobs::encode(&data, &mut frame).unwrap();
-//! frame[n] = 0x00; // append sentinel for framing
+//! frame[n] = 0x00; // append sentinel
+//! let wire = &frame[..n + 1];
 //!
-//! // Decode — strip sentinel before decoding
+//! // On the receiving end, strip the sentinel before decoding
+//! let cobs_data = &wire[..wire.len() - 1];
 //! let mut out = [0u8; 16];
-//! let m = ucobs::decode(&frame[..n], &mut out).unwrap();
+//! let m = ucobs::decode(cobs_data, &mut out).unwrap();
 //! assert_eq!(&out[..m], &data);
 //! ```
+//!
+//! # Parsing a stream of frames
+//!
+//! In a byte stream, split on `0x00` to extract individual COBS frames,
+//! then decode each one:
+//!
+//! ```
+//! // Simulate a stream containing two frames separated by 0x00 sentinels
+//! let stream = [
+//!     0x02, 0x11, 0x02, 0x33, 0x00,  // frame 1: encodes [0x11, 0x00, 0x33]
+//!     0x03, 0xAA, 0xBB, 0x00,        // frame 2: encodes [0xAA, 0xBB]
+//! ];
+//!
+//! let mut out = [0u8; 16];
+//! let frames: Vec<&[u8]> = stream.split(|&b| b == 0x00)
+//!     .filter(|f| !f.is_empty())
+//!     .collect();
+//!
+//! let n = ucobs::decode(frames[0], &mut out).unwrap();
+//! assert_eq!(&out[..n], &[0x11, 0x00, 0x33]);
+//!
+//! let n = ucobs::decode(frames[1], &mut out).unwrap();
+//! assert_eq!(&out[..n], &[0xAA, 0xBB]);
+//! ```
+//!
+//! # Compile-time encoding
+//!
+//! [`encode`] is `const fn`, so you can build COBS-encoded lookup tables
+//! or protocol headers at compile time with zero runtime cost:
+//!
+//! ```
+//! // Pre-encode a command table at compile time
+//! const PING: [u8; 2] = {
+//!     let mut buf = [0u8; 2];
+//!     match ucobs::encode(&[0x01], &mut buf) {
+//!         Some(_) => buf,
+//!         None => panic!("buffer too small"),
+//!     }
+//! };
+//!
+//! const ACK: [u8; 3] = {
+//!     let mut buf = [0u8; 3];
+//!     match ucobs::encode(&[0x06, 0x00], &mut buf) {
+//!         Some(_) => buf,
+//!         None => panic!("buffer too small"),
+//!     }
+//! };
+//!
+//! // No runtime encoding needed — these are baked into the binary
+//! assert_eq!(PING, [0x02, 0x01]);
+//! assert_eq!(ACK, [0x02, 0x06, 0x01]);
+//! ```
+//!
+//! # Error handling
+//!
+//! Both [`encode`] and [`decode`] return `Option<usize>` — `None` on
+//! failure, never panic.
+//!
+//! ```
+//! let mut buf = [0u8; 8];
+//!
+//! // Destination too small for encode
+//! assert_eq!(ucobs::encode(&[1, 2, 3], &mut buf[..1]), None);
+//!
+//! // Malformed COBS data (zero byte in encoded stream)
+//! assert_eq!(ucobs::decode(&[0x00], &mut buf), None);
+//!
+//! // Truncated frame (code byte promises more data than exists)
+//! assert_eq!(ucobs::decode(&[0x05, 0x11], &mut buf), None);
+//!
+//! // Empty input is valid for both
+//! assert_eq!(ucobs::encode(&[], &mut buf), Some(1)); // encodes to [0x01]
+//! assert_eq!(ucobs::decode(&[], &mut buf), Some(0)); // decodes to []
+//! ```
 
 #![no_std]
 #![forbid(unsafe_code)]
 #![warn(missing_docs)]
 
-/// Encode `src` into `dest` using COBS. Returns the number of bytes written,
+/// Copy `src` into `dest[offset..]` in a const-compatible way.
-/// or `None` if `dest` is too small.
+/// Uses `copy_from_slice` (Rust 1.93+, compiles to memcpy).
+/// With `legacy-msrv` feature, falls back to a manual loop for Rust 1.83+.
+#[inline(always)]
+#[cfg(not(feature = "legacy-msrv"))]
+const fn const_copy(dest: &mut [u8], di: usize, src: &[u8], n: usize) {
+    let (_, dt) = dest.split_at_mut(di);
+    let (dc, _) = dt.split_at_mut(n);
+    let (sc, _) = src.split_at(n);
+    dc.copy_from_slice(sc);
+}
+
+/// Fallback const copy for toolchains older than Rust 1.93.
+#[cfg_attr(test, mutants::skip)] // dead code under default features
+#[inline(always)]
+#[cfg(feature = "legacy-msrv")]
+const fn const_copy(dest: &mut [u8], di: usize, src: &[u8], n: usize) {
+    let mut i = 0;
+    while i < n {
+        dest[di + i] = src[i];
+        i += 1;
+    }
+}
+
+/// COBS-encode `src` into `dest`.
 ///
-/// The output does **not** include a trailing `0x00` sentinel — the caller
+/// Returns the number of bytes written, or `None` if `dest` is too small.
-/// should append it as a frame delimiter.
+/// Use [`max_encoded_len`] to size the destination buffer.
+///
+/// The output does **not** include a trailing `0x00` sentinel — append it
+/// yourself when framing for transport.
+///
+/// This function is `const fn`, so it can encode data at compile time.
+///
+/// # Examples
+///
+/// ```
+/// let mut buf = [0u8; 8];
+/// let n = ucobs::encode(&[0x00], &mut buf).unwrap();
+/// assert_eq!(&buf[..n], &[0x01, 0x01]);
+/// ```
+///
+/// Compile-time encoding:
+///
+/// ```
+/// const ENCODED: [u8; 4] = {
+///     let mut buf = [0u8; 4];
+///     match ucobs::encode(&[0x11, 0x00, 0x33], &mut buf) {
+///         Some(_) => buf,
+///         None => panic!("buffer too small"),
+///     }
+/// };
+/// assert_eq!(ENCODED, [0x02, 0x11, 0x02, 0x33]);
+/// ```
 #[must_use]
-pub fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
+pub const fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
     let mut si = 0;
     let mut di = 0;
 
     loop {
-        // Scan for next zero (or 254-byte cap).
         let remaining = src.len() - si;
         let max_run = if remaining < 254 { remaining } else { 254 };
-        let run = if max_run > 0 && src[si] == 0x00 {
+
-            0
+        // Fast path: leading zero byte — skip split_at overhead entirely.
-        } else {
+        if max_run > 0 && src[si] == 0x00 {
-            src[si..si + max_run]
+            if di >= dest.len() {
-                .iter()
+                return None;
-                .position(|&b| b == 0x00)
+            }
-                .unwrap_or(max_run)
+            dest[di] = 0x01;
-        };
+            di += 1;
+            si += 1;
+            continue;
+        }
+
+        // Create bounded source window for scan + copy.
+        let (_, src_tail) = src.split_at(si);
+        let (src_window, _) = src_tail.split_at(max_run);
+
+        // Scan for next zero byte.
+        let mut run = 0;
+        while run < src_window.len() && src_window[run] != 0x00 {
+            run += 1;
+        }
         let full_block = run == 254;
 
         // Write code byte.
@@ -66,27 +246,20 @@ pub fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
         dest[di] = if full_block { 0xFF } else { (run + 1) as u8 };
         di += 1;
 
-        // Bulk-copy the run.
+        // Copy the run.
         if run > 0 {
             if di + run > dest.len() {
                 return None;
             }
-            dest[di..di + run].copy_from_slice(&src[si..si + run]);
+            const_copy(dest, di, src_window, run);
             di += run;
         }
         si += run;
 
-        if full_block {
+        if si >= src.len() {
-            // 254 non-zero bytes emitted; no implicit zero.
+            break;
-            // If input is exhausted exactly here, we're done.
+        }
-            if si >= src.len() {
+        if !full_block {
-                break;
-            }
-        } else {
-            // Run ended at a zero or end-of-input.
-            if si >= src.len() {
-                break;
-            }
             si += 1; // skip the zero byte
         }
     }
@@ -94,34 +267,72 @@ pub fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
     Some(di)
 }
 
-/// Decode a COBS-encoded buffer. Returns the number of decoded bytes
+/// Decode a COBS-encoded buffer.
-/// written to `dest`, or `None` if the input is malformed or `dest` is
+///
-/// too small.
+/// Returns the number of decoded bytes written to `dest`, or `None` if the
+/// input is malformed or `dest` is too small.
 ///
 /// `src` should **not** include the trailing `0x00` frame sentinel — strip
 /// it before calling this function.
+///
+/// # Examples
+///
+/// ```
+/// let mut out = [0u8; 8];
+/// let n = ucobs::decode(&[0x02, 0x11, 0x02, 0x33], &mut out).unwrap();
+/// assert_eq!(&out[..n], &[0x11, 0x00, 0x33]);
+/// ```
+///
+/// Malformed input returns `None`:
+///
+/// ```
+/// let mut out = [0u8; 8];
+/// assert_eq!(ucobs::decode(&[0x00], &mut out), None); // unexpected zero
+/// assert_eq!(ucobs::decode(&[0x05, 0x11], &mut out), None); // truncated
+/// ```
 #[must_use]
 pub fn decode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
-    let mut si = 0;
+    let mut src = src;
     let mut di = 0;
-    while si < src.len() {
+    while let Some((&code_byte, rest)) = src.split_first() {
-        let code = src[si] as usize;
+        src = rest;
-        si += 1;
+        let code = code_byte as usize;
         if code == 0 {
-            return None; // unexpected zero in COBS data
+            return None;
         }
-        // Bulk-copy `code - 1` data bytes.
+
-        let n = code - 1;
+        if code == 1 {
-        if n > 0 {
+            // Batch consecutive 0x01 codes and fill zeros in bulk.
-            if si + n > src.len() || di + n > dest.len() {
+            let mut count = 1usize;
+            while let Some((&next, rest)) = src.split_first() {
+                if next != 0x01 {
+                    break;
+                }
+                src = rest;
+                count += 1;
+            }
+            // If more data follows, every 0x01 represents a zero byte.
+            // If the stream ends here, the final 0x01 is the end-of-data
+            // marker, not a zero — so emit one fewer.
+            let zeros = if !src.is_empty() { count } else { count - 1 };
+            if di + zeros > dest.len() {
                 return None;
             }
-            dest[di..di + n].copy_from_slice(&src[si..si + n]);
+            dest[di..di + zeros].fill(0);
-            di += n;
+            di += zeros;
-            si += n;
+            continue;
         }
-        // If code < 0xFF, an implicit zero follows (unless we're at the end).
+
-        if code < 0xFF && si < src.len() {
+        // Bulk-copy `code - 1` data bytes.
+        let n = code - 1;
+        if src.len() < n || di + n > dest.len() {
+            return None;
+        }
+        let (block, rest) = src.split_at(n);
+        dest[di..di + n].copy_from_slice(block);
+        di += n;
+        src = rest;
+        if code != 0xFF && !src.is_empty() {
             if di >= dest.len() {
                 return None;
             }
@@ -133,6 +344,17 @@ pub fn decode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
 }
 
 /// Maximum encoded size for a given source length (excluding sentinel).
+///
+/// Use this to allocate a destination buffer for [`encode`]. The formula
+/// is `src_len + (src_len / 254) + 1`.
+///
+/// # Examples
+///
+/// ```
+/// assert_eq!(ucobs::max_encoded_len(0), 1);
+/// assert_eq!(ucobs::max_encoded_len(1), 2);
+/// assert_eq!(ucobs::max_encoded_len(254), 256);
+/// ```
 #[must_use]
 pub const fn max_encoded_len(src_len: usize) -> usize {
     src_len + (src_len / 254) + 1
@@ -155,6 +377,21 @@ mod tests {
         encode(input, &mut buf).map(|n| buf[..n].to_vec())
     }
 
+    // ── Compile-time (const fn) encode ────────────────────────────
+
+    const _: () = {
+        let input = [0x11, 0x00, 0x33];
+        let mut buf = [0u8; 8];
+        let Some(n) = encode(&input, &mut buf) else {
+            panic!("const encode failed");
+        };
+        assert!(n == 4);
+        assert!(buf[0] == 0x02);
+        assert!(buf[1] == 0x11);
+        assert!(buf[2] == 0x02);
+        assert!(buf[3] == 0x33);
+    };
+
     // ── Wikipedia canonical decode vectors (Cheshire & Baker 1999) ──
 
     #[test]
@@ -597,6 +834,49 @@ mod tests {
         assert_eq!(enc(&input), Some(expected));
     }
 
+    #[test]
+    fn encode_254_block_boundary_exact() {
+        // 254 nonzero bytes MUST produce code 0xFF (full block), NOT 0xFF+1=0x00.
+        // This catches max_run being 255 instead of 254.
+        let input: Vec<u8> = (1..=254).map(|i| i as u8).collect();
+        let encoded = enc(&input).unwrap();
+        assert_eq!(encoded[0], 0xFF); // code byte for full 254-byte block
+        assert_eq!(encoded.len(), 255); // 1 code byte + 254 data bytes
+                                        // Verify it round-trips
+        let mut dec = vec![0u8; 256];
+        let n = decode(&encoded, &mut dec).unwrap();
+        assert_eq!(&dec[..n], &input[..]);
+    }
+
+    #[test]
+    fn encode_255_nonzero_needs_two_blocks() {
+        // 255 nonzero bytes must split into a 254-byte full block + 1-byte block.
+        // If max_run were 255 (bug), this would be a single block — wrong.
+        let input: Vec<u8> = (1..=255).map(|i| (i % 255 + 1) as u8).collect();
+        let encoded = enc(&input).unwrap();
+        assert_eq!(encoded[0], 0xFF); // first block: full 254
+        assert_eq!(encoded[255], 0x02); // second block: 1 byte (code = 2)
+        assert_eq!(encoded.len(), 257); // 1 + 254 + 1 + 1
+    }
+
+    #[test]
+    fn encode_block_boundary_matches_corncobs() {
+        // Verify byte-for-byte match with corncobs at the critical 254-byte
+        // block boundary. Catches max_run being off by one.
+        for len in [254, 255, 256, 508, 512] {
+            let input: Vec<u8> = (0..len).map(|i| (i % 254 + 1) as u8).collect();
+            let ours = enc(&input).unwrap();
+            let mut their_buf = vec![0u8; corncobs::max_encoded_len(input.len())];
+            let their_len = corncobs::encode_buf(&input, &mut their_buf);
+            // corncobs appends a sentinel; strip it for comparison
+            assert_eq!(
+                &ours[..],
+                &their_buf[..their_len - 1],
+                "mismatch at len={len}"
+            );
+        }
+    }
+
     // ── Round-trip tests ────────────────────────────────────────
 
     #[test]
@@ -729,11 +1009,11 @@ mod tests {
 
     #[test]
     fn interop_random_payloads() {
-        use rand::Rng;
+        use rand::RngExt;
-        let mut rng = rand::thread_rng();
+        let mut rng = rand::rng();
         for _ in 0..500 {
-            let len = rng.gen_range(0..=512);
+            let len = rng.random_range(0..=512);
-            let data: Vec<u8> = (0..len).map(|_| rng.gen()).collect();
+            let data: Vec<u8> = (0..len).map(|_| rng.random()).collect();
             cross_encode(&data);
             cross_decode(&data);
         }
@@ -800,6 +1080,29 @@ mod tests {
         assert_eq!(out, [0x11, 0x22]);
     }
 
+    #[test]
+    fn decode_zeros_dest_exact_size() {
+        // Decode all-zeros into an exact-fit buffer.
+        // Encoded: N+1 bytes of 0x01 → N bytes of 0x00.
+        // This catches `> dest.len()` mutated to `>= dest.len()`.
+        for n in [1, 2, 10, 64, 254, 255, 256] {
+            let encoded: Vec<u8> = vec![0x01; n + 1];
+            let mut out = vec![0u8; n]; // exact fit
+            assert_eq!(decode(&encoded, &mut out), Some(n), "failed at n={n}");
+            assert!(out.iter().all(|&b| b == 0x00), "wrong data at n={n}");
+        }
+    }
+
+    #[test]
+    fn decode_zeros_dest_one_too_small() {
+        // Buffer one byte too small for all-zeros decode.
+        for n in [2, 10, 64, 256] {
+            let encoded: Vec<u8> = vec![0x01; n + 1];
+            let mut out = vec![0u8; n - 1]; // one too small
+            assert_eq!(decode(&encoded, &mut out), None, "should fail at n={n}");
+        }
+    }
+
     // ── Multi-block boundary tests ──────────────────────────────
 
     #[test]
@@ -955,7 +1258,7 @@ mod proptests {
             shrink in 1usize..256,
         ) {
             let max = max_encoded_len(data.len());
-            let dest_size = if shrink >= max { 0 } else { max - shrink };
+            let dest_size = max.saturating_sub(shrink);
             let mut buf = vec![0u8; dest_size];
             let _ = encode(&data, &mut buf); // may return None, must not panic
         }