feat: add benchmarks, justfile, optimize throughput, harden quality gates

Benchmarking & comparison:
- Add criterion benchmarks (ucobs vs cobs 0.3 vs corncobs 0.1)
- Add Python cobs benchmark script
- Add performance comparison section to README with full results

Performance:
- Rewrite encode to scan-and-bulk-copy via copy_from_slice
- Rewrite decode to bulk-copy via copy_from_slice
- ~2-4x throughput improvement; now matches corncobs on most workloads

Quality hardening:
- Add #![forbid(unsafe_code)], #![warn(missing_docs)], #[must_use]
- Add [lints.rust] unsafe_code = "forbid" in Cargo.toml
- Add rust-version (MSRV) = "1.56"
- Add docs.rs metadata
- 106 tests: 80 unit/corpus + 25 property tests (25k cases each = 625k total)
- External test vectors from: Wikipedia, Cheshire & Baker paper, nanocobs,
  cobs-c, cobs2-rs, Jacques Fortier, Python cobs package
- Add fuzz_small_dest fuzz target (encode/decode with undersized buffers)
- Improve doc examples to show sentinel append/strip convention

Infrastructure:
- Add justfile with test (nextest), check, bench, fuzz, miri, fmt targets
- cargo-nextest for parallel test execution
- Update .gitignore for Python cache files

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

README.md

@@ -20,6 +20,70 @@ allowing `0x00` to serve as an unambiguous frame delimiter. Overhead is exactly
 - **Insanely tested** — canonical vectors, property-based tests, fuzz targets,
   cross-crate interop, randomized payloads, and more coming
 
+## Performance
+
+Benchmarked against the two other `no_std` COBS crates using
+[Criterion](https://github.com/bheisler/criterion.rs) across three payload
+patterns: **all zeros** (worst case — every byte triggers a code emit),
+**no zeros**, and **mixed** (0x00–0xFF cycling). Throughput in MB/s, higher
+is better.
+
+### Encode
+
+| Payload | Pattern | μCOBS | `cobs` 0.3 | `corncobs` 0.1 |
+|---------|---------|------:|-----------:|---------------:|
+| 64 B    | zeros   |   595 |        748 |            180 |
+| 64 B    | mixed   | 1,853 |      1,011 |          1,734 |
+| 64 B    | nonzero | 1,884 |      1,012 |          2,110 |
+| 256 B   | zeros   |   654 |        817 |            184 |
+| 256 B   | mixed   | 1,888 |      1,002 |          1,887 |
+| 256 B   | nonzero | 1,964 |      1,004 |          1,968 |
+| 4096 B  | zeros   |   675 |        845 |            188 |
+| 4096 B  | mixed   | 1,734 |      1,016 |          1,746 |
+| 4096 B  | nonzero | 1,775 |      1,020 |          2,245 |
+
+### Decode
+
+| Payload | Pattern | μCOBS | `cobs` 0.3 | `corncobs` 0.1 |
+|---------|---------|------:|-----------:|---------------:|
+| 64 B    | zeros   |   595 |        440 |            732 |
+| 64 B    | mixed   | 7,041 |        504 |          7,991 |
+| 64 B    | nonzero | 9,900 |        502 |          9,807 |
+| 256 B   | zeros   |   656 |        472 |            742 |
+| 256 B   | mixed   |17,024 |        645 |         16,207 |
+| 256 B   | nonzero |18,456 |        647 |         19,549 |
+| 4096 B  | zeros   |   675 |        483 |            768 |
+| 4096 B  | mixed   |17,522 |        656 |         22,959 |
+| 4096 B  | nonzero |37,987 |        623 |         40,005 |
+
+### Code size
+
+Measured from `.text` section of release-optimized symbols (`encode` + `decode`
+combined):
+
+| Crate          | Code size |
+|----------------|----------:|
+| μCOBS          |   541 B   |
+| `cobs` 0.3     |   537 B   |
+| `corncobs` 0.1 |   637 B   |
+
+All three crates are tiny. The difference is negligible for any target.
+
+### Crate properties
+
+| Property       | μCOBS  | `cobs` 0.3       | `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       |
+
+[^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 `cargo bench` to
+> reproduce. Full Criterion reports are generated in `target/criterion/report/`.
+
 ## Usage
 
 ```rust