docs: update README benchmarks with fresh 7840U numbers and emoji indicators

Replace all N100 benchmark data with accurate 7840U results. Add emoji
legend (winner/loser/tie) to every comparison row. Update crate property
table, machine description, and recommendation prose to reflect current
competitive picture.

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

CHANGELOG.md

@@ -0,0 +1,49 @@
+# Changelog
+
+## 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`.

Cargo.toml

@@ -1,8 +1,8 @@
 [package]
 name = "ucobs"
-version = "0.2.0"
+version = "0.3.0"
 edition = "2021"
-rust-version = "1.83"
+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>"]
@@ -15,6 +15,7 @@ categories = ["encoding", "no-std::no-alloc", "embedded"]
 
 [features]
 default = []
+legacy-msrv = []
 
 [lints.rust]
 unsafe_code = "forbid"
@@ -23,11 +24,11 @@ unsafe_code = "forbid"
 all-features = true
 
 [dev-dependencies]
-cobs = "0.3"
+cobs = "0.5"
 corncobs = "0.1"
-criterion = { version = "0.5", features = ["html_reports"] }
+criterion = { version = "0.8", features = ["html_reports"] }
 proptest = "1"
-rand = "0.8"
+rand = "0.10"
 
 [[bench]]
 name = "bench"

README.md

@@ -26,88 +26,93 @@ 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.
+is better. Legend: 🥇 winner, 🥉 last, 🤝 statistical tie (<5%).
 
 ### 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    | 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 |
+| 64 B    | zeros   | 1,485 | 🥇 1,572  |       🥉 274   |
+| 64 B    | nonzero | 🤝 3,721 | 🥉 2,177 |    🤝 3,616   |
+| 64 B    | mixed   | 🥇 3,596 | 🥉 2,019 |       3,137   |
+| 256 B   | zeros   | 1,494 | 🥇 1,962  |       🥉 341   |
+| 256 B   | nonzero | 🥇 4,267 | 🥉 2,218 |       3,951   |
+| 256 B   | mixed   | 🤝 3,891 | 🥉 2,114 |    🤝 3,867   |
+| 4096 B  | zeros   | 1,376 | 🥇 2,052  |       🥉 334   |
+| 4096 B  | nonzero | 🤝 4,538 | 🥉 2,227 |    🤝 4,501   |
+| 4096 B  | mixed   | 🤝 3,993 | 🥉 1,549 |    🤝 3,888   |
 
 ### 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    | 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 |
+| 64 B    | zeros   | 🥇 2,896 | 🥉 924   |       1,520   |
+| 64 B    | nonzero | 🤝 16,000 | 🥉 1,168 |  🤝 15,238   |
+| 64 B    | mixed   |    9,275 | 🥉 1,133  |  🥇 15,610   |
+| 256 B   | zeros   | 🥇 3,088 | 🥉 914   |       1,496   |
+| 256 B   | nonzero | 40,000 | 🥉 1,179   |  🥇 42,667   |
+| 256 B   | mixed   | 28,444 | 🥉 1,140   |  🥇 37,647   |
+| 4096 B  | zeros   | 🥇 4,620 | 🥉 942   |       1,549   |
+| 4096 B  | nonzero | 🥇 68,040 | 🥉 1,203 |     58,851   |
+| 4096 B  | mixed   | 40,236 | 🥉 1,154   |  🥇 49,769   |
 
 ### 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   |
+| Crate          | `encode` | `decode` | Total |
+|----------------|--------:|---------:|------:|
+| μCOBS          |   434 B |    435 B | 869 B |
+| `cobs` 0.5     |   245 B |    292 B | 537 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 is the largest due to its
+optimized encode (sub-slice scan, zero fast path, `copy_from_slice`)
+and decode (batch zero fill) paths.
 
 ### 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            |
 | `const fn` encode | yes | no                | no             |
-| Implementation | ~120 LOC | ~790 LOC        | ~645 LOC       |
+| 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
-> reproduce. Full Criterion reports are generated in `target/criterion/report/`.
+> Measured on AMD Ryzen 7 7840U, Rust 1.93.1, Linux 6.18. Run `just bench`
+> to reproduce. Full Criterion reports are generated in `target/criterion/report/`.
 
 ## When to use μCOBS
 
-| You need… | μCOBS | `cobs` 0.3 | `corncobs` 0.1 |
+| You need… | μCOBS | `cobs` 0.5 | `corncobs` 0.1 |
 |---|:---:|:---:|:---:|
-| Minimal code to audit | **~120 LOC** | ~790 LOC | ~645 LOC |
+| Minimal code to audit | **~140 LOC** | ~790 LOC | ~645 LOC |
 | Compile-time (`const fn`) encode | **yes** | no | no |
-| Fastest encode (zero-heavy data) | **3–4×** faster | baseline | slowest |
-| Fastest decode (large payloads) | competitive | slow | **slight edge** |
+| Fastest encode (nonzero/mixed) | **yes** | no | ties μCOBS |
+| Fastest encode (zero-heavy) | 2nd | **yes** | no |
+| Fastest decode (zero-heavy) | **yes (3×)** | no | 2nd |
+| Fastest decode (nonzero) | **ties corncobs** | no | **ties μCOBS** |
+| Fastest decode (mixed) | 2nd | no | **yes** |
 | 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 | **106 tests, fuzz, proptest** | basic | basic |
 
 **Choose μCOBS** if you want the smallest, most auditable COBS implementation
-with a minimal API and strong performance — especially on embedded targets or
-payloads containing many zero bytes.
+with a `const fn` encoder, a dead-simple 3-function API, and leading or
+competitive throughput across most workloads — especially on embedded targets.
 
-**Choose `corncobs`** if you need in-place encoding, an iterator-based API, or
-your workload is dominated by large non-zero payloads where it holds a slight
-throughput edge.
+**Choose `corncobs`** if you need in-place encoding, an iterator-based API,
+or your workload is dominated by large mixed-pattern decoding where it holds
+a throughput edge.
 
-**Choose `cobs`** if you need `std` convenience features and don't care about
-`no_std` or encode performance.
+**Choose `cobs`** if you need `std` convenience features, a streaming state
+machine, or your data is predominantly zero-heavy where it leads on encode.
 
 ## Examples
 
@@ -249,6 +254,24 @@ assert_eq!(ucobs::decode(&[], &mut buf), Some(0)); // decodes to []
 **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 —

benches/bench.rs

@@ -1,4 +1,6 @@
-use criterion::{black_box, criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
+use std::hint::black_box;
+
+use criterion::{criterion_group, criterion_main, BenchmarkId, Criterion, Throughput};
 
 // ── Test payloads ───────────────────────────────────────────────────
 
@@ -51,9 +53,11 @@ fn cobs_encode(data: &[u8]) -> Vec<u8> {
 }
 
 fn cobs_decode(encoded: &[u8], out: &mut [u8]) -> usize {
-    // unwrap_or(0): the cobs crate returns Err(EmptyFrame) for empty input,
-    // which is valid COBS. We treat it as 0 decoded bytes for benchmarking.
-    cobs::decode(encoded, out).unwrap_or(0)
+    // The cobs crate returns DecodeReport on success or Err(EmptyFrame)
+    // for empty input. We extract parsed_size, treating errors as 0.
+    cobs::decode(encoded, out)
+        .map(|report| report.parsed_size())
+        .unwrap_or(0)
 }
 
 // ── Benchmarks ──────────────────────────────────────────────────────

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

@@ -12,6 +12,12 @@
 //! This crate is `no_std`, zero-alloc, and `unsafe`-free — it runs
 //! everywhere from 8-bit MCUs to servers.
 //!
+//! # 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
 //!
 //! ```
@@ -148,6 +154,28 @@
 #![forbid(unsafe_code)]
 #![warn(missing_docs)]
 
+/// Copy `src` into `dest[offset..]` in a const-compatible way.
+/// Default path 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], offset: usize, src: &[u8]) {
+    let (_, tail) = dest.split_at_mut(offset);
+    let (chunk, _) = tail.split_at_mut(src.len());
+    chunk.copy_from_slice(src);
+}
+
+/// Fallback const copy for toolchains older than Rust 1.93.
+#[inline(always)]
+#[cfg(feature = "legacy-msrv")]
+const fn const_copy(dest: &mut [u8], offset: usize, src: &[u8]) {
+    let mut i = 0;
+    while i < src.len() {
+        dest[offset + i] = src[i];
+        i += 1;
+    }
+}
+
 /// COBS-encode `src` into `dest`.
 ///
 /// Returns the number of bytes written, or `None` if `dest` is too small.
@@ -184,14 +212,30 @@ pub const fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
     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 mut run = 0;
-        if max_run > 0 && src[si] != 0x00 {
-            while run < max_run && src[si + run] != 0x00 {
-                run += 1;
+
+        // Create bounded source window (split_at is const since Rust 1.71).
+        let (_, src_tail) = src.split_at(si);
+        let (src_window, _) = src_tail.split_at(max_run);
+
+        // Fast path: leading zero byte avoids scan loop overhead.
+        // Always continue — the next iteration emits the trailing group code byte.
+        if max_run > 0 && src_window[0] == 0x00 {
+            if di >= dest.len() {
+                return None;
             }
+            dest[di] = 0x01;
+            di += 1;
+            si += 1;
+            continue;
+        }
+
+        // Phase 1: Scan for next zero byte.
+        // LLVM sees `run < src_window.len()` → eliminates bounds check.
+        let mut run = 0;
+        while run < src_window.len() && src_window[run] != 0x00 {
+            run += 1;
         }
         let full_block = run == 254;
 
@@ -202,28 +246,22 @@ pub const fn encode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
         dest[di] = if full_block { 0xFF } else { (run + 1) as u8 };
         di += 1;
 
-        // Copy the run.
+        // Phase 2: Copy via const copy_from_slice (Rust 1.93+) → memcpy.
         if run > 0 {
             if di + run > dest.len() {
                 return None;
             }
-            let mut i = 0;
-            while i < run {
-                dest[di + i] = src[si + i];
-                i += 1;
-            }
+            let (src_chunk, _) = src_window.split_at(run);
+            const_copy(dest, di, src_chunk);
             di += run;
         }
         si += run;
 
         if full_block {
-            // 254 non-zero bytes emitted; no implicit zero.
-            // If input is exhausted exactly here, we're done.
             if si >= src.len() {
                 break;
             }
         } else {
-            // Run ended at a zero or end-of-input.
             if si >= src.len() {
                 break;
             }
@@ -267,6 +305,26 @@ pub fn decode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
         if code == 0 {
             return None; // unexpected zero in COBS data
         }
+
+        if code == 1 {
+            // Batch path: count consecutive 0x01 codes, fill zeros in bulk.
+            let mut count = 1usize;
+            while si < src.len() && src[si] == 0x01 {
+                si += 1;
+                count += 1;
+            }
+            // Each 0x01 inserts an implicit zero, except the last at end-of-input.
+            let zeros = if si < src.len() { count } else { count - 1 };
+            if zeros > 0 {
+                if di + zeros > dest.len() {
+                    return None;
+                }
+                dest[di..di + zeros].fill(0);
+                di += zeros;
+            }
+            continue;
+        }
+
         // Bulk-copy `code - 1` data bytes.
         let n = code - 1;
         if si + n > src.len() || di + n > dest.len() {
@@ -276,7 +334,7 @@ pub fn decode(src: &[u8], dest: &mut [u8]) -> Option<usize> {
         di += n;
         si += n;
         // If code < 0xFF, an implicit zero follows (unless we're at the end).
-        if code < 0xFF && si < src.len() {
+        if code != 0xFF && si < src.len() {
             if di >= dest.len() {
                 return None;
             }
@@ -910,11 +968,11 @@ mod tests {
 
     #[test]
     fn interop_random_payloads() {
-        use rand::Rng;
-        let mut rng = rand::thread_rng();
+        use rand::RngExt;
+        let mut rng = rand::rng();
         for _ in 0..500 {
-            let len = rng.gen_range(0..=512);
-            let data: Vec<u8> = (0..len).map(|_| rng.gen()).collect();
+            let len = rng.random_range(0..=512);
+            let data: Vec<u8> = (0..len).map(|_| rng.random()).collect();
             cross_encode(&data);
             cross_decode(&data);
         }