lev-rs 0.2.2__tar.gz → 0.2.4__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (52) hide show
  1. lev_rs-0.2.4/.claude/agents/code-cartographer.md +42 -0
  2. lev_rs-0.2.4/.claude/agents/competitor-scout.md +38 -0
  3. lev_rs-0.2.4/.claude/agents/microarch-checker.md +46 -0
  4. lev_rs-0.2.4/.claude/agents/profile-runner.md +32 -0
  5. lev_rs-0.2.4/.claude/settings.json +22 -0
  6. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/dependabot.yaml +4 -0
  7. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/workflows/benchmark.yml +10 -5
  8. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/workflows/docs.yml +2 -2
  9. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/workflows/prek.yaml +2 -2
  10. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/workflows/release.yml +20 -36
  11. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/workflows/test.yml +3 -3
  12. {lev_rs-0.2.2 → lev_rs-0.2.4}/.gitignore +2 -0
  13. {lev_rs-0.2.2 → lev_rs-0.2.4}/.vscode/settings.json +3 -1
  14. {lev_rs-0.2.2 → lev_rs-0.2.4}/AGENTS.md +34 -5
  15. {lev_rs-0.2.2 → lev_rs-0.2.4}/CHANGELOG.md +17 -0
  16. {lev_rs-0.2.2 → lev_rs-0.2.4}/Cargo.lock +11 -11
  17. {lev_rs-0.2.2 → lev_rs-0.2.4}/Cargo.toml +1 -1
  18. {lev_rs-0.2.2 → lev_rs-0.2.4}/PKG-INFO +9 -10
  19. {lev_rs-0.2.2 → lev_rs-0.2.4}/README.md +7 -1
  20. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_ascii_dark.svg +219 -133
  21. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_ascii_light.svg +219 -133
  22. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_cjk_dark.svg +265 -279
  23. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_cjk_light.svg +265 -279
  24. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_emoji_dark.svg +212 -154
  25. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_emoji_light.svg +212 -154
  26. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_latin_1_dark.svg +264 -156
  27. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/assets/benchmark_latin_1_light.svg +264 -156
  28. lev_rs-0.2.4/docs/assets/benchmark_realistic_dark.svg +1283 -0
  29. lev_rs-0.2.4/docs/assets/benchmark_realistic_light.svg +1283 -0
  30. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/index.md +10 -2
  31. {lev_rs-0.2.2 → lev_rs-0.2.4}/prek.toml +10 -1
  32. {lev_rs-0.2.2 → lev_rs-0.2.4}/pyproject.toml +8 -20
  33. lev_rs-0.2.4/scripts/benchmark.py +284 -0
  34. {lev_rs-0.2.2 → lev_rs-0.2.4}/scripts/benchmark_by_length.py +105 -4
  35. {lev_rs-0.2.2 → lev_rs-0.2.4}/src/lib.rs +797 -117
  36. lev_rs-0.2.4/tests/conftest.py +21 -0
  37. lev_rs-0.2.4/tests/test_lev.py +401 -0
  38. {lev_rs-0.2.2 → lev_rs-0.2.4}/uv.lock +8 -892
  39. lev_rs-0.2.4/version +1 -0
  40. {lev_rs-0.2.2 → lev_rs-0.2.4}/zensical.toml +0 -4
  41. lev_rs-0.2.2/scripts/benchmark.py +0 -149
  42. lev_rs-0.2.2/scripts/install_dev.sh +0 -38
  43. lev_rs-0.2.2/tests/test_lev.py +0 -118
  44. lev_rs-0.2.2/version +0 -1
  45. {lev_rs-0.2.2 → lev_rs-0.2.4}/.cargo/config.toml +0 -0
  46. {lev_rs-0.2.2 → lev_rs-0.2.4}/.gemini/settings.json +0 -0
  47. {lev_rs-0.2.2 → lev_rs-0.2.4}/.github/CODEOWNERS +0 -0
  48. {lev_rs-0.2.2 → lev_rs-0.2.4}/LICENSE +0 -0
  49. {lev_rs-0.2.2 → lev_rs-0.2.4}/build.rs +0 -0
  50. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/api.md +0 -0
  51. {lev_rs-0.2.2 → lev_rs-0.2.4}/docs/overrides/partials/integrations/analytics/umami.html +0 -0
  52. {lev_rs-0.2.2 → lev_rs-0.2.4}/lev.pyi +0 -0
@@ -0,0 +1,42 @@
1
+ ---
2
+ name: code-cartographer
3
+ description: Maps the lev Rust source, identifies the hot function per CPython string kind, and reports algorithm, allocations, branch structure, and PyO3 boundary work. Read-only.
4
+ tools: Read, Glob, Grep
5
+ model: opus
6
+ ---
7
+ You map the codebase. Never edit source files.
8
+
9
+ Deliverables, in this order:
10
+
11
+ 1. **Entry-point map.** For each public Python function (`lev.distance`,
12
+ `lev.ratio`, anything else exposed via PyO3), trace the call path from
13
+ the `#[pyfunction]` down to the innermost DP loop. Note the file and
14
+ line range for each hop.
15
+
16
+ 2. **Per-kind specialization.** CPython strings come in four kinds
17
+ (ASCII, UCS-1 Latin-1, UCS-2, UCS-4). For each kind, report:
18
+ - Is there a dedicated code path, or does everything funnel into one
19
+ generic implementation?
20
+ - How is the raw buffer accessed (PyUnicode_KIND + raw pointer, or
21
+ `to_str` with UTF-8 decode)?
22
+ - What integer width is the DP row using?
23
+
24
+ 3. **Hot loop anatomy.** For the innermost distance loop:
25
+ - Algorithm (Wagner-Fischer, Myers bit-parallel, Ukkonen banded, other).
26
+ - Row width and reuse (allocated per call? thread-local? stack?).
27
+ - Branch structure of the min-of-three update.
28
+ - Any `unsafe`, `unreachable_unchecked`, SIMD intrinsics, or explicit
29
+ unrolling.
30
+ - Any early-exit conditions (max distance, length diff, common
31
+ prefix/suffix trim).
32
+
33
+ 4. **PyO3 boundary work.** Any per-call cost outside the DP: argument
34
+ extraction, GIL handling, allocation for the return value, string kind
35
+ dispatch.
36
+
37
+ 5. **Test/bench inventory.** List every criterion bench and every
38
+ `scripts/benchmark.py` scenario, one line each: name, inputs, what it
39
+ measures.
40
+
41
+ Format as markdown with file:line references in the form `src/lib.rs:42`.
42
+ Do not propose changes.
@@ -0,0 +1,38 @@
1
+ ---
2
+ name: competitor-scout
3
+ description: Reads rapidfuzz, edlib, and editdistance source to summarize the algorithms and low-level tricks each uses for Levenshtein distance. Returns a comparison table plus notable ideas not present in lev.
4
+ tools: Read, Glob, Grep, Bash, WebFetch
5
+ model: opus
6
+ ---
7
+ You survey competing Levenshtein implementations. Read-only on our repo.
8
+
9
+ For each of rapidfuzz (C++), edlib (C++), editdistance (Cython):
10
+ 1. Locate the source. If not already vendored somewhere in this repo, fetch
11
+ the relevant file(s) from GitHub via WebFetch. Prefer:
12
+ - rapidfuzz-cpp: `rapidfuzz/distance/Levenshtein_impl.hpp` and related.
13
+ - edlib: `edlib/src/edlib.cpp`.
14
+ - editdistance: `editdistance/_editdistance.h` / `.cpp`.
15
+ 2. Identify the core algorithm and the specific tricks used:
16
+ - Bit-parallel (Myers / Hyyrö)? What word width, single-word vs
17
+ multi-word (block) variant?
18
+ - Banded / Ukkonen cutoff?
19
+ - Common prefix/suffix trimming?
20
+ - Length-based short-circuits?
21
+ - SIMD (SSE/AVX/NEON)?
22
+ - Per-encoding or per-alphabet-size specializations?
23
+ - Small-string special cases (len 0, 1, equal length)?
24
+ - Scratch buffer strategy (stack, thread-local, pool)?
25
+ 3. Note any correctness subtleties (Unicode normalization, code point vs
26
+ code unit, surrogate handling).
27
+
28
+ Return two artifacts:
29
+
30
+ **Comparison table** — one row per library, columns for each trick above
31
+ (yes/no/details), plus a "notes" column.
32
+
33
+ **Ideas worth stealing** — a bulleted list of specific techniques used by
34
+ competitors that (based on your read of the code) are likely absent or
35
+ weaker in lev. Rank by expected impact on 100-char inputs. Do not propose
36
+ implementation — just name the idea and cite the competitor's file:line.
37
+
38
+ Do not edit lev's source.
@@ -0,0 +1,46 @@
1
+ ---
2
+ name: microarch-checker
3
+ description: Inspects generated assembly for lev's hot loop on Apple Silicon (aarch64), reports whether NEON is used, checks for spills and suboptimal codegen. Read-only.
4
+ tools: Read, Glob, Grep, Bash
5
+ model: opus
6
+ ---
7
+ You inspect codegen. Never edit source files.
8
+
9
+ Setup:
10
+ - Target is aarch64-apple-darwin (Mac Mini M2 Pro benchmark box).
11
+ - Ensure `cargo-show-asm` is available: `cargo install cargo-show-asm` only
12
+ if the user has confirmed; otherwise fall back to
13
+ `RUSTFLAGS="--emit=asm" cargo build --release` and read the `.s` files
14
+ under `target/release/deps/`.
15
+
16
+ For the innermost distance function identified by code-cartographer (if
17
+ that report is available in the conversation, use it; otherwise find it
18
+ yourself via Grep for the DP loop):
19
+
20
+ 1. Dump the assembly with `cargo asm --lib --rust --target aarch64-apple-darwin
21
+ <fully::qualified::path>` (or the emit-asm fallback). Save to
22
+ `/tmp/lev-hot.s`.
23
+
24
+ 2. Report:
25
+ - Is NEON in use? (Look for `v0.16b`, `v0.8h`, `ld1`, `st1`, `umin`,
26
+ `smin`, `add v`, etc.) If scalar-only, say so plainly.
27
+ - Are the min-of-three updates vectorized, or done in scalar with
28
+ `csel` / `cmp` / `b.lt`?
29
+ - Register pressure: any stack spills in the inner loop
30
+ (`str`/`ldr` targeting `[sp, #...]`)?
31
+ - Loop structure: unrolled? bounds-checked on every iteration
32
+ (`bl __rust_panic` or similar in-loop)?
33
+ - Any surprising library calls inside the loop (memcpy, alloc, panic
34
+ handlers).
35
+
36
+ 3. Cross-check against the Cargo release profile in `Cargo.toml`: `opt-level`,
37
+ `lto`, `codegen-units`, `panic`. Report current values and flag if any
38
+ are suboptimal for a hot numeric loop (but do not change them).
39
+
40
+ Return a markdown report:
41
+ - **Codegen summary** — 3-5 bullets, plain language.
42
+ - **Evidence** — 10-20 lines of the actual assembly with annotations.
43
+ - **Flags to consider** — profile settings or RUSTFLAGS worth testing in
44
+ a *separate* experiment. Explicitly do not recommend changing them here.
45
+
46
+ If tools are missing, report which and stop.
@@ -0,0 +1,32 @@
1
+ ---
2
+ name: profile-runner
3
+ description: Profiles scripts/benchmark.py and the criterion benches with samply and cargo flamegraph, returns top self-time frames per encoding. Read-only, never edits source.
4
+ tools: Bash, Read, Glob
5
+ model: haiku
6
+ ---
7
+ You profile only. Never edit source files.
8
+
9
+ For each encoding kind in {ascii, latin1, cjk, emoji}:
10
+
11
+ 1. Profile the Python-level benchmark for that kind under samply:
12
+ samply record -o profile-<kind>.json -- \
13
+ uv run scripts/benchmark.py --kind <kind> --save /tmp/bench-<kind>.json
14
+ If samply is not installed, report and stop — do not try to install it.
15
+
16
+ 2. Profile the Rust-level criterion benches under cargo flamegraph, one
17
+ flamegraph per bench binary that exists in `benches/`:
18
+ cargo flamegraph -o flame-<bench>.svg --bench <bench>
19
+ Do not pass `--profile-time` — criterion may not recognize it and it
20
+ is not required for profiling. If a per-kind bench does not exist,
21
+ profile whichever benches do and note the coverage gap in your report.
22
+ If cargo-flamegraph is not installed, report and stop.
23
+
24
+ Return one markdown table per artifact:
25
+ | Frame | %self | %total | notes |
26
+ Include the top 5 frames by self-time. In the `notes` column flag anything
27
+ that looks like: PyO3 boundary work, UTF-8 decode, allocation, memcpy,
28
+ unexpected library functions.
29
+
30
+ End with a 3-bullet summary: (1) hottest single frame across all kinds,
31
+ (2) biggest per-kind difference, (3) anything that looks like low-hanging
32
+ fruit. Do not propose fixes — that is the main session's job.
@@ -0,0 +1,22 @@
1
+ {
2
+ "$schema": "https://json.schemastore.org/claude-code-settings.json",
3
+ "hooks": {
4
+ "PostToolUse": [
5
+ {
6
+ "matcher": "Write|Edit|MultiEdit",
7
+ "hooks": [
8
+ {
9
+ "type": "command",
10
+ "command": "cd \"$CLAUDE_PROJECT_DIR\" && uv run --no-sync ruff check --fix . >&2 && uv run --no-sync ruff format . >&2 || exit 2",
11
+ "timeout": 60
12
+ },
13
+ {
14
+ "type": "command",
15
+ "command": "cd \"$CLAUDE_PROJECT_DIR\" && cargo fmt >&2 || exit 2",
16
+ "timeout": 60
17
+ }
18
+ ]
19
+ }
20
+ ]
21
+ }
22
+ }
@@ -12,6 +12,8 @@ updates:
12
12
  interval: "daily"
13
13
  assignees:
14
14
  - "KarelZe"
15
+ cooldown:
16
+ default-days: 7
15
17
  # Maintain dependencies for GitHub Actions
16
18
  - package-ecosystem: "github-actions"
17
19
  directory: "/"
@@ -19,3 +21,5 @@ updates:
19
21
  interval: "daily"
20
22
  assignees:
21
23
  - "KarelZe"
24
+ cooldown:
25
+ default-days: 7
@@ -13,14 +13,19 @@ on:
13
13
 
14
14
  jobs:
15
15
  benchmarks:
16
- name: CodSpeed benchmarks
16
+ name: CodSpeed benchmarks (py${{ matrix.python-version }})
17
17
  runs-on: ubuntu-latest
18
+ strategy:
19
+ fail-fast: false
20
+ matrix:
21
+ python-version: ["3.13", "3.14", "3.15"]
18
22
  steps:
19
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
23
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
20
24
 
21
- - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
25
+ - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
22
26
  with:
23
- python-version: "3.13"
27
+ python-version: ${{ matrix.python-version }}
28
+ allow-prereleases: true
24
29
 
25
30
  - uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
26
31
 
@@ -30,7 +35,7 @@ jobs:
30
35
  - name: Install benchmark deps
31
36
  run: pip install pytest pytest-codspeed pytest-cov
32
37
 
33
- - uses: CodSpeedHQ/action@c381be0bfd20e844fb45594f6aa182ffcd94545c # v4.15.0
38
+ - uses: CodSpeedHQ/action@4296e51e7041e24dadb86d1d6e8b9320d223dbe8 # v5.0.3
34
39
  with:
35
40
  mode: simulation
36
41
  token: ${{ secrets.CODSPEED_TOKEN }}
@@ -18,9 +18,9 @@ jobs:
18
18
  build:
19
19
  runs-on: ubuntu-latest
20
20
  steps:
21
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
21
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
22
22
 
23
- - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
23
+ - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
24
24
  with:
25
25
  python-version: "3.13"
26
26
 
@@ -6,8 +6,8 @@ jobs:
6
6
  prek:
7
7
  runs-on: ubuntu-latest
8
8
  steps:
9
- - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6
9
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v6
10
10
  - uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
11
11
  with:
12
12
  components: rustfmt
13
- - uses: j178/prek-action@6ad80277337ad479fe43bd70701c3f7f8aa74db3 # v2
13
+ - uses: j178/prek-action@4e14d07f9231acabce116ccfca13b13dd9755ece # v3.0.0
@@ -15,36 +15,20 @@ jobs:
15
15
  runs-on: ubuntu-latest
16
16
  strategy:
17
17
  matrix:
18
- include:
19
- - target: x86_64
20
- # x86_64: explicit CPython paths to skip PyPy in the manylinux container.
21
- args: >-
22
- --release --out dist
23
- --interpreter
24
- /opt/python/cp310-cp310/bin/python
25
- /opt/python/cp311-cp311/bin/python
26
- /opt/python/cp312-cp312/bin/python
27
- /opt/python/cp313-cp313/bin/python
28
- /opt/python/cp314-cp314/bin/python
29
- # FIXME: currently disabled as paths cannot be found.
30
- # - target: aarch64
31
- # # aarch64: explicit CPython paths (cross-compiled).
32
- # # Consistent with x86_64 to ensure specific versions and skip PyPy.
33
- # args: >-
34
- # --release --out dist
35
- # --interpreter
36
- # /opt/python/cp310-cp310/bin/python
37
- # /opt/python/cp311-cp311/bin/python
38
- # /opt/python/cp312-cp312/bin/python
39
- # /opt/python/cp313-cp313/bin/python
40
- # /opt/python/cp314-cp314/bin/python
18
+ target: [x86_64, aarch64]
41
19
  steps:
42
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
20
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
43
21
  - uses: PyO3/maturin-action@e83996d129638aa358a18fbd1dfb82f0b0fb5d3b # v1
44
22
  with:
45
23
  target: ${{ matrix.target }}
46
24
  manylinux: auto
47
- args: ${{ matrix.args }}
25
+ # Version-only specifiers (not full paths). In the manylinux-cross
26
+ # container these resolve via sysconfig data files, which lets the
27
+ # aarch64 build read interpreter metadata without exec'ing the
28
+ # aarch64 binaries on the x86_64 host.
29
+ args: >-
30
+ --release --out dist
31
+ -i python3.10 python3.11 python3.12 python3.13 python3.14
48
32
  - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
49
33
  with:
50
34
  name: wheels-linux-${{ matrix.target }}
@@ -54,8 +38,8 @@ jobs:
54
38
  name: Build — Windows (x64)
55
39
  runs-on: windows-latest
56
40
  steps:
57
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
58
- - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
41
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
42
+ - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
59
43
  with:
60
44
  python-version: |
61
45
  3.10
@@ -76,8 +60,8 @@ jobs:
76
60
  name: Build — macOS (universal2)
77
61
  runs-on: macos-latest
78
62
  steps:
79
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
80
- - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
63
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
64
+ - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
81
65
  with:
82
66
  python-version: |
83
67
  3.10
@@ -98,7 +82,7 @@ jobs:
98
82
  name: Build — sdist
99
83
  runs-on: ubuntu-latest
100
84
  steps:
101
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
85
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
102
86
  - uses: PyO3/maturin-action@e83996d129638aa358a18fbd1dfb82f0b0fb5d3b # v1
103
87
  with:
104
88
  command: sdist
@@ -116,12 +100,12 @@ jobs:
116
100
  permissions:
117
101
  id-token: write
118
102
  steps:
119
- - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
103
+ - uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
120
104
  with:
121
105
  pattern: wheels-*
122
106
  merge-multiple: true
123
107
  path: dist
124
- - uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b # release/v1
108
+ - uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # release/v1
125
109
  with:
126
110
  repository-url: https://test.pypi.org/legacy/
127
111
 
@@ -133,12 +117,12 @@ jobs:
133
117
  permissions:
134
118
  id-token: write
135
119
  steps:
136
- - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
120
+ - uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
137
121
  with:
138
122
  pattern: wheels-*
139
123
  merge-multiple: true
140
124
  path: dist
141
- - uses: pypa/gh-action-pypi-publish@cef221092ed1bacb1cc03d23a2d87d1d172e277b # release/v1
125
+ - uses: pypa/gh-action-pypi-publish@dc37677b2e1c63e2034f94d8a5b11f265b73ba33 # release/v1
142
126
 
143
127
  github-release:
144
128
  name: GitHub Release
@@ -147,12 +131,12 @@ jobs:
147
131
  permissions:
148
132
  contents: write
149
133
  steps:
150
- - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
134
+ - uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
151
135
  with:
152
136
  pattern: wheels-*
153
137
  merge-multiple: true
154
138
  path: dist
155
- - uses: softprops/action-gh-release@3bb12739c298aeb8a4eeaf626c5b8d85266b0e65 # v2
139
+ - uses: softprops/action-gh-release@3d0d9888cb7fd7b750713d6e236d1fcb99157228 # v3.0.2
156
140
  with:
157
141
  files: dist/*
158
142
  generate_release_notes: true
@@ -16,16 +16,16 @@ jobs:
16
16
  runs-on: ubuntu-latest
17
17
  strategy:
18
18
  matrix:
19
- python-version: ["3.10", "3.11", "3.12", "3.13"]
19
+ python-version: ["3.11", "3.12", "3.13", "3.14"]
20
20
 
21
21
  steps:
22
- - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
22
+ - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
23
23
 
24
24
  - name: Set up Rust
25
25
  uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
26
26
 
27
27
  - name: Set up Python ${{ matrix.python-version }}
28
- uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
28
+ uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
29
29
  with:
30
30
  python-version: ${{ matrix.python-version }}
31
31
 
@@ -196,3 +196,5 @@ Cargo.lock
196
196
 
197
197
  ## benchmarks
198
198
  bench*.json
199
+ *json.gz
200
+ baselines/
@@ -1,11 +1,12 @@
1
1
  {
2
2
  "cSpell.words": [
3
3
  "Bilz",
4
+ "cooldown",
4
5
  "Damerau",
5
6
  "duftify",
6
7
  "geht",
7
- "Hyyrö's",
8
8
  "Hyyrö",
9
+ "Hyyrö's",
9
10
  "labelrotation",
10
11
  "manylinux",
11
12
  "Mathers",
@@ -17,6 +18,7 @@
17
18
  "pytest",
18
19
  "rapidfuzz",
19
20
  "résumé",
21
+ "standalone",
20
22
  "woulda",
21
23
  "xticklabels",
22
24
  "xticks",
@@ -9,17 +9,46 @@ prefer Rust-side optimizations over Python wrappers.
9
9
  ## Environment Setup
10
10
  - Python >= 3.10
11
11
  - Install: `uv sync --all-extras`
12
- - After editing Rust, rebuild the extension: `bash scripts/install_dev.sh`
12
+ - After editing Rust, rebuild the extension: `uv run maturin develop --release`
13
+
14
+ ## Performance workflow
15
+
16
+ All performance changes follow this loop. One change per experiment,
17
+ measured on the same hardware, kept only if it wins.
18
+
19
+ 1. Pin a baseline before any change:
20
+ - Python-level:
21
+ `uv run scripts/benchmark.py --save baselines/<name>.json`
22
+ Add `--kind ascii` (or `latin1`, `cjk`, `emoji`, `realistic`) to baseline a
23
+ single string kind instead of all five.
24
+ - Rust-level (if criterion benches exist for the change surface):
25
+ `cargo bench -- --save-baseline <name>`
26
+ 2. State a single, testable hypothesis before editing (e.g. "prefix/suffix
27
+ trim removes ~30% of DP work on ASCII-100"). One change per experiment.
28
+ 3. Re-run the same benchmark on identical hardware, quiet system, `taskset`
29
+ or `sudo nice -n -20` where reasonable. Report median + IQR, not mean.
30
+ 4. If the change is under 3% or within noise on every kind, REVERT.
31
+ Do not commit tuning that only wins in one microbenchmark.
32
+ 5. Verify correctness against the property tests and a diff test vs.
33
+ rapidfuzz / edlib on random inputs before keeping any win.
34
+ 6. Check codegen for the hot loop with `cargo asm` (or `--emit=asm`) when
35
+ claiming SIMD or auto-vectorization wins.
36
+ 7. Never edit the release profile in `Cargo.toml` (LTO, codegen-units,
37
+ panic) as part of an algorithmic experiment. Separate PR.
38
+
39
+ Profiling is a read-only step and is delegated to the `profile-runner`
40
+ sub-agent (`.claude/agents/profile-runner.md`). It uses samply for the
41
+ Python-level benchmark and cargo flamegraph for the criterion benches;
42
+ neither needs `--profile-time`.
13
43
 
14
44
  ## Testing
15
45
  - Python: `pytest tests/`
16
46
  - Rust: `cargo test`
17
- - Rust changes require a rebuild before pytest will pick them up. You can trigger a rebuild with `bash scripts/install_dev.sh`.
47
+ - Rust changes require a rebuild before pytest will pick them up. You can trigger a rebuild with `uv run maturin develop --release`.
18
48
 
19
49
  ## Benchmarking
20
- - Install deps: `uv sync --extra benchmark`
21
- - Compare against other libraries: `uv run python scripts/benchmark.py`
22
- - Compare across string lengths: `uv run python scripts/benchmark_by_length.py`
50
+ - Compare against other libraries: `uv run scripts/benchmark.py`
51
+ - Compare across string lengths: `uv run scripts/benchmark_by_length.py`
23
52
  - For performance-critical changes, run `benchmark.py` before and after,
24
53
  then diff `benchmark_results.json`.
25
54
 
@@ -1,3 +1,20 @@
1
+ ## 0.2.4 (2026-08-25)
2
+
3
+ ### Perf
4
+
5
+ - improve performance on long cjk and emoji strings🚀 (#136)
6
+
7
+ ## 0.2.3 (2026-07-09)
8
+
9
+ ### Fix
10
+
11
+ - correct typo in project.description🐞 (#106)
12
+
13
+ ### Perf
14
+
15
+ - improve performance on exceptionally long strings🚀 (#131)
16
+ - improve speed for >= python 3.14🚀 (#129)
17
+
1
18
  ## 0.2.2 (2026-05-05)
2
19
 
3
20
  ### Perf
@@ -10,7 +10,7 @@ checksum = "2304e00983f87ffb38b55b444b5e3b60a884b5d30c0fca7d82fe33449bbe55ea"
10
10
 
11
11
  [[package]]
12
12
  name = "lev"
13
- version = "0.2.2"
13
+ version = "0.2.4"
14
14
  dependencies = [
15
15
  "pyo3",
16
16
  "pyo3-build-config",
@@ -18,9 +18,9 @@ dependencies = [
18
18
 
19
19
  [[package]]
20
20
  name = "libc"
21
- version = "0.2.186"
21
+ version = "0.2.189"
22
22
  source = "registry+https://github.com/rust-lang/crates.io-index"
23
- checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
23
+ checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
24
24
 
25
25
  [[package]]
26
26
  name = "once_cell"
@@ -30,15 +30,15 @@ checksum = "9f7c3e4beb33f85d45ae3e3a1792185706c8e16d043238c593331cc7cd313b50"
30
30
 
31
31
  [[package]]
32
32
  name = "portable-atomic"
33
- version = "1.13.1"
33
+ version = "1.15.0"
34
34
  source = "registry+https://github.com/rust-lang/crates.io-index"
35
- checksum = "c33a9471896f1c69cecef8d20cbe2f7accd12527ce60845ff44c153bb2a21b49"
35
+ checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"
36
36
 
37
37
  [[package]]
38
38
  name = "proc-macro2"
39
- version = "1.0.106"
39
+ version = "1.0.107"
40
40
  source = "registry+https://github.com/rust-lang/crates.io-index"
41
- checksum = "8fd00f0bb2e90d81d1044c2b32617f68fcb9fa3bb7640c23e9c748e53fb30934"
41
+ checksum = "985e7ec9bb745e6ce6535b544d84d6cd6f7ad8bd711c398938ae983b91a766d9"
42
42
  dependencies = [
43
43
  "unicode-ident",
44
44
  ]
@@ -103,18 +103,18 @@ dependencies = [
103
103
 
104
104
  [[package]]
105
105
  name = "quote"
106
- version = "1.0.45"
106
+ version = "1.0.47"
107
107
  source = "registry+https://github.com/rust-lang/crates.io-index"
108
- checksum = "41f2619966050689382d2b44f664f4bc593e129785a36d6ee376ddf37259b924"
108
+ checksum = "1fbf4db142a473a8d80c26bbf18454ed458bf8d26c8219c331daecfdbd079001"
109
109
  dependencies = [
110
110
  "proc-macro2",
111
111
  ]
112
112
 
113
113
  [[package]]
114
114
  name = "syn"
115
- version = "2.0.117"
115
+ version = "2.0.119"
116
116
  source = "registry+https://github.com/rust-lang/crates.io-index"
117
- checksum = "e665b8803e7b1d2a727f4023456bbbbe74da67099c585258af0ad9c5013b9b99"
117
+ checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
118
118
  dependencies = [
119
119
  "proc-macro2",
120
120
  "quote",
@@ -1,7 +1,7 @@
1
1
 
2
2
  [package]
3
3
  name = "lev"
4
- version = "0.2.2"
4
+ version = "0.2.4"
5
5
  edition = "2021"
6
6
  readme = "README.md"
7
7
 
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: lev-rs
3
- Version: 0.2.2
3
+ Version: 0.2.4
4
4
  Classifier: Development Status :: 3 - Alpha
5
5
  Classifier: Programming Language :: Python :: 3 :: Only
6
6
  Classifier: Programming Language :: Python :: 3.10
@@ -13,12 +13,6 @@ Classifier: Intended Audience :: Information Technology
13
13
  Classifier: Intended Audience :: Science/Research
14
14
  Classifier: Topic :: Scientific/Engineering
15
15
  Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
16
- Requires-Dist: edlib ; extra == 'benchmark'
17
- Requires-Dist: editdistance>=0.8.1 ; extra == 'benchmark'
18
- Requires-Dist: matplotlib>=3.10.9 ; extra == 'benchmark'
19
- Requires-Dist: matplotx>=0.3.10 ; extra == 'benchmark'
20
- Requires-Dist: numpy>=2.0 ; extra == 'benchmark'
21
- Requires-Dist: rapidfuzz>=3.14.5 ; extra == 'benchmark'
22
16
  Requires-Dist: commitizen ; extra == 'dev'
23
17
  Requires-Dist: maturin ; extra == 'dev'
24
18
  Requires-Dist: prek ; extra == 'dev'
@@ -28,11 +22,10 @@ Requires-Dist: pytest ; extra == 'dev'
28
22
  Requires-Dist: ruff ; extra == 'dev'
29
23
  Requires-Dist: zensical ; extra == 'docs'
30
24
  Requires-Dist: mkdocstrings[python] ; extra == 'docs'
31
- Provides-Extra: benchmark
32
25
  Provides-Extra: dev
33
26
  Provides-Extra: docs
34
27
  License-File: LICENSE
35
- Summary: A extremely fast Python library for the Levenshtein distance, written in Rust 🦀.
28
+ Summary: An extremely fast Python library for the Levenshtein distance and similarity ratio, written in Rust 🦀.
36
29
  Keywords: distance,levenshtein,strings
37
30
  Author-email: Markus Bilz <github@markusbilz.com>
38
31
  Requires-Python: >=3.10
@@ -79,7 +72,7 @@ For more details on the API see the [API Reference](https://karelze.github.io/le
79
72
  [edlib](https://github.com/Martinsos/edlib). We excluded slower implementations like
80
73
  [pylev](https://github.com/toastdriven/pylev/tree/main) and [python-Levenshtein](https://github.com/ztane/python-Levenshtein).
81
74
 
82
- > Benchmarks were run on an Apple Mac Mini M2 Pro (macOS 26.2) using Python 3.13. Each string pair is exactly 100 characters long. Results represent the total wall time for 1,000 repetitions using Python's `timeit`. To reproduce, run [`uv run python scripts/benchmark.py`](https://github.com/KarelZe/lev/blob/main/scripts/benchmark.py).
75
+ > Benchmarks were run on an Apple Mac Mini M2 Pro (macOS 26.2) using Python 3.13. The ASCII, Latin-1, CJK, and Emoji pairs are exactly 100 characters long; the realistic-text pair is natural-length prose. Results represent the total wall time for 1,000 repetitions using Python's `timeit`. To reproduce, run [`uv run scripts/benchmark.py`](https://github.com/KarelZe/lev/blob/main/scripts/benchmark.py).
83
76
 
84
77
  ### ASCII
85
78
 
@@ -103,6 +96,12 @@ For more details on the API see the [API Reference](https://karelze.github.io/le
103
96
 
104
97
  ![Emoji benchmark – dark](https://raw.githubusercontent.com/KarelZe/lev/main/docs/assets/benchmark_emoji_dark.svg)
105
98
 
99
+ ### Realistic Text
100
+
101
+ This pair is a short customer support message with a handful of natural typos.
102
+
103
+ ![Realistic text benchmark – dark](https://raw.githubusercontent.com/KarelZe/lev/main/docs/assets/benchmark_realistic_dark.svg)
104
+
106
105
  ## Contact
107
106
 
108
107
  To get in contact, please open an issue or contact me via `github@markusbilz.com`.
@@ -37,7 +37,7 @@ For more details on the API see the [API Reference](https://karelze.github.io/le
37
37
  [edlib](https://github.com/Martinsos/edlib). We excluded slower implementations like
38
38
  [pylev](https://github.com/toastdriven/pylev/tree/main) and [python-Levenshtein](https://github.com/ztane/python-Levenshtein).
39
39
 
40
- > Benchmarks were run on an Apple Mac Mini M2 Pro (macOS 26.2) using Python 3.13. Each string pair is exactly 100 characters long. Results represent the total wall time for 1,000 repetitions using Python's `timeit`. To reproduce, run [`uv run python scripts/benchmark.py`](https://github.com/KarelZe/lev/blob/main/scripts/benchmark.py).
40
+ > Benchmarks were run on an Apple Mac Mini M2 Pro (macOS 26.2) using Python 3.13. The ASCII, Latin-1, CJK, and Emoji pairs are exactly 100 characters long; the realistic-text pair is natural-length prose. Results represent the total wall time for 1,000 repetitions using Python's `timeit`. To reproduce, run [`uv run scripts/benchmark.py`](https://github.com/KarelZe/lev/blob/main/scripts/benchmark.py).
41
41
 
42
42
  ### ASCII
43
43
 
@@ -61,6 +61,12 @@ For more details on the API see the [API Reference](https://karelze.github.io/le
61
61
 
62
62
  ![Emoji benchmark – dark](https://raw.githubusercontent.com/KarelZe/lev/main/docs/assets/benchmark_emoji_dark.svg)
63
63
 
64
+ ### Realistic Text
65
+
66
+ This pair is a short customer support message with a handful of natural typos.
67
+
68
+ ![Realistic text benchmark – dark](https://raw.githubusercontent.com/KarelZe/lev/main/docs/assets/benchmark_realistic_dark.svg)
69
+
64
70
  ## Contact
65
71
 
66
72
  To get in contact, please open an issue or contact me via `github@markusbilz.com`.