new v0.9.0 release

.agents/MIRI_PLAN.md

@@ -150,14 +150,18 @@ algorithms (RANSAC, ORB, Lucas-Kanade), and **none of them contain or reach
 | `features2d::tests::test_orb_pyramid_dimensions` | 30.2s | `src/features2d/tests.rs` |
 
 **Why so few exclusions suffice.** The distribution is extremely skewed: the top 5
-tests are 82% of total runtime, while the remaining **293 tests complete in 139s
-combined**. Nine annotations take the suite from 41 minutes to roughly 4.
+tests are 82% of total runtime, while the remaining **341 tests complete in well
+under two minutes combined**. Nine annotations take the suite from 41 minutes to roughly 4.
 
 **Coverage check on the one borderline case.** `test_build_pyramid_with_derivatives`
-exercises the Sobel path, which under `simd` reaches the `unsafe` in
-`derivatives.rs:346,349`. That coverage is **not** lost: `imgproc::tests::test_sobel`
-calls `sobel(&src_f32, 1, 0, 3, …)` (`src/imgproc/tests.rs:138`), matching
-`fast_deriv_3x3`'s `TypeId == f32 && ksize == 3` trigger, and runs in 0.8s under Miri.
+exercises the Scharr path (Sobel before #130, switched to Scharr by that fix), which
+under `simd` reaches the `unsafe` in `derivatives.rs:346,349`. That coverage is
+**not** lost: `imgproc::tests::test_scharr` calls `scharr(&src, 1, 0, …)`
+(`src/imgproc/tests.rs:154`), matching `fast_deriv_3x3`'s `TypeId == f32 &&
+kernel-length == 3` trigger (true for both Sobel `ksize=3` and Scharr `ksize=-1`,
+since `get_deriv_kernel` returns a 3-element kernel for both), and runs in ~0.75s
+under Miri — same order of magnitude as `test_sobel`, which still covers the
+Sobel path used elsewhere (`canny`, `hough_circles`, …).
 
 ### Annotation convention
 
@@ -483,7 +487,7 @@ This turns a claim Miri would contradict into one Miri actively backs.
 | # | Decision | Alternatives considered | Rationale |
 |---|----------|-------------------------|-----------|
 | 10 | Set `MIRIFLAGS: -Zmiri-deterministic-floats` — **revises #4** | `#[cfg_attr(miri, ignore)]` on `test_randn_determinism`; leave the failure | #4 said "no flags", but measurement found a genuine need. The flag keeps the RNG determinism contract under test instead of disabling it, and targets a documented Miri behaviour rather than a real defect |
-| 11 | Exclude the 9 tests over 30s — **implements #2** | Exclude >10s (15 tests); shrink inputs under `cfg(miri)`; exclude nothing and raise the timeout | Measurement showed an extreme skew: 5 tests = 82% of runtime, 293 tests = 139s. Nine annotations buy a 10× speedup; none of the nine touch `unsafe`, so no UB coverage is lost |
+| 11 | Exclude the 9 tests over 30s — **implements #2** | Exclude >10s (15 tests); shrink inputs under `cfg(miri)`; exclude nothing and raise the timeout | Measurement showed an extreme skew: 5 tests = 82% of runtime, the rest (341 tests as of #131) well under two minutes combined. Nine annotations buy a 10× speedup; none of the nine touch `unsafe`, so no UB coverage is lost |
 
 ---
 
@@ -493,5 +497,5 @@ This turns a claim Miri would contradict into one Miri actively backs.
 |-----------------------|--------------|
 | Miri CI job passes on `dev` | ✅ Both legs green locally (§9). Pending confirmation on `ubuntu-latest`. |
 | All existing unsafe verified UB-free, or documented exceptions | ✅ 6 of 6 reachable production blocks verified clean. Exception: the `parallel`-only pair, documented in §5. |
-| Incompatible tests annotated `#[cfg_attr(miri, ignore)]` | ✅ 9 tests, each with a reason comment (§4). Excluded for runtime, not incompatibility — nothing in the suite proved Miri-incompatible. |
+| Incompatible tests annotated `#[cfg_attr(miri, ignore)]` | ✅ 9 tests, each with a reason comment (§4). Excluded for runtime, not incompatibility — nothing in the suite proved Miri-incompatible. #130 and #138 together added three more tests exercising the same unsafe Scharr fast path, but all three measured well under the 30s threshold (§4) and are not excluded. |
 | Plan document identifying included/excluded code with rationale | ✅ This document, tracked in git via a `.gitignore` exception. |

.github/workflows/release.yml

@@ -108,18 +108,38 @@ jobs:
       - name: Publish to Crates.io
         run: cargo publish --token ${{ secrets.CRATES_TOKEN }} -p purecv
 
+  # Publishes @webarkit/purecv-wasm via npm Trusted Publishing (OIDC), not a
+  # long-lived NPM_TOKEN (#129). Requires a Trusted Publisher configured on
+  # npmjs.com for this package: owner `webarkit`, repository `purecv`,
+  # workflow `release.yml`, no environment.
   publish-npm:
     needs: [release]
     runs-on: ubuntu-latest
+    # Job-level permissions replace the workflow-level block for this job:
+    # checkout only needs to read, and the OIDC token is what npm exchanges
+    # for a short-lived publish credential (and signs provenance with).
+    permissions:
+      contents: read
+      id-token: write
     steps:
       - uses: actions/checkout@v6
       - name: Install Rust
         uses: dtolnay/rust-toolchain@1.98.0
       - name: Install wasm-pack
         run: curl https://rustwasm.github.io/wasm-pack/installer/init.sh -sSf | sh
+      - name: Set up Node
+        uses: actions/setup-node@v7
+        with:
+          node-version: '24'
+      # Trusted publishing needs npm >= 11.5.1, and Node 24's bundled npm is
+      # not guaranteed to be new enough (24.18.0 ships npm 10.9.4, see #129).
+      - name: Ensure npm >= 11.5.1
+        run: |
+          npm install -g npm@11
+          npm --version
+          node -e 'const [a,b,c]=process.argv[1].split(".").map(Number); if (a<11 || (a===11 && (b<5 || (b===5 && c<1)))) { console.error(`npm ${process.argv[1]} < 11.5.1`); process.exit(1); }' "$(npm --version)"
       - name: Build and Publish to NPM
         run: |
           npm run build:wasm
           cd crates/wasm/pkg
-          npm config set //registry.npmjs.org/:_authToken ${{ secrets.NPM_TOKEN }}
-          npm publish --access public
+          npm publish --access public --provenance

CHANGELOG.md

@@ -2,6 +2,54 @@
 
 All notable changes to this project will be documented in this file.
 
+## [0.9.0] - 2026-09-27
+
+### ⚙️ Miscellaneous Tasks
+
+- *(video)* Fix formatting/lint from oscillation half-step change (#131)
+
+### 🐛 Bug Fixes
+
+- *(video)* Use Scharr derivatives in build_optical_flow_pyramid (#130)
+- *(video)* Use Scharr derivatives in calc_optical_flow_pyramid_lk (#130)
+- *(video)* Match OpenCV LK eigenvalue scale
+- *(video)* [**breaking**] Correct LK eigenvalue divisor and pin OpenCV's scale
+- *(video)* Add OpenCV's oscillation half-step fallback to LK (#131)
+- *(video)* Scale the LK determinant guard to match OpenCV (#142)
+- *(video)* Use OpenCV's signed determinant test in the LK guard (#142)
+- *(video)* Only lose an LK point when level 0 is degenerate (#145)
+- *(video)* Sample exactly win_size pixels in the LK window, like OpenCV (#144)
+- *(video)* Scale LK's temporal difference to the Scharr gain (#149)
+- *(ci)* Publish to npm via trusted publishing (OIDC) instead of NPM_TOKEN (#129)
+
+### 📚 Documentation
+
+- *(video)* Correct stale Sobel references after Scharr switch (#130)
+- Update stale unit test count in README (342 -> 348)
+- *(video)* Address Copilot review findings on PR #137
+- *(video)* Correct the LK threshold migration factor
+- *(video)* Fix dangling trace reference and comment alignment (#131)
+- Fix cross-fix staleness found in pre-merge review (#130, #131, #138)
+- Address final-review findings for #142 (README count, doc polish)
+
+### 🚜 Refactor
+
+- *(video)* Extract lk_iterate from lk_single_level (#131)
+
+### 🧪 Testing
+
+- *(video)* Pin build_optical_flow_pyramid to Scharr derivatives (#130)
+- *(video)* Pin calc_optical_flow_pyramid_lk to Scharr derivatives (#130)
+- *(video)* Use relative tolerance for f32 min-eigen comparison (#130)
+- *(video)* Pin lk_iterate's oscillation half-step fallback (#131)
+- *(video)* Pin near-degenerate LK window rejection to OpenCV's scale (#142)
+- *(video)* Correct effective threshold value in test doc comment (#142)
+- *(video)* Pin the LK determinant guard from the accept side (#142)
+- *(video)* LK must survive a degenerate coarse pyramid level (#145)
+- *(video)* Make the #145 propagated-flow test step-size independent
+- *(video)* Pin LK window sampling for even and non-square win_size (#144)
+- *(video)* Pin the LK Newton step to the full Gauss-Newton step (#149)
+
 ## [0.8.0] - 2026-09-02
 
 ### ⚙️ Miscellaneous Tasks

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "purecv"
-version = "0.8.0"
+version = "0.9.0"
 authors = ["Walter Perdan <https://github.com/kalwalt>"]
 edition = "2021"
 rust-version = "1.88"
@@ -86,7 +86,7 @@ members = ["crates/wasm"]
 exclude = ["crates/no-std-smoke"]
 
 [workspace.package]
-version = "0.8.0"
+version = "0.9.0"
 authors = ["Walter Perdan <https://github.com/kalwalt>"]
 edition = "2021"
 description = "A pure Rust, high-performance computer vision library focused on safety and portability."

README.md

@@ -87,7 +87,7 @@ Add the following to your `Cargo.toml`:
 
 ```toml
 [dependencies]
-purecv = "0.8"
+purecv = "0.9"
 ```
 
 PureCV's minimum supported Rust version (MSRV) is **1.88**.
@@ -105,7 +105,7 @@ PureCV's minimum supported Rust version (MSRV) is **1.88**.
 ### `no_std` / embedded support
 
 Build with `--no-default-features` to run on bare-metal targets such as the
-ESP32 (`purecv = { version = "0.8", default-features = false }`). Only `core`
+ESP32 (`purecv = { version = "0.9", default-features = false }`). Only `core`
 and `alloc` are required (an allocator must be provided by the target).
 
 | Module | `no_std` | Notes |
@@ -121,7 +121,7 @@ features gives the scalar, single-threaded code paths.
 
 ```toml
 [dependencies]
-purecv = { version = "0.8", default-features = false }
+purecv = { version = "0.9", default-features = false }
 ```
 
 ```rust
@@ -150,14 +150,14 @@ To enable the `ndarray` feature:
 
 ```toml
 [dependencies]
-purecv = { version = "0.8", features = ["ndarray"] }
+purecv = { version = "0.9", features = ["ndarray"] }
 ```
 
 To enable SIMD + Parallel for maximum performance:
 
 ```toml
 [dependencies]
-purecv = { version = "0.8", features = ["parallel", "simd"] }
+purecv = { version = "0.9", features = ["parallel", "simd"] }
 ```
 
 ### Usage Example
@@ -348,7 +348,7 @@ cargo run --example rectification
 ## 🧪 Testing & Benchmarking
 
 ### Running Tests
-PureCV uses a comprehensive suite of unit tests to ensure correctness and parity with OpenCV. The test suite currently includes **342 unit tests** (plus **40 doc-tests**) covering:
+PureCV uses a comprehensive suite of unit tests to ensure correctness and parity with OpenCV. The test suite currently includes **359 unit tests** (plus **40 doc-tests**) covering:
 
 - **Core module:** Matrix factories, scalar arithmetic variants, bitwise scalar ops, min/max, comparison ops (`compare`, `in_range`), reduction (`reduce`, `count_non_zero`), polar/cartesian conversions, linear algebra (`determinant`, `invert`, `solve`), channel ops (`extract_channel`, `insert_channel`), `DynamicMatrix`, transforms, sorting, clustering, and RNG.
 - **Imgproc module:** Filters, derivatives, edge detection, color conversions (including gray-to-RGB/BGR/RGBA/BGRA), thresholding, morphology (`erode`, `dilate`), pyramids (`pyr_down`, `pyr_up`), kernel helpers (`get_gaussian_kernel`, `get_sobel_kernels`), and histograms/CLAHE (`calc_hist`, `calc_back_project`, `compare_hist`, `equalize_hist`, `Clahe`).

benches/benchmark_results.md

@@ -106,10 +106,13 @@ Parallel + SIMD achieves the best throughput.
 
 #### `build_optical_flow_pyramid` with derivatives
 
-When `with_derivatives = true`, the per-level Sobel (Ix, Iy) passes are
-independent across pyramid levels and run concurrently via Rayon with the
-`parallel` feature.  For a 4-level pyramid the expected speedup is up to
-4× (level count) times the per-level Sobel speedup.  In practice the
+When `with_derivatives = true`, the per-level Scharr (Ix, Iy) passes
+(Sobel before #130, switched to Scharr to match OpenCV) are independent
+across pyramid levels and run concurrently via Rayon with the `parallel`
+feature.  Scharr shares the same `fast_deriv_3x3` code path as Sobel, so
+the performance characteristics below are unchanged by that switch.  For
+a 4-level pyramid the expected speedup is up to 4× (level count) times
+the per-level Scharr speedup.  In practice the
 coarser levels are very small so the scaling is sub-linear, but a 2–3×
 wall-clock gain is typical.
 

benches/video_bench.rs

@@ -111,7 +111,7 @@ fn make_grid_points(size: usize, step: usize) -> Vec<Point2f> {
 ///
 /// Benchmark names:
 /// * `build_optical_flow_pyramid/no_deriv` — pyramid only
-/// * `build_optical_flow_pyramid/with_deriv` — pyramid + Sobel Ix, Iy per level
+/// * `build_optical_flow_pyramid/with_deriv` — pyramid + Scharr Ix, Iy per level
 ///   (this is where the `parallel` feature gives the most gain in this function)
 fn bench_build_pyramid(c: &mut Criterion) {
     let size = 512usize;

crates/wasm/pkg/package.json

@@ -5,7 +5,7 @@
                           "Walter Perdan \u003chttps://github.com/kalwalt\u003e"
                       ],
     "description":  "A pure Rust, high-performance computer vision library focused on safety and portability.",
-    "version":  "0.8.0",
+    "version":  "0.9.0",
     "license":  "LGPL-2.1-or-later",
     "repository":  {
                        "type":  "git",

crates/wasm/src/lib.rs

@@ -1966,7 +1966,7 @@ pub fn morph_blackhat() -> i32 {
 ///
 /// * `win_w`, `win_h` – Tracking window size.
 /// * `max_level`      – Maximum number of additional pyramid levels.
-/// * `with_derivatives`– Compute Sobel derivatives alongside the pyramid.
+/// * `with_derivatives`– Compute Scharr derivatives alongside the pyramid.
 /// * `pyr_border`     – Border interpolation for downsampling.
 /// * `deriv_border`   – Border interpolation for derivatives.
 ///
@@ -2038,7 +2038,15 @@ pub fn wasm_build_optical_flow_pyramid(
 /// * `epsilon`        – Convergence threshold.
 /// * `flags`          – Combine `OPTFLOW_USE_INITIAL_FLOW()` and/or
 ///   `OPTFLOW_LK_GET_MIN_EIGENVALS()`.
-/// * `min_eigen_threshold` – Min eigenvalue below which a point is lost.
+/// * `min_eigen_threshold` – Min eigenvalue below which a point is lost. The
+///   gradient matrix is built from Scharr derivatives and normalized by
+///   `FLT_SCALE = 2^-20` and the window area, matching
+///   `cv::calcOpticalFlowPyrLK` (see #130 and #138), so this value is on the
+///   same scale as OpenCV's and its `1e-4` default transfers directly.
+///   Thresholds tuned against a purecv build predating both fixes must be
+///   retuned: Scharr multiplies the gradient matrix by 16 relative to the
+///   Sobel derivatives used then, and `FLT_SCALE` divides by `2^20`, so such
+///   thresholds read about `2^20 / 16 = 65536` times larger on this scale.
 ///
 /// ```js
 /// const gray0 = Mat.fromU8Data(h, w, 1, frameData0);

examples/optical_flow.rs

@@ -429,7 +429,7 @@ fn draw_line(
 /// The output has one header row followed by one data row per point:
 /// ```text
 /// idx,prev_x,prev_y,next_x,next_y,flow_x,flow_y,status,min_eigen
-/// 0,158.000,124.000,162.010,127.000,4.010,3.000,1,88873.757813
+/// 0,158.000,124.000,161.922,126.883,3.922,2.883,1,1.470835
 /// …
 /// ```
 ///

package.json

@@ -1,6 +1,6 @@
 {
   "name": "purecv",
-  "version": "0.8.0",
+  "version": "0.9.0",
   "description": "A pure Rust, high-performance computer vision library focused on safety and portability.",
   "private": true,
   "scripts": {