chore(deps): bump actions/github-script from 8 to 9

Bumps [actions/github-script](https://github.com/actions/github-script) from 8 to 9.
- [Release notes](https://github.com/actions/github-script/releases)
- [Commits](https://github.com/actions/github-script/compare/v8...v9)

---
updated-dependencies:
- dependency-name: actions/github-script
  dependency-version: '9'
  dependency-type: direct:production
  update-type: version-update:semver-major
...

Signed-off-by: dependabot[bot] <support@github.com>

.github/dependabot.yml

@@ -34,3 +34,14 @@ updates:
     labels:
       - "dependencies"
       - "github-actions"
+
+  # WASM JavaScript package
+  - package-ecosystem: "npm"
+    directory: "/wasm"
+    schedule:
+      interval: "weekly"
+      day: "monday"
+    open-pull-requests-limit: 5
+    labels:
+      - "dependencies"
+      - "javascript"

.github/workflows/CI.yml

@@ -87,7 +87,7 @@ jobs:
     steps:
       - name: Deploy to GitHub Pages
         id: deployment
-        uses: actions/deploy-pages@v4
+        uses: actions/deploy-pages@v5
 
   # -------------------------------------------------------------------------
   # CI gate — all required jobs must pass before this job succeeds.
@@ -136,88 +136,123 @@ jobs:
   # Keep release jobs here because trusted-publisher configuration points to
   # CI.yml specifically.
   # -------------------------------------------------------------------------
+  # One abi3 wheel per (platform, arch). abi3-py310 means a single wheel
+  # covers CPython 3.10+ on each target, so there is no python-version axis.
+  # extension-module + abi3 link no libpython, which is what lets the linux
+  # jobs cross-compile aarch64/musl from an x86_64 runner via maturin-action.
+
   build-wheels-linux:
-    name: Build wheels (linux / py${{ matrix.python-version }})
+    name: Build wheels (linux-gnu / ${{ matrix.target }})
     runs-on: ubuntu-latest
     if: (github.event_name == 'release' && github.event.action == 'published') || (github.event_name == 'workflow_dispatch' && inputs.release)
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13"]
+        target: [x86_64, aarch64]
 
     steps:
       - uses: actions/checkout@v6
 
-      - name: Build wheel
+      - name: Build abi3 manylinux wheel
         uses: PyO3/maturin-action@v1
         with:
           command: build
-          args: --release --out dist --compatibility pypi -i python${{ matrix.python-version }}
+          target: ${{ matrix.target }}
+          args: --release --out dist
           manylinux: "2_17"
 
       - name: Upload wheels as artifact
         uses: actions/upload-artifact@v7
         with:
-          name: wheels-linux-py${{ matrix.python-version }}
+          name: wheels-linux-gnu-${{ matrix.target }}
+          path: dist/*.whl
+
+  build-wheels-musllinux:
+    name: Build wheels (linux-musl / ${{ matrix.target }})
+    runs-on: ubuntu-latest
+    if: (github.event_name == 'release' && github.event.action == 'published') || (github.event_name == 'workflow_dispatch' && inputs.release)
+    strategy:
+      fail-fast: false
+      matrix:
+        target: [x86_64, aarch64]
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Build abi3 musllinux wheel
+        uses: PyO3/maturin-action@v1
+        with:
+          command: build
+          target: ${{ matrix.target }}
+          args: --release --out dist
+          manylinux: musllinux_1_2
+
+      - name: Upload wheels as artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: wheels-linux-musl-${{ matrix.target }}
           path: dist/*.whl
 
   build-wheels-macos:
-    name: Build wheels (macos / py${{ matrix.python-version }})
+    name: Build wheels (macos / universal2)
     runs-on: macos-latest
     if: (github.event_name == 'release' && github.event.action == 'published') || (github.event_name == 'workflow_dispatch' && inputs.release)
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13"]
-
     steps:
       - uses: actions/checkout@v6
 
-      - name: Set up Python ${{ matrix.python-version }}
+      - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.10"
 
-      - name: Build universal2 wheel
+      - name: Build universal2 abi3 wheel
         uses: PyO3/maturin-action@v1
         with:
           command: build
-          args: --release --out dist --compatibility pypi -i python
           target: universal2-apple-darwin
+          args: --release --out dist
 
       - name: Upload wheels as artifact
         uses: actions/upload-artifact@v7
         with:
-          name: wheels-macos-py${{ matrix.python-version }}
+          name: wheels-macos-universal2
           path: dist/*.whl
 
   build-wheels-windows:
-    name: Build wheels (windows / py${{ matrix.python-version }})
+    name: Build wheels (windows / ${{ matrix.platform.arch }})
-    runs-on: windows-latest
+    runs-on: ${{ matrix.platform.runner }}
     if: (github.event_name == 'release' && github.event.action == 'published') || (github.event_name == 'workflow_dispatch' && inputs.release)
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.10", "3.11", "3.12", "3.13"]
+        platform:
+          - runner: windows-latest
+            arch: x64
+            target: x64
+          - runner: windows-11-arm
+            arch: arm64
+            target: aarch64
 
     steps:
       - uses: actions/checkout@v6
 
-      - name: Set up Python ${{ matrix.python-version }}
+      - name: Set up Python
         uses: actions/setup-python@v6
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.11"
+          architecture: ${{ matrix.platform.arch }}
 
-      - name: Build wheel
+      - name: Build abi3 wheel
         uses: PyO3/maturin-action@v1
         with:
           command: build
-          args: --release --out dist --compatibility pypi -i python
+          target: ${{ matrix.platform.target }}
+          args: --release --out dist
 
       - name: Upload wheels as artifact
         uses: actions/upload-artifact@v7
         with:
-          name: wheels-windows-py${{ matrix.python-version }}
+          name: wheels-windows-${{ matrix.platform.arch }}
           path: dist/*.whl
 
   build-sdist:
@@ -247,6 +282,7 @@ jobs:
     runs-on: ubuntu-latest
     needs:
       - build-wheels-linux
+      - build-wheels-musllinux
       - build-wheels-macos
       - build-wheels-windows
       - build-sdist
@@ -256,6 +292,8 @@ jobs:
       url: https://pypi.org/p/ferro-ta
     permissions:
       id-token: write
+      attestations: write
+      contents: read
     steps:
       - name: Download all wheels
         uses: actions/download-artifact@v8
@@ -270,6 +308,11 @@ jobs:
           name: sdist
           path: dist
 
+      - name: Generate SLSA build provenance attestations
+        uses: actions/attest-build-provenance@v2
+        with:
+          subject-path: "dist/*"
+
       - name: Verify distribution coverage
         run: |
           python3 - <<'PY'
@@ -283,19 +326,15 @@ jobs:
           for name in files:
               print(f" - {name}")
 
+          # One abi3 wheel (cp310-abi3) per platform/arch — covers CPython 3.10+.
           expected = [
-              "ferro_ta-*-cp310-cp310-manylinux*_x86_64.whl",
+              "ferro_ta-*-cp310-abi3-manylinux*_x86_64.whl",
-              "ferro_ta-*-cp311-cp311-manylinux*_x86_64.whl",
+              "ferro_ta-*-cp310-abi3-manylinux*_aarch64.whl",
-              "ferro_ta-*-cp312-cp312-manylinux*_x86_64.whl",
+              "ferro_ta-*-cp310-abi3-musllinux*_x86_64.whl",
-              "ferro_ta-*-cp313-cp313-manylinux*_x86_64.whl",
+              "ferro_ta-*-cp310-abi3-musllinux*_aarch64.whl",
-              "ferro_ta-*-cp310-cp310-win_amd64.whl",
+              "ferro_ta-*-cp310-abi3-macosx*_universal2.whl",
-              "ferro_ta-*-cp311-cp311-win_amd64.whl",
+              "ferro_ta-*-cp310-abi3-win_amd64.whl",
-              "ferro_ta-*-cp312-cp312-win_amd64.whl",
+              "ferro_ta-*-cp310-abi3-win_arm64.whl",
-              "ferro_ta-*-cp313-cp313-win_amd64.whl",
-              "ferro_ta-*-cp310-cp310-macosx*_universal2.whl",
-              "ferro_ta-*-cp311-cp311-macosx*_universal2.whl",
-              "ferro_ta-*-cp312-cp312-macosx*_universal2.whl",
-              "ferro_ta-*-cp313-cp313-macosx*_universal2.whl",
               "ferro_ta-*.tar.gz",
           ]
 
@@ -346,6 +385,7 @@ jobs:
     permissions:
       contents: write
       id-token: write
+      attestations: write
     steps:
       - uses: actions/checkout@v6
 
@@ -380,12 +420,40 @@ jobs:
       - name: Generate Rust SBOM (CycloneDX)
         run: cargo cyclonedx --format json --override-filename ferro-ta-rust-sbom.cdx
 
-      - name: Upload Python SBOM to release
+      - name: Validate SBOMs (schema check)
-        uses: softprops/action-gh-release@v2
+        run: |
+          python3 -c "import json, sys; json.load(open('ferro-ta-python-sbom.spdx.json')); print('Python SBOM valid JSON')"
+          python3 -c "import json, sys; data=json.load(open('ferro-ta-rust-sbom.cdx.json')); assert data.get('bomFormat')=='CycloneDX', 'not CycloneDX'; print('Rust SBOM valid CycloneDX')"
+
+      - name: Install cosign
+        uses: sigstore/cosign-installer@v3
+
+      - name: Sign SBOMs with cosign (keyless)
+        env:
+          COSIGN_EXPERIMENTAL: "1"
+        run: |
+          cosign sign-blob --yes ferro-ta-python-sbom.spdx.json --output-signature ferro-ta-python-sbom.spdx.json.sig --output-certificate ferro-ta-python-sbom.spdx.json.pem
+          cosign sign-blob --yes ferro-ta-rust-sbom.cdx.json     --output-signature ferro-ta-rust-sbom.cdx.json.sig     --output-certificate ferro-ta-rust-sbom.cdx.json.pem
+
+      - name: Attest SBOM provenance
+        uses: actions/attest-build-provenance@v2
         with:
-          files: ferro-ta-python-sbom.spdx.json
+          subject-path: |
+            ferro-ta-python-sbom.spdx.json
+            ferro-ta-rust-sbom.cdx.json
+
+      - name: Upload Python SBOM to release
+        uses: softprops/action-gh-release@v3
+        with:
+          files: |
+            ferro-ta-python-sbom.spdx.json
+            ferro-ta-python-sbom.spdx.json.sig
+            ferro-ta-python-sbom.spdx.json.pem
 
       - name: Upload Rust SBOM to release
-        uses: softprops/action-gh-release@v2
+        uses: softprops/action-gh-release@v3
         with:
-          files: ferro-ta-rust-sbom.cdx.json
+          files: |
+            ferro-ta-rust-sbom.cdx.json
+            ferro-ta-rust-sbom.cdx.json.sig
+            ferro-ta-rust-sbom.cdx.json.pem

.github/workflows/ci-docs.yml

@@ -44,6 +44,6 @@ jobs:
 
       - name: Upload GitHub Pages artifact
         if: ${{ inputs.upload-pages-artifact }}
-        uses: actions/upload-pages-artifact@v4
+        uses: actions/upload-pages-artifact@v5
         with:
           path: docs/_build/

.github/workflows/ci-python.yml

@@ -7,151 +7,157 @@ permissions:
   contents: read
 
 jobs:
+  # -------------------------------------------------------------------------
+  # Lint — no Rust, no wheel build, very fast
+  # -------------------------------------------------------------------------
   lint:
     name: Lint (ruff)
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-python@v6
-      - name: Set up Python 3.12
-        uses: actions/setup-python@v6
         with:
           python-version: "3.12"
+      - run: pip install uv
+      - run: uv run --with ruff ruff check python/ tests/
+      - run: uv run --with ruff ruff format --check python/ tests/
 
-      - name: Install uv
+  # -------------------------------------------------------------------------
-        run: pip install uv
+  # Type checking — needs wheel; build once with cache
-
+  # -------------------------------------------------------------------------
-      - name: Run ruff check via uv
-        run: uv run --with ruff ruff check python/ tests/
-      - name: Run ruff format check via uv
-        run: uv run --with ruff ruff format --check python/ tests/
-
   typecheck:
     name: Type checking (mypy + pyright)
     runs-on: ubuntu-latest
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-python@v6
-      - name: Set up Python 3.12
-        uses: actions/setup-python@v6
         with:
           python-version: "3.12"
+      - uses: dtolnay/rust-toolchain@v1
+        with:
+          toolchain: stable
+      - uses: Swatinem/rust-cache@v2
+      - run: pip install maturin numpy
+      - run: maturin build --release --out dist && pip install dist/*.whl
+      - run: pip install uv
+      - run: uv run --with mypy --with numpy python -m mypy python/ferro_ta --ignore-missing-imports --no-error-summary
+      - run: uv run --with pyright python -m pyright python/ferro_ta
 
-      - name: Install uv
+  # -------------------------------------------------------------------------
-        run: pip install uv
+  # Build wheel artifact (once per run, reused by all test matrix entries)
-
+  # -------------------------------------------------------------------------
-      - name: Run mypy on ferro_ta via uv
+  build-wheel:
-        run: uv run --with mypy --with numpy python -m mypy python/ferro_ta --ignore-missing-imports --no-error-summary
+    name: Build wheel (ubuntu / Python ${{ matrix.python-version }})
-
-      - name: Run pyright on ferro_ta via uv
-        run: uv run --with pyright python -m pyright python/ferro_ta
-
-  test:
-    name: Test (ubuntu-latest / Python ${{ matrix.python-version }})
     runs-on: ubuntu-latest
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
+    strategy:
+      matrix:
+        python-version: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-python@v6
+        with:
+          python-version: ${{ matrix.python-version }}
+      - uses: dtolnay/rust-toolchain@v1
+        with:
+          toolchain: stable
+      - uses: Swatinem/rust-cache@v2
+        with:
+          shared-key: "wheel-${{ matrix.python-version }}"
+      - run: pip install maturin numpy
+      - run: maturin build --release --out dist
+      - uses: actions/upload-artifact@v7
+        with:
+          name: wheel-${{ matrix.python-version }}
+          path: dist/*.whl
+
+  # -------------------------------------------------------------------------
+  # Test matrix — downloads pre-built wheel, no Rust compilation
+  # -------------------------------------------------------------------------
+  test:
+    name: Test (Python ${{ matrix.python-version }})
+    runs-on: ubuntu-latest
+    needs: build-wheel
     strategy:
       fail-fast: false
       matrix:
         python-version: ["3.10", "3.11", "3.12", "3.13"]
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-python@v6
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v6
         with:
           python-version: ${{ matrix.python-version }}
-
+      - uses: actions/download-artifact@v8
-      - name: Install maturin and test dependencies
+        with:
+          name: wheel-${{ matrix.python-version }}
+          path: dist
+      - name: Install wheel + test deps
         run: |
-          pip install maturin numpy pytest pytest-cov pandas polars hypothesis pyyaml mcp
-
-      - name: Build and install ferro_ta (dev mode)
-        run: |
-          maturin build --release --out dist
           pip install dist/*.whl
-
+          pip install pytest pytest-cov pandas polars hypothesis pyyaml mcp scipy
-      - name: Run unit tests with coverage
+      - name: Run tests with coverage
         run: pytest tests/unit/ tests/integration/ -v --cov=ferro_ta --cov-report=xml --cov-report=term-missing --cov-fail-under=65
-
       - name: Check API manifest is current
         if: matrix.python-version == '3.12'
         run: python scripts/check_api_manifest.py
-
+      - uses: actions/upload-artifact@v7
-      - name: Upload coverage report
-        uses: actions/upload-artifact@v7
         if: matrix.python-version == '3.12'
         with:
           name: coverage-python-${{ matrix.python-version }}
           path: coverage.xml
 
+  # -------------------------------------------------------------------------
+  # Benchmark vs TA-Lib (downloads wheel from build-wheel job)
+  # -------------------------------------------------------------------------
   benchmark-vs-talib:
     name: Benchmark vs TA-Lib
     runs-on: ubuntu-latest
+    needs: build-wheel
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-python@v6
-      - name: Set up Python 3.12
-        uses: actions/setup-python@v6
         with:
           python-version: "3.12"
-
+      - uses: actions/download-artifact@v8
-      - name: Install TA-Lib C library (Ubuntu)
+        with:
+          name: wheel-3.12
+          path: dist
+      - name: Install TA-Lib C library
         run: |
           sudo apt-get update
           sudo apt-get install -y build-essential curl
           curl -sL https://sourceforge.net/projects/ta-lib/files/ta-lib/0.4.0/ta-lib-0.4.0-src.tar.gz/download -o ta-lib-0.4.0-src.tar.gz
           tar -xzf ta-lib-0.4.0-src.tar.gz
-          cd ta-lib
+          cd ta-lib && ./configure --prefix=/usr && make && sudo make install && sudo ldconfig
-          ./configure --prefix=/usr
+      - run: pip install dist/*.whl numpy ta-lib
-          make
+      - run: python benchmarks/bench_vs_talib.py --sizes 10000 100000 --json benchmark_vs_talib.json
-          sudo make install
+      - run: python3 benchmarks/check_vs_talib_regression.py --input benchmark_vs_talib.json
-          sudo ldconfig
+      - uses: actions/upload-artifact@v7
-
-      - name: Install maturin and ta-lib Python package
-        run: |
-          pip install maturin numpy
-          pip install ta-lib
-
-      - name: Build and install ferro_ta
-        run: |
-          maturin build --release --out dist
-          pip install dist/*.whl
-
-      - name: Run benchmark comparison
-        run: |
-          python benchmarks/bench_vs_talib.py --sizes 10000 100000 --json benchmark_vs_talib.json
-
-      - name: Enforce benchmark regression policy
-        run: python3 benchmarks/check_vs_talib_regression.py --input benchmark_vs_talib.json
-
-      - name: Upload benchmark results
-        uses: actions/upload-artifact@v7
         with:
           name: benchmark-vs-talib
           path: benchmark_vs_talib.json
 
+  # -------------------------------------------------------------------------
+  # Performance smoke (downloads wheel from build-wheel job)
+  # -------------------------------------------------------------------------
   perf-smoke:
     name: Performance smoke and contracts
     runs-on: ubuntu-latest
+    needs: build-wheel
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-python@v6
-      - name: Set up Python 3.12
-        uses: actions/setup-python@v6
         with:
           python-version: "3.12"
-
+      - uses: actions/download-artifact@v8
-      - name: Install maturin and perf dependencies
+        with:
-        run: |
+          name: wheel-3.12
-          pip install maturin numpy pytest
+          path: dist
-
+      - run: pip install dist/*.whl numpy pytest
-      - name: Build and install ferro_ta
+      - run: |
-        run: |
-          maturin build --release --out dist
-          pip install dist/*.whl
-
-      - name: Generate reproducible perf artifacts
-        run: |
           python benchmarks/run_perf_contract.py \
             --output-dir perf-contract \
             --skip-talib \
@@ -161,12 +167,8 @@ jobs:
             --streaming-bars 20000 \
             --price-bars 20000 \
             --iv-bars 50000
-
+      - run: python benchmarks/check_hotspot_regression.py --input perf-contract/runtime_hotspots.json
-      - name: Enforce hotspot regression policy
+      - uses: actions/upload-artifact@v7
-        run: python benchmarks/check_hotspot_regression.py --input perf-contract/runtime_hotspots.json
-
-      - name: Upload perf artifacts
-        uses: actions/upload-artifact@v7
         with:
           name: perf-contract
           path: perf-contract/

.github/workflows/ci-rust.yml

@@ -7,102 +7,119 @@ permissions:
   contents: read
 
 jobs:
-  audit:
+  # -------------------------------------------------------------------------
-    name: Dependency audit (cargo + pip)
+  # cargo-deny — uses the official pre-built action (no cargo install needed)
+  # -------------------------------------------------------------------------
+  cargo-deny:
+    name: cargo deny check
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v6
+      - uses: EmbarkStudios/cargo-deny-action@v2
 
-      - name: Install cargo-audit and run cargo audit
+  # -------------------------------------------------------------------------
-        run: |
+  # pip-audit + uv.lock freshness check (lightweight, no Rust needed)
-          cargo install cargo-audit
+  # -------------------------------------------------------------------------
-          cargo audit
+  pip-audit:
-
+    name: pip-audit + uv.lock check
-      - name: Install cargo-deny and run cargo deny check
+    runs-on: ubuntu-latest
-        run: |
+    steps:
-          cargo install cargo-deny --locked
+      - uses: actions/checkout@v6
-          cargo deny check
+      - uses: actions/setup-python@v6
-
-      - name: Set up Python 3.12
-        uses: actions/setup-python@v6
         with:
           python-version: "3.12"
+      - run: pip install uv
+      - run: uv run --with pip-audit pip-audit --skip-editable
+      - run: uv lock --check
 
-      - name: Install uv
+  # -------------------------------------------------------------------------
-        run: pip install uv
+  # fmt + clippy (with Rust cache so subsequent runs skip recompilation)
-
+  # -------------------------------------------------------------------------
-      - name: Install pip-audit via uv and run pip-audit
+  rust-lint:
-        run: uv run --with pip-audit pip-audit --skip-editable
+    name: Rust fmt + clippy
-
-      - name: Verify uv.lock is up-to-date
-        run: uv lock --check
-
-  rust:
-    name: Rust (fmt + clippy)
     runs-on: ubuntu-latest
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
       - uses: dtolnay/rust-toolchain@v1
         with:
           toolchain: stable
           components: rustfmt, clippy
+      - uses: Swatinem/rust-cache@v2
       - run: cargo fmt --all -- --check
       - run: cargo clippy --release -- -D warnings
-      - name: Verify benchmarks compile (ferro_ta_core)
+      - name: Verify benchmarks compile
         run: cargo bench -p ferro_ta_core --no-run
 
+  # -------------------------------------------------------------------------
+  # Build + test ferro_ta_core (with Rust cache)
+  # -------------------------------------------------------------------------
   rust-core:
-    name: Rust core library (ferro_ta_core)
+    name: Rust core (build + test)
     runs-on: ubuntu-latest
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
       - uses: dtolnay/rust-toolchain@v1
         with:
           toolchain: stable
-      - name: Build core crate
+      - uses: Swatinem/rust-cache@v2
-        run: cargo build -p ferro_ta_core
+      - run: cargo build -p ferro_ta_core
-      - name: Test core crate
+      - run: cargo test -p ferro_ta_core
-        run: cargo test -p ferro_ta_core
 
+  # -------------------------------------------------------------------------
+  # Coverage (optional, cached)
+  # -------------------------------------------------------------------------
   rust-coverage:
-    name: Rust coverage (tarpaulin, optional)
+    name: Rust coverage (tarpaulin)
     runs-on: ubuntu-latest
     continue-on-error: true
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
       - uses: dtolnay/rust-toolchain@v1
         with:
           toolchain: stable
-      - name: Install cargo-tarpaulin
+      - uses: Swatinem/rust-cache@v2
-        run: cargo install cargo-tarpaulin --locked
+      - uses: taiki-e/install-action@v2
-      - name: Collect Rust coverage (ferro_ta_core)
+        with:
-        run: cargo tarpaulin -p ferro_ta_core --out Xml --output-dir coverage/
+          tool: cargo-tarpaulin
-      - name: Upload Rust coverage artifact
+      - run: cargo tarpaulin -p ferro_ta_core --out Xml --output-dir coverage/
-        uses: actions/upload-artifact@v7
+      - uses: actions/upload-artifact@v7
         with:
           name: rust-coverage
           path: coverage/
           if-no-files-found: ignore
 
+  # -------------------------------------------------------------------------
+  # Fuzz (optional, nightly, cached)
+  # -------------------------------------------------------------------------
   fuzz:
-    name: Fuzz targets (short CI run, optional)
+    name: Fuzz targets (short CI run)
     runs-on: ubuntu-latest
     continue-on-error: true
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
       - uses: dtolnay/rust-toolchain@v1
         with:
           toolchain: nightly
-      - name: Install cargo-fuzz
+          targets: x86_64-unknown-linux-gnu
-        run: cargo install cargo-fuzz --locked
+      - uses: Swatinem/rust-cache@v2
+      - uses: taiki-e/install-action@v2
+        with:
+          tool: cargo-fuzz
       - name: Run fuzz_sma (10000 iterations)
         working-directory: fuzz
-        run: cargo fuzz run fuzz_sma -- -runs=10000 -max_len=512
+        run: cargo fuzz run fuzz_sma --target x86_64-unknown-linux-gnu -- -runs=10000 -max_len=512
       - name: Run fuzz_rsi (10000 iterations)
         working-directory: fuzz
-        run: cargo fuzz run fuzz_rsi -- -runs=10000 -max_len=512
+        run: cargo fuzz run fuzz_rsi --target x86_64-unknown-linux-gnu -- -runs=10000 -max_len=512
-      - name: Upload fuzz artifacts on crash
+      - uses: actions/upload-artifact@v7
-        uses: actions/upload-artifact@v7
         if: always()
         with:
           name: fuzz-artifacts

.github/workflows/ci-wasm.yml

@@ -8,54 +8,43 @@ permissions:
 
 jobs:
   wasm:
-    name: WASM binding (wasm-pack test --node)
+    name: WASM (test + build + bench)
     runs-on: ubuntu-latest
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
     steps:
       - uses: actions/checkout@v6
-
+      - uses: actions/setup-node@v6
-      - name: Set up Node.js
-        uses: actions/setup-node@v6
         with:
           node-version: "20"
-
+      - uses: dtolnay/rust-toolchain@v1
-      - name: Install Rust (stable)
-        uses: dtolnay/rust-toolchain@v1
         with:
           toolchain: stable
           targets: wasm32-unknown-unknown
-
+      - uses: Swatinem/rust-cache@v2
-      - name: Install wasm-pack
+      - uses: taiki-e/install-action@v2
-        run: curl https://rustwasm.github.io/wasm-pack/installer/init.sh -sSf | sh
+        with:
-
+          tool: wasm-pack
-      - name: Build and test WASM binding
+      - name: Test WASM binding
         working-directory: wasm
         run: wasm-pack test --node
-
+      - name: Build WASM packages (node + web)
-      - name: Build WASM package (both targets)
         working-directory: wasm
         run: npm run build
-
       - name: Check API manifest is current
         run: python3 scripts/check_api_manifest.py
-
+      - name: Benchmark WASM
-      - name: Benchmark WASM package
         working-directory: wasm
         run: node bench.js --json ../wasm_benchmark.json
-
+      - uses: actions/upload-artifact@v7
-      - name: Upload WASM node package artifact
-        uses: actions/upload-artifact@v7
         with:
           name: wasm-pkg-node
           path: wasm/node/
-
+      - uses: actions/upload-artifact@v7
-      - name: Upload WASM web package artifact
-        uses: actions/upload-artifact@v7
         with:
           name: wasm-pkg-web
           path: wasm/web/
-
+      - uses: actions/upload-artifact@v7
-      - name: Upload WASM benchmark artifact
-        uses: actions/upload-artifact@v7
         with:
           name: wasm-benchmark
           path: wasm_benchmark.json

.github/workflows/nightly-bench.yml

@@ -0,0 +1,100 @@
+name: Nightly benchmarks
+
+on:
+  schedule:
+    # 03:15 UTC every day — off peak, avoids colliding with the regular CI surge.
+    - cron: "15 3 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  bench:
+    name: Run perf contract and regression checks
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      RUSTFLAGS: ""  # override target-cpu=native from .cargo/config.toml
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: dtolnay/rust-toolchain@v1
+        with:
+          toolchain: stable
+
+      - uses: Swatinem/rust-cache@v2
+
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        run: pip install uv
+
+      - name: Build ferro-ta wheel
+        run: |
+          uv sync --extra dev
+          uv run maturin build --release --out dist
+          uv run pip install --force-reinstall dist/*.whl
+
+      - name: Run vs-TA-Lib benchmark
+        run: uv run python benchmarks/bench_vs_talib.py --output benchmarks/artifacts/nightly_vs_talib.json
+
+      - name: Run SIMD benchmark matrix
+        run: uv run python benchmarks/bench_simd.py --output benchmarks/artifacts/nightly_simd.json
+
+      - name: Check hotspot regression
+        id: hotspot
+        continue-on-error: true
+        run: uv run python benchmarks/check_hotspot_regression.py --tolerance 0.05
+
+      - name: Check vs-TA-Lib regression
+        id: vs_talib
+        continue-on-error: true
+        run: uv run python benchmarks/check_vs_talib_regression.py --tolerance 0.05
+
+      - name: Run criterion bench (build only)
+        run: cargo bench -p ferro_ta_core --no-run
+
+      - name: Upload nightly artifacts
+        uses: actions/upload-artifact@v7
+        if: always()
+        with:
+          name: nightly-bench-${{ github.run_id }}
+          path: benchmarks/artifacts/
+          retention-days: 30
+
+      - name: Open issue on regression
+        if: steps.hotspot.outcome == 'failure' || steps.vs_talib.outcome == 'failure'
+        uses: actions/github-script@v9
+        env:
+          HOTSPOT_OUTCOME: ${{ steps.hotspot.outcome }}
+          VS_TALIB_OUTCOME: ${{ steps.vs_talib.outcome }}
+          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        with:
+          script: |
+            const hotspot = process.env.HOTSPOT_OUTCOME;
+            const vsTalib = process.env.VS_TALIB_OUTCOME;
+            const runUrl = process.env.RUN_URL;
+            const failed = [];
+            if (hotspot === 'failure') failed.push('hotspot');
+            if (vsTalib === 'failure') failed.push('vs-TA-Lib');
+            const title = `Nightly benchmark regression: ${failed.join(' + ')}`;
+            const body = [
+              `The nightly benchmark workflow reported a >5% regression in: **${failed.join(', ')}**.`,
+              '',
+              `Run: ${runUrl}`,
+              '',
+              'Artifacts with the raw JSON are attached to the run above.',
+              'If this is a known regression, update the baseline JSON and close.',
+              'If it is unexpected, investigate the last commit on `main` before the nightly ran.',
+            ].join('\n');
+            await github.rest.issues.create({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              title,
+              body,
+              labels: ['performance', 'regression', 'nightly'],
+            });

.github/workflows/release.yml

@@ -20,6 +20,7 @@ jobs:
       - uses: actions/checkout@v6
         with:
           fetch-depth: 0
+          token: ${{ secrets.GH_PAT }}
 
       - name: Extract version from tag
         id: version
@@ -51,10 +52,11 @@ jobs:
           PY
 
       - name: Create GitHub Release
-        uses: softprops/action-gh-release@v2
+        uses: softprops/action-gh-release@v3
         with:
           tag_name: ${{ github.ref_name }}
           name: v${{ steps.version.outputs.version }}
           body: ${{ steps.changelog.outputs.body }}
           draft: false
           prerelease: ${{ contains(github.ref_name, '-') }}
+          token: ${{ secrets.GH_PAT }}

CHANGELOG.md

@@ -9,6 +9,74 @@ and the project uses [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+### Added
+
+- **Runtime CPU-feature dispatch** for SIMD hot paths via the `multiversion`
+  crate (`ferro_ta_core/src/simd.rs`). One binary now selects baseline /
+  AVX2-FMA / AVX-512 / NEON kernels at load time via CPUID instead of being
+  pinned at build time, so it runs on any CPU of the target architecture with
+  no illegal-instruction crashes on older chips. The `simd` feature is on by
+  default; `--no-default-features` gives a pure-scalar build.
+- **Broader wheel coverage**: release builds now also produce Linux `aarch64`
+  (manylinux) and `musllinux` (x86_64 + aarch64) wheels and a Windows `arm64`
+  wheel, alongside the existing Linux x86_64, macOS universal2, and Windows
+  x64 wheels.
+- **abi3 wheels** (`cp310-abi3`): a single stable-ABI wheel per platform now
+  covers CPython 3.10+ (including future 3.14+), replacing the per-version
+  wheel matrix.
+- **Dynamic Time Warping** (`DTW`, `DTW_DISTANCE`, `BATCH_DTW`): Euclidean-cost
+  DTW with optional Sakoe-Chiba band (`window=` parameter). `DTW()` returns
+  `(distance, path)` with the optimal warping path as an `(N, 2)` index array;
+  `DTW_DISTANCE()` is the faster distance-only variant; `BATCH_DTW()` computes
+  distances from each row of a 2-D matrix to a reference series in parallel
+  via rayon. Distance convention matches `dtaidistance.dtw.distance()`.
+- **Indicator-specific exception types** (`FerroTaError`, `InvalidPeriodError`,
+  `InsufficientDataError`, `LengthMismatchError`, `NumericConvergenceError`,
+  `InvalidInputError`): finer-grained errors for catching specific failure
+  modes. All subclass `ValueError` so existing `except ValueError` code keeps
+  working (backward compatible).
+
+### Changed
+
+- **SIMD is now enabled by default and runtime-dispatched.** Replaced the
+  compile-time `wide` crate (which was never actually enabled in published
+  wheels) with `multiversion`. The `wide` Cargo feature is removed; use the
+  default `simd` feature instead.
+- **Dependency bumps.** Rust: `log` 0.4.32, `serde_json` 1.0.150,
+  `rayon` 1.12.0. API service (`api/requirements.txt`): `uvicorn>=0.49.0`,
+  `pydantic>=2.13.4`, `ferro-ta>=1.1.4`. CI actions: `actions/deploy-pages` v5,
+  `actions/upload-pages-artifact` v5, `softprops/action-gh-release` v3.
+- Python coverage threshold raised from 65% to 80% and enforced in CI.
+
+### Fixed
+
+- **aarch64 Linux containers can now install ferro-ta.** Previously no Linux
+  `aarch64` wheel was published, so arm64 images (e.g. AWS Graviton) fell back
+  to an sdist build that failed without a Rust toolchain. A manylinux/musllinux
+  aarch64 wheel is now published.
+- Published wheels now actually ship SIMD-accelerated kernels; the prior build
+  enabled no SIMD feature at all.
+
+### Security
+
+- **SLSA build provenance** attestations are now generated for every PyPI
+  wheel and sdist via `actions/attest-build-provenance`. Verify with
+  `gh attestation verify <wheel>`.
+- **Sigstore keyless signatures** are now published alongside both
+  CycloneDX/SPDX SBOMs on every GitHub Release (`.sig` + `.pem` files).
+- `ferro_ta_core` crate now declares `#![forbid(unsafe_code)]` to prevent
+  regression — the pure-logic layer has no unsafe code and never will.
+- Dependabot now covers the WASM npm package in addition to pip, cargo,
+  and GitHub Actions.
+- **pip-audit fixes**: bumped dev lockfile deps `idna` 3.18, `pytest` 9.1.1,
+  and `urllib3` 2.7.0 to clear PYSEC-2026-215, CVE-2025-71176, and
+  PYSEC-2026-141/142.
+- **pyo3 advisories triaged**: RUSTSEC-2026-0176 and RUSTSEC-2026-0177 are
+  ignored in `deny.toml` with rationale — ferro-ta uses neither affected code
+  path (PyList/PyTuple `nth` iterators; `PyCFunction::new_closure`). The
+  upstream fix requires pyo3 >=0.29 (a large API migration), tracked for a
+  follow-up.
+
 ## [1.1.3] — 2026-04-02
 
 ### Added

Cargo.lock

@@ -34,9 +34,9 @@ checksum = "940b3a0ca603d1eade50a4846a2afffd5ef57a9feac2c0e2ec2e14f9ead76000"
 
 [[package]]
 name = "arc-swap"
-version = "1.9.0"
+version = "1.9.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a07d1f37ff60921c83bdfc7407723bdefe89b44b98a9b772f225c8f9d67141a6"
+checksum = "6a3a1fd6f75306b68087b831f025c712524bcb19aad54e557b1129cfa0a2b207"
 dependencies = [
  "rustversion",
 ]
@@ -53,12 +53,6 @@ version = "3.20.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "5d20789868f4b01b2f2caec9f5c4e0213b41e3e5702a50157d699ae31ced2fcb"
 
-[[package]]
-name = "bytemuck"
-version = "1.25.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c8efb64bd706a16a1bdde310ae86b351e4d21550d98d056f22f8a7f7a2183fec"
-
 [[package]]
 name = "cast"
 version = "0.3.0"
@@ -67,9 +61,9 @@ checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
 [[package]]
 name = "cc"
-version = "1.2.57"
+version = "1.2.59"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a0dd1ca384932ff3641c8718a02769f1698e7563dc6974ffd03346116310423"
+checksum = "b7a4d3ec6524d28a329fc53654bbadc9bdd7b0431f5d65f1a56ffb28a1ee5283"
 dependencies = [
  "find-msvc-tools",
  "shlex",
@@ -207,7 +201,7 @@ checksum = "48c757948c5ede0e46177b7add2e67155f70e33c07fea8284df6576da70b3719"
 
 [[package]]
 name = "ferro_ta"
-version = "1.1.3"
+version = "1.1.4"
 dependencies = [
  "criterion",
  "ferro_ta_core",
@@ -222,12 +216,12 @@ dependencies = [
 
 [[package]]
 name = "ferro_ta_core"
-version = "1.1.3"
+version = "1.1.4"
 dependencies = [
  "criterion",
+ "multiversion",
  "serde",
  "serde_json",
- "wide",
 ]
 
 [[package]]
@@ -279,9 +273,9 @@ checksum = "8f42a60cbdf9a97f5d2305f08a87dc4e09308d1276d28c869c684d7777685682"
 
 [[package]]
 name = "js-sys"
-version = "0.3.91"
+version = "0.3.94"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b49715b7073f385ba4bc528e5747d02e66cb39c6146efb66b781f131f0fb399c"
+checksum = "2e04e2ef80ce82e13552136fabeef8a5ed1f985a96805761cbb9a2c34e7664d9"
 dependencies = [
  "once_cell",
  "wasm-bindgen",
@@ -289,15 +283,15 @@ dependencies = [
 
 [[package]]
 name = "libc"
-version = "0.2.183"
+version = "0.2.184"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5b646652bf6661599e1da8901b3b9522896f01e736bad5f723fe7a3a27f899d"
+checksum = "48f5d2a454e16a5ea0f4ced81bd44e4cfc7bd3a507b61887c99fd3538b28e4af"
 
 [[package]]
 name = "log"
-version = "0.4.29"
+version = "0.4.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
+checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
 
 [[package]]
 name = "matrixmultiply"
@@ -324,6 +318,28 @@ dependencies = [
  "autocfg",
 ]
 
+[[package]]
+name = "multiversion"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7edb7f0ff51249dfda9ab96b5823695e15a052dc15074c9dbf3d118afaf2c201"
+dependencies = [
+ "multiversion-macros",
+ "target-features",
+]
+
+[[package]]
+name = "multiversion-macros"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b093064383341eb3271f42e381cb8f10a01459478446953953c75d24bd339fc0"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+ "target-features",
+]
+
 [[package]]
 name = "ndarray"
 version = "0.16.1"
@@ -546,9 +562,9 @@ checksum = "60a357793950651c4ed0f3f52338f53b2f809f32d83a07f72909fa13e4c6c1e3"
 
 [[package]]
 name = "rayon"
-version = "1.11.0"
+version = "1.12.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "368f01d005bf8fd9b1206fb6fa653e6c4a81ceb1466406b81792d87c5677a58f"
+checksum = "fb39b166781f92d482534ef4b4b1b2568f42613b53e5b6c160e24cfbfa30926d"
 dependencies = [
  "either",
  "rayon-core",
@@ -595,9 +611,9 @@ checksum = "dc897dd8d9e8bd1ed8cdad82b5966c3e0ecae09fb1907d58efaa013543185d0a"
 
 [[package]]
 name = "rustc-hash"
-version = "2.1.1"
+version = "2.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "357703d41365b4b27c590e3ed91eabb1b663f07c4c084095e60cbed4362dff0d"
+checksum = "94300abf3f1ae2e2b8ffb7b58043de3d399c73fa6f4b73826402a5c457614dbe"
 
 [[package]]
 name = "rustversion"
@@ -605,15 +621,6 @@ version = "1.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b39cdef0fa800fc44525c84ccb54a029961a8215f9619753635a9c0d2538d46d"
 
-[[package]]
-name = "safe_arch"
-version = "1.0.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1f7caad094bd561859bcd467734a720c3c1f5d1f338995351fefe2190c45efed"
-dependencies = [
- "bytemuck",
-]
-
 [[package]]
 name = "same-file"
 version = "1.0.6"
@@ -655,9 +662,9 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.149"
+version = "1.0.150"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "83fc039473c5595ace860d8c4fafa220ff474b3fc6bfdb4293327f1a37e94d86"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
 dependencies = [
  "itoa",
  "memchr",
@@ -689,6 +696,12 @@ version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "609409d472a0a7d8d4dd9e19891bbdef546b9dce670c3057d0e02192dc541226"
 
+[[package]]
+name = "target-features"
+version = "0.1.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c1bbb9f3c5c463a01705937a24fdabc5047929ac764b2d5b9cf681c1f5041ed5"
+
 [[package]]
 name = "target-lexicon"
 version = "0.13.5"
@@ -729,9 +742,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen"
-version = "0.2.114"
+version = "0.2.117"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6532f9a5c1ece3798cb1c2cfdba640b9b3ba884f5db45973a6f442510a87d38e"
+checksum = "0551fc1bb415591e3372d0bc4780db7e587d84e2a7e79da121051c5c4b89d0b0"
 dependencies = [
  "cfg-if",
  "once_cell",
@@ -742,9 +755,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.114"
+version = "0.2.117"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "18a2d50fcf105fb33bb15f00e7a77b772945a2ee45dcf454961fd843e74c18e6"
+checksum = "7fbdf9a35adf44786aecd5ff89b4563a90325f9da0923236f6104e603c7e86be"
 dependencies = [
  "quote",
  "wasm-bindgen-macro-support",
@@ -752,9 +765,9 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.114"
+version = "0.2.117"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "03ce4caeaac547cdf713d280eda22a730824dd11e6b8c3ca9e42247b25c631e3"
+checksum = "dca9693ef2bab6d4e6707234500350d8dad079eb508dca05530c85dc3a529ff2"
 dependencies = [
  "bumpalo",
  "proc-macro2",
@@ -765,33 +778,23 @@ dependencies = [
 
 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.114"
+version = "0.2.117"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "75a326b8c223ee17883a4251907455a2431acc2791c98c26279376490c378c16"
+checksum = "39129a682a6d2d841b6c429d0c51e5cb0ed1a03829d8b3d1e69a011e62cb3d3b"
 dependencies = [
  "unicode-ident",
 ]
 
 [[package]]
 name = "web-sys"
-version = "0.3.91"
+version = "0.3.94"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "854ba17bb104abfb26ba36da9729addc7ce7f06f5c0f90f3c391f8461cca21f9"
+checksum = "cd70027e39b12f0849461e08ffc50b9cd7688d942c1c8e3c7b22273236b4dd0a"
 dependencies = [
  "js-sys",
  "wasm-bindgen",
 ]
 
-[[package]]
-name = "wide"
-version = "1.2.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "198f6abc41fab83526d10880fa5c17e2b4ee44e763949b4bb34e2fd1e8ca48e4"
-dependencies = [
- "bytemuck",
- "safe_arch",
-]
-
 [[package]]
 name = "winapi"
 version = "0.3.9"
@@ -840,18 +843,18 @@ dependencies = [
 
 [[package]]
 name = "zerocopy"
-version = "0.8.47"
+version = "0.8.48"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "efbb2a062be311f2ba113ce66f697a4dc589f85e78a4aea276200804cea0ed87"
+checksum = "eed437bf9d6692032087e337407a86f04cd8d6a16a37199ed57949d415bd68e9"
 dependencies = [
  "zerocopy-derive",
 ]
 
 [[package]]
 name = "zerocopy-derive"
-version = "0.8.47"
+version = "0.8.48"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e8bc7269b54418e7aeeef514aa68f8690b8c0489a06b0136e5f57c4c5ccab89"
+checksum = "70e3cd084b1788766f53af483dd21f93881ff30d7320490ec3ef7526d203bad4"
 dependencies = [
  "proc-macro2",
  "quote",

Cargo.toml

@@ -5,7 +5,7 @@ resolver = "2"
 
 [package]
 name = "ferro_ta"
-version = "1.1.3"
+version = "1.1.4"
 edition = "2021"
 description = "Rust-powered Python technical analysis library with a TA-Lib-compatible API"
 license = "MIT"
@@ -22,7 +22,9 @@ name = "ferro_ta"
 crate-type = ["cdylib"]
 
 [dependencies]
-pyo3 = { version = "0.25", features = ["extension-module"] }
+# abi3-py310: build a single stable-ABI wheel that runs on CPython 3.10+
+# (including future 3.14+), instead of one wheel per minor version.
+pyo3 = { version = "0.25", features = ["extension-module", "abi3-py310"] }
 ta = "0.5.0"
 numpy = "0.25"
 # Must be < 0.17 while numpy 0.25 is used (numpy's IntoPyArray is for its own ndarray only).
@@ -30,7 +32,11 @@ ndarray = "0.16"
 rayon = "1.10"
 log = "0.4"
 pyo3-log = "0.12"
-ferro_ta_core = { path = "crates/ferro_ta_core", version = "1.1.3", features = ["serde"] }
+# default-features = false so the `simd` toggle is forwarded explicitly via
+# this crate's own `simd` feature (below). Without this, core's default `simd`
+# would always be on and `--no-default-features` could never produce a true
+# pure-scalar build (used by the SIMD benchmark baseline).
+ferro_ta_core = { path = "crates/ferro_ta_core", version = "1.1.4", default-features = false, features = ["serde"] }
 
 [dev-dependencies]
 criterion = { version = "0.8", features = ["html_reports"] }
@@ -40,5 +46,7 @@ lto = true
 codegen-units = 1
 
 [features]
-default = []
+# SIMD runtime dispatch ON by default → published wheels ship the adaptive
+# fast path with no extra flags. `--no-default-features` yields pure scalar.
+default = ["simd"]
 simd = ["ferro_ta_core/simd"]

api/Dockerfile

@@ -8,21 +8,36 @@
 #
 # Environment variables (override at runtime):
 #   MAX_SERIES_LENGTH=100000   # maximum data-point count per request
+#
+# CPU portability
+# ---------------
+# This image installs the PRE-BUILT ferro-ta wheel from PyPI — we do NOT
+# recompile from sdist with `RUSTFLAGS=-C target-cpu=...`. The wheel is built
+# at the manylinux baseline (x86-64-v1) and selects AVX2/AVX-512/NEON kernels
+# at RUNTIME via CPU dispatch. One image therefore runs on any node — old or
+# new CPU, x86_64 or arm64 — with no illegal-instruction (SIGILL) crashes.
+# Pinning a target-cpu would be faster on a uniform fleet but would crash on
+# any older/heterogeneous node, which is the opposite of broad coverage.
+#
+# Build this image for whichever arch your nodes use:
+#   docker build --platform linux/amd64 -t ferro-ta-api .
+#   docker build --platform linux/arm64 -t ferro-ta-api .   # Graviton/Ampere
+# Both resolve a matching manylinux wheel — no Rust toolchain needed here.
 
 FROM python:3.11-slim
 
 WORKDIR /app
 
-# Install system dependencies required to build ferro_ta (Rust is pre-compiled
+# Copy and install dependencies first (cache layer). No compiler is needed:
-# into the wheel, so only pip + wheel tooling is needed at runtime).
+# ferro-ta, numpy, and pydantic-core all ship prebuilt wheels for linux
-RUN apt-get update && apt-get install -y --no-install-recommends \
+# x86_64 and aarch64.
-    gcc \
-    && rm -rf /var/lib/apt/lists/*
-
-# Copy and install dependencies first (cache layer)
 COPY requirements.txt ./
 RUN pip install --no-cache-dir -r requirements.txt
 
+# Fail the build immediately if the wheel did not resolve for this arch
+# (e.g. an exotic platform that fell back to an sdist build without Rust).
+RUN python -c "import ferro_ta, numpy as np; ferro_ta.SMA(np.arange(10.0), 3); print('ferro_ta', ferro_ta.__version__, 'import OK')"
+
 # Copy API source
 COPY main.py ./
 

api/requirements.txt

@@ -1,6 +1,6 @@
 # Runtime dependencies for ferro-ta API
-ferro_ta>=1.0.0
+ferro_ta>=1.1.4
 fastapi>=0.110.0
-uvicorn[standard]>=0.27.0
+uvicorn[standard]>=0.49.0
-pydantic>=2.0.0
+pydantic>=2.13.4
 numpy>=1.20

benchmarks/bench_simd.py

@@ -55,8 +55,11 @@ def run_simd_benchmark(
     iv_bars: int = 50_000,
     window: int = 252,
 ) -> dict[str, Any]:
+    # `simd` is a default feature, so a pure-scalar baseline must explicitly
+    # opt out via --no-default-features; otherwise both builds would be
+    # identical and every reported speedup would collapse to 1.0.
     variants = [
-        ("portable_release", []),
+        ("portable_release", ["--no-default-features"]),
         ("simd_release", ["--features", "simd"]),
     ]
     reports = {

conda/meta.yaml

@@ -1,5 +1,5 @@
 {% set name = "ferro-ta" %}
-{% set version = "1.1.3" %}
+{% set version = "1.1.4" %}
 
 package:
   name: {{ name|lower }}

crates/ferro_ta_core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "ferro_ta_core"
-version = "1.1.3"
+version = "1.1.4"
 edition = "2021"
 description = "Pure Rust core indicator library — no PyO3, no numpy dependency"
 license = "MIT"
@@ -16,7 +16,7 @@ name = "ferro_ta_core"
 crate-type = ["lib"]
 
 [dependencies]
-wide = { version = "1.1.1", optional = true }
+multiversion = { version = "0.8", optional = true }
 serde = { version = "1.0", features = ["derive"], optional = true }
 serde_json = { version = "1.0", optional = true }
 
@@ -28,6 +28,12 @@ name = "indicators"
 harness = false
 
 [features]
-wide = ["dep:wide"]
+# Runtime CPU-feature dispatch (multiversion). Default ON so `cargo add
-simd = ["wide"]
+# ferro_ta_core` and the published wheels get SIMD-accelerated reductions
+# that adapt to the running CPU (baseline .. AVX-512 / NEON) WITHOUT pinning
+# a target-cpu — one binary runs on any CPU of the target arch, with no
+# illegal-instruction crashes on older chips. Disable with
+# `--no-default-features` for a pure-scalar build.
+default = ["simd"]
+simd = ["dep:multiversion"]
 serde = ["dep:serde", "dep:serde_json"]

crates/ferro_ta_core/README.md

@@ -13,7 +13,7 @@ PyO3, NumPy, or Python runtime dependency, which makes it a good fit for:
 
 ```toml
 [dependencies]
-ferro_ta_core = "1.1.3"
+ferro_ta_core = "1.1.4"
 ```
 
 ## Design

crates/ferro_ta_core/src/lib.rs

@@ -1,3 +1,5 @@
+#![forbid(unsafe_code)]
+
 /*!
 ferro_ta_core — Pure Rust indicator library.
 
@@ -49,6 +51,8 @@ pub mod price_transform;
 pub mod regime;
 pub mod resampling;
 pub mod signals;
+/// Runtime-dispatched SIMD reduction primitives (internal).
+pub(crate) mod simd;
 pub mod statistic;
 pub mod streaming;
 pub mod volatility;

crates/ferro_ta_core/src/overlap.rs

@@ -38,27 +38,10 @@ pub fn sma_into(src: &[f64], timeperiod: usize, dest: &mut [f64], dest_offset: u
         return;
     }
 
-    #[cfg(feature = "simd")]
+    // Seed the rolling window with a runtime-dispatched reduction. The O(n)
-    let window_sum_init = {
+    // streaming recurrence below is inherently sequential, so SIMD only ever
-        use wide::f64x4;
+    // applies to this initial window sum.
-        let p_data = &src[..timeperiod];
+    let mut window_sum = crate::simd::sum(&src[..timeperiod]);
-        let mut sum = f64x4::splat(0.0);
-        let mut chunks = p_data.chunks_exact(4);
-        for chunk in &mut chunks {
-            sum += f64x4::new([chunk[0], chunk[1], chunk[2], chunk[3]]);
-        }
-        let arr = sum.to_array();
-        let mut total = arr[0] + arr[1] + arr[2] + arr[3];
-        for &v in chunks.remainder() {
-            total += v;
-        }
-        total
-    };
-
-    #[cfg(not(feature = "simd"))]
-    let window_sum_init: f64 = src[..timeperiod].iter().sum();
-
-    let mut window_sum = window_sum_init;
     let tp_f64 = timeperiod as f64;
     dest[dest_offset + timeperiod - 1] = window_sum / tp_f64;
 
@@ -124,46 +107,9 @@ pub fn wma(close: &[f64], timeperiod: usize) -> Vec<f64> {
     let denom: f64 = (timeperiod * (timeperiod + 1) / 2) as f64;
     let p = timeperiod as f64;
 
-    // Seed: compute T and S for the first window.
+    // Seed: compute T and S for the first window via a runtime-dispatched
-    #[cfg(feature = "simd")]
+    // reduction (the streaming recurrence below is sequential).
-    let (mut t, mut s) = {
+    let (mut t, mut s) = crate::simd::wma_seed(&close[..timeperiod]);
-        use wide::f64x4;
-        let p_data = &close[..timeperiod];
-        let mut t_simd = f64x4::splat(0.0);
-        let mut s_simd = f64x4::splat(0.0);
-        let mut chunks = p_data.chunks_exact(4);
-        let mut idx = 1.0;
-        let step = f64x4::new([0.0, 1.0, 2.0, 3.0]);
-
-        for chunk in &mut chunks {
-            let vals = f64x4::new([chunk[0], chunk[1], chunk[2], chunk[3]]);
-            let mults = f64x4::splat(idx) + step;
-            t_simd += vals * mults;
-            s_simd += vals;
-            idx += 4.0;
-        }
-        let t_arr = t_simd.to_array();
-        let s_arr = s_simd.to_array();
-        let mut t = t_arr[0] + t_arr[1] + t_arr[2] + t_arr[3];
-        let mut s = s_arr[0] + s_arr[1] + s_arr[2] + s_arr[3];
-        for &v in chunks.remainder() {
-            t += v * idx;
-            s += v;
-            idx += 1.0;
-        }
-        (t, s)
-    };
-
-    #[cfg(not(feature = "simd"))]
-    let (mut t, mut s) = {
-        let t_val: f64 = close[..timeperiod]
-            .iter()
-            .enumerate()
-            .map(|(k, &v)| v * (k + 1) as f64)
-            .sum();
-        let s_val: f64 = close[..timeperiod].iter().sum();
-        (t_val, s_val)
-    };
 
     result[timeperiod - 1] = t / denom;
 

crates/ferro_ta_core/src/simd.rs

@@ -0,0 +1,161 @@
+//! Runtime-dispatched SIMD primitives.
+//!
+//! Each public reduction here is compiled into several CPU-feature-specific
+//! variants (baseline, SSE, AVX2/FMA, AVX-512 on x86_64; NEON on aarch64; …)
+//! by [`multiversion`]. The fastest variant the *current* CPU supports is
+//! chosen at runtime via CPUID. This gives one binary that:
+//!
+//! * runs on **any** CPU of the target architecture — no illegal-instruction
+//!   (SIGILL) crashes on pre-AVX2 chips, unlike a static `-C target-cpu=…`;
+//! * still uses wide vector units where the hardware has them.
+//!
+//! The hot loops accumulate into **independent lanes** before a final
+//! horizontal combine. That is what lets the optimizer auto-vectorize them:
+//! a plain sequential `iter().sum()` is a dependency chain LLVM may not
+//! reorder (doing so would change floating-point rounding). As a consequence
+//! these results differ from a strict left-to-right sum by a few ULPs — well
+//! inside every indicator's documented tolerance.
+
+/// Number of independent accumulator lanes. Eight `f64` lanes cover the
+/// widest target we dispatch to (AVX-512 = 8×f64); narrower targets (AVX2,
+/// NEON) simply use a subset.
+#[cfg(feature = "simd")]
+const LANES: usize = 8;
+
+/// Sum of a slice of `f64`, runtime-dispatched.
+#[cfg(feature = "simd")]
+#[multiversion::multiversion(targets = "simd")]
+pub(crate) fn sum(data: &[f64]) -> f64 {
+    let mut acc = [0.0f64; LANES];
+    let mut chunks = data.chunks_exact(LANES);
+    for chunk in &mut chunks {
+        for (a, &v) in acc.iter_mut().zip(chunk) {
+            *a += v;
+        }
+    }
+    let remainder: f64 = chunks.remainder().iter().sum();
+    remainder + acc.iter().sum::<f64>()
+}
+
+/// Pure-scalar fallback when the `simd` feature is disabled.
+#[cfg(not(feature = "simd"))]
+pub(crate) fn sum(data: &[f64]) -> f64 {
+    data.iter().sum()
+}
+
+/// Weighted-moving-average seed for the first window.
+///
+/// Returns `(t, s)` where `t = Σ data[k] * (k + 1)` (1-based linear weights)
+/// and `s = Σ data[k]`. Used to seed the O(n) WMA recurrence.
+#[cfg(feature = "simd")]
+#[multiversion::multiversion(targets = "simd")]
+pub(crate) fn wma_seed(data: &[f64]) -> (f64, f64) {
+    // Lane-local accumulation (same idea as `sum`) so each CPU-feature clone
+    // can vectorize: `t` weights each value by its 1-based global index.
+    let mut t_acc = [0.0f64; LANES];
+    let mut s_acc = [0.0f64; LANES];
+    let mut chunks = data.chunks_exact(LANES);
+    let mut base = 0.0f64; // global index of this chunk's first element
+    for chunk in &mut chunks {
+        for (lane, ((t, s), &v)) in t_acc
+            .iter_mut()
+            .zip(s_acc.iter_mut())
+            .zip(chunk)
+            .enumerate()
+        {
+            *t += v * (base + lane as f64 + 1.0);
+            *s += v;
+        }
+        base += LANES as f64;
+    }
+    let mut t = 0.0;
+    let mut s = 0.0;
+    for (i, &v) in chunks.remainder().iter().enumerate() {
+        t += v * (base + i as f64 + 1.0);
+        s += v;
+    }
+    (t + t_acc.iter().sum::<f64>(), s + s_acc.iter().sum::<f64>())
+}
+
+/// Pure-scalar fallback when the `simd` feature is disabled.
+#[cfg(not(feature = "simd"))]
+pub(crate) fn wma_seed(data: &[f64]) -> (f64, f64) {
+    let mut t = 0.0;
+    let mut s = 0.0;
+    for (k, &v) in data.iter().enumerate() {
+        t += v * (k + 1) as f64;
+        s += v;
+    }
+    (t, s)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Strict sequential reference — the ground truth we compare against.
+    fn naive_sum(data: &[f64]) -> f64 {
+        data.iter().sum()
+    }
+
+    fn naive_wma_seed(data: &[f64]) -> (f64, f64) {
+        let t = data
+            .iter()
+            .enumerate()
+            .map(|(k, &v)| v * (k + 1) as f64)
+            .sum();
+        let s = data.iter().sum();
+        (t, s)
+    }
+
+    /// Deterministic test vectors spanning the lane boundaries: empty, a
+    /// partial chunk (< LANES), an exact multiple, and an exact-multiple +
+    /// remainder. This exercises every branch of the chunked reduction.
+    fn cases() -> Vec<Vec<f64>> {
+        let big: Vec<f64> = (0..1000).map(|i| (i as f64) * 0.5 - 123.0).collect();
+        vec![
+            vec![],
+            vec![42.0],
+            vec![1.0, 2.0, 3.0],                  // < LANES
+            (1..=8).map(|i| i as f64).collect(),  // exactly LANES
+            (1..=17).map(|i| i as f64).collect(), // LANES*2 + 1
+            big,
+        ]
+    }
+
+    #[test]
+    fn sum_matches_sequential_within_tolerance() {
+        for data in cases() {
+            let got = sum(&data);
+            let want = naive_sum(&data);
+            assert!(
+                (got - want).abs() <= 1e-9 * want.abs().max(1.0),
+                "sum mismatch: got {got}, want {want}, len {}",
+                data.len()
+            );
+        }
+    }
+
+    #[test]
+    fn wma_seed_matches_sequential_within_tolerance() {
+        for data in cases() {
+            let (t, s) = wma_seed(&data);
+            let (wt, ws) = naive_wma_seed(&data);
+            assert!(
+                (t - wt).abs() <= 1e-9 * wt.abs().max(1.0),
+                "wma t mismatch: got {t}, want {wt}, len {}",
+                data.len()
+            );
+            assert!(
+                (s - ws).abs() <= 1e-9 * ws.abs().max(1.0),
+                "wma s mismatch: got {s}, want {ws}, len {}",
+                data.len()
+            );
+        }
+    }
+
+    #[test]
+    fn sum_empty_is_zero() {
+        assert_eq!(sum(&[]), 0.0);
+    }
+}

crates/ferro_ta_core/src/statistic.rs

@@ -247,6 +247,118 @@ pub fn correl(real0: &[f64], real1: &[f64], timeperiod: usize) -> Vec<f64> {
     result
 }
 
+// ---------------------------------------------------------------------------
+// Dynamic Time Warping (DTW)
+// ---------------------------------------------------------------------------
+
+/// Internal helper: build the full DTW accumulated-cost matrix.
+///
+/// Local cost: `|s1[i] - s2[j]|` (Euclidean / L1 for 1-D series).
+/// This matches the convention used by `dtaidistance.dtw.distance()`.
+///
+/// Out-of-band cells (Sakoe-Chiba constraint) are set to `f64::INFINITY`.
+fn dtw_matrix(s1: &[f64], s2: &[f64], window: Option<usize>) -> Vec<Vec<f64>> {
+    let n = s1.len();
+    let m = s2.len();
+    let mut dp = vec![vec![f64::INFINITY; m]; n];
+    for i in 0..n {
+        // Window convention matches dtaidistance: window=w means |i-j| < w.
+        // None = unconstrained (full matrix).
+        let (j_lo, j_hi) = match window {
+            None => (0, m),
+            Some(w) => {
+                let lo = i.saturating_sub(w.saturating_sub(1));
+                let hi = i.saturating_add(w).min(m);
+                (lo, hi)
+            }
+        };
+        for j in j_lo..j_hi {
+            // Squared Euclidean local cost — matches dtaidistance convention.
+            // The final sqrt is applied only once at the top level (not per-step).
+            let cost = (s1[i] - s2[j]).powi(2);
+            let prev = if i == 0 && j == 0 {
+                0.0
+            } else if i == 0 {
+                dp[0][j - 1]
+            } else if j == 0 {
+                dp[i - 1][0]
+            } else {
+                dp[i - 1][j - 1].min(dp[i - 1][j]).min(dp[i][j - 1])
+            };
+            dp[i][j] = cost + prev;
+        }
+    }
+    dp
+}
+
+/// Compute the Dynamic Time Warping distance between two 1-D series.
+///
+/// Returns the accumulated Euclidean cost along the optimal warping path.
+/// Uses `|s1[i] - s2[j]|` as the local cost, matching `dtaidistance` convention.
+///
+/// # Arguments
+/// * `s1` - First time series.
+/// * `s2` - Second time series.
+/// * `window` - Optional Sakoe-Chiba band width. `None` = unconstrained.
+///
+/// Returns `f64::NAN` if either input is empty.
+pub fn dtw_distance(s1: &[f64], s2: &[f64], window: Option<usize>) -> f64 {
+    if s1.is_empty() || s2.is_empty() {
+        return f64::NAN;
+    }
+    let dp = dtw_matrix(s1, s2, window);
+    // sqrt applied once at the end — matches dtaidistance.dtw.distance() convention.
+    dp[s1.len() - 1][s2.len() - 1].sqrt()
+}
+
+/// Compute the DTW distance and the optimal warping path between two 1-D series.
+///
+/// The warping path is a `Vec<(usize, usize)>` of `(i, j)` index pairs,
+/// starting at `(0, 0)` and ending at `(n-1, m-1)`, monotonically non-decreasing.
+///
+/// # Arguments
+/// * `s1` - First time series.
+/// * `s2` - Second time series.
+/// * `window` - Optional Sakoe-Chiba band width. `None` = unconstrained.
+///
+/// Returns `(f64::NAN, vec![])` if either input is empty.
+pub fn dtw_path(s1: &[f64], s2: &[f64], window: Option<usize>) -> (f64, Vec<(usize, usize)>) {
+    if s1.is_empty() || s2.is_empty() {
+        return (f64::NAN, vec![]);
+    }
+    let dp = dtw_matrix(s1, s2, window);
+    let dist = dp[s1.len() - 1][s2.len() - 1].sqrt();
+
+    // Backtrace from (n-1, m-1) to (0, 0)
+    let mut path = Vec::new();
+    let (mut i, mut j) = (s1.len() - 1, s2.len() - 1);
+    path.push((i, j));
+    while i > 0 || j > 0 {
+        let (ni, nj) = match (i, j) {
+            (0, _) => (0, j - 1),
+            (_, 0) => (i - 1, 0),
+            _ => {
+                let diag = dp[i - 1][j - 1];
+                let up = dp[i - 1][j];
+                let left = dp[i][j - 1];
+                let best = diag.min(up).min(left);
+                if best == diag {
+                    (i - 1, j - 1)
+                } else if best == up {
+                    (i - 1, j)
+                } else {
+                    (i, j - 1)
+                }
+            }
+        };
+        i = ni;
+        j = nj;
+        path.push((i, j));
+    }
+    path.reverse();
+    (dist, path)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -259,4 +371,122 @@ mod tests {
             assert!(v.abs() < 1e-10);
         }
     }
+
+    #[test]
+    fn dtw_identical_series_is_zero() {
+        let a = vec![1.0, 2.0, 3.0, 4.0, 5.0];
+        assert_eq!(dtw_distance(&a, &a, None), 0.0);
+    }
+
+    #[test]
+    fn dtw_known_shifted_series() {
+        // [0,1,2] vs [1,2,3]: DTW uses squared Euclidean local cost + final sqrt.
+        // Optimal path (0,0)→(1,0)→(2,1)→(2,2), accumulated cost = 1+0+0+1 = 2, sqrt(2).
+        // Matches dtaidistance.dtw.distance([0,1,2],[1,2,3]) = 1.4142...
+        let a = vec![0.0, 1.0, 2.0];
+        let b = vec![1.0, 2.0, 3.0];
+        let expected = 2.0_f64.sqrt();
+        let result = dtw_distance(&a, &b, None);
+        assert!(
+            (result - expected).abs() < 1e-12,
+            "got {result}, expected {expected}"
+        );
+    }
+
+    #[test]
+    fn dtw_known_even_shift() {
+        // [0,2,4] vs [1,3,5]: diagonal path, squared costs 1+1+1=3, sqrt(3).
+        // Matches dtaidistance.dtw.distance([0,2,4],[1,3,5]) = 1.7320...
+        let a = vec![0.0, 2.0, 4.0];
+        let b = vec![1.0, 3.0, 5.0];
+        let expected = 3.0_f64.sqrt();
+        let result = dtw_distance(&a, &b, None);
+        assert!(
+            (result - expected).abs() < 1e-12,
+            "got {result}, expected {expected}"
+        );
+    }
+
+    #[test]
+    fn dtw_single_element() {
+        let a = vec![3.0];
+        let b = vec![7.0];
+        assert_eq!(dtw_distance(&a, &b, None), 4.0);
+    }
+
+    #[test]
+    fn dtw_empty_returns_nan() {
+        assert!(dtw_distance(&[], &[1.0, 2.0], None).is_nan());
+        assert!(dtw_distance(&[1.0, 2.0], &[], None).is_nan());
+    }
+
+    #[test]
+    fn dtw_path_endpoints() {
+        let a = vec![1.0, 2.0, 3.0, 4.0];
+        let b = vec![1.5, 2.5, 3.5, 4.5];
+        let (_, path) = dtw_path(&a, &b, None);
+        assert_eq!(path.first(), Some(&(0, 0)));
+        assert_eq!(path.last(), Some(&(3, 3)));
+    }
+
+    #[test]
+    fn dtw_path_is_monotone() {
+        let a = vec![1.0, 3.0, 2.0, 5.0, 4.0];
+        let b = vec![2.0, 1.0, 4.0, 3.0, 6.0];
+        let (_, path) = dtw_path(&a, &b, None);
+        for k in 1..path.len() {
+            assert!(path[k].0 >= path[k - 1].0);
+            assert!(path[k].1 >= path[k - 1].1);
+        }
+    }
+
+    #[test]
+    fn dtw_path_distance_matches_distance_only() {
+        let a = vec![1.0, 4.0, 2.0, 8.0, 3.0];
+        let b = vec![2.0, 3.0, 7.0, 4.0, 5.0];
+        let d1 = dtw_distance(&a, &b, None);
+        let (d2, _) = dtw_path(&a, &b, None);
+        assert!((d1 - d2).abs() < 1e-12);
+    }
+
+    #[test]
+    fn dtw_nan_in_input_propagates() {
+        // NaN in either input must propagate to the distance (IEEE 754 semantics).
+        let a = vec![1.0, 2.0, f64::NAN, 4.0];
+        let b = vec![1.0, 2.0, 3.0, 4.0];
+        assert!(dtw_distance(&a, &b, None).is_nan());
+        assert!(dtw_distance(&b, &a, None).is_nan());
+    }
+
+    #[test]
+    fn dtw_is_symmetric() {
+        let a = vec![1.0, 4.0, 2.0, 8.0, 3.0, 6.0, 5.0];
+        let b = vec![2.0, 3.0, 7.0, 4.0, 5.0, 1.0, 9.0];
+        let d_ab = dtw_distance(&a, &b, None);
+        let d_ba = dtw_distance(&b, &a, None);
+        assert!((d_ab - d_ba).abs() < 1e-12);
+    }
+
+    #[test]
+    fn dtw_path_length_bounded() {
+        // A valid warp path has length between max(n, m) and n + m - 1.
+        let a: Vec<f64> = (0..7).map(|x| x as f64).collect();
+        let b: Vec<f64> = (0..10).map(|x| (x as f64).sin()).collect();
+        let (_, path) = dtw_path(&a, &b, None);
+        let n = a.len();
+        let m = b.len();
+        assert!(path.len() >= n.max(m));
+        assert!(path.len() <= n + m - 1);
+    }
+
+    #[test]
+    fn dtw_window_constrained_ge_unconstrained() {
+        // window convention matches dtaidistance: Some(w) means |i-j| < w.
+        // A narrow window restricts warping, so constrained distance >= unconstrained.
+        let a: Vec<f64> = (0..20).map(|x| x as f64).collect();
+        let b: Vec<f64> = (0..20).map(|x| x as f64 + 3.0).collect();
+        let d_full = dtw_distance(&a, &b, None);
+        let d_narrow = dtw_distance(&a, &b, Some(3));
+        assert!(d_narrow >= d_full - 1e-12);
+    }
 }

deny.toml

@@ -53,7 +53,19 @@ skip = []
 db-urls = ["https://github.com/rustsec/advisory-db"]
 # Deny known security vulnerabilities
 version = 2
-ignore = []
+ignore = [
+    # pyo3 0.25 advisories — fix is pyo3 >=0.29, which is a large API
+    # migration (IntoPy/ToPyObject were removed in 0.26). Ignored here
+    # because ferro-ta does NOT use either affected code path:
+    #   * RUSTSEC-2026-0176 — OOB read in PyList/PyTuple nth/nth_back
+    #     iterators: the crate has zero PyList/PyTuple iterator usage.
+    #   * RUSTSEC-2026-0177 — missing Sync bound on
+    #     PyCFunction::new_closure: the crate uses #[pyfunction], never
+    #     new_closure.
+    # Tracked for removal once the pyo3 0.29 upgrade lands.
+    "RUSTSEC-2026-0176",
+    "RUSTSEC-2026-0177",
+]
 
 # ---------------------------------------------------------------------------
 # Sources — only allow crates from crates.io and our own path deps

docs/api_manifest.json

@@ -1,8 +1,8 @@
 {
   "surfaces": {
     "python": {
-      "indicator_count": 208,
+      "indicator_count": 211,
-      "method_count": 464,
+      "method_count": 467,
       "categories": [
         "aggregation",
         "alerts",
@@ -131,6 +131,13 @@
           "doc": "",
           "params": []
         },
+        {
+          "name": "BATCH_DTW",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
         {
           "name": "BBANDS",
           "category": "overlap",
@@ -656,6 +663,20 @@
           "doc": "",
           "params": []
         },
+        {
+          "name": "DTW",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
+        {
+          "name": "DTW_DISTANCE",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
         {
           "name": "DX",
           "category": "momentum",
@@ -3283,6 +3304,13 @@
           "doc": "",
           "params": []
         },
+        {
+          "name": "BATCH_DTW",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
         {
           "name": "BETA",
           "category": "statistic",
@@ -3297,6 +3325,20 @@
           "doc": "",
           "params": []
         },
+        {
+          "name": "DTW",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
+        {
+          "name": "DTW_DISTANCE",
+          "category": "statistic",
+          "module": "ferro_ta.indicators.statistic",
+          "doc": "",
+          "params": []
+        },
         {
           "name": "LINEARREG",
           "category": "statistic",
@@ -4735,7 +4777,7 @@
       ]
     },
     "rust_core": {
-      "public_function_count": 349,
+      "public_function_count": 351,
       "functions": [
         {
           "module": "aggregation",
@@ -6187,6 +6229,16 @@
           "function": "correl",
           "file": "statistic.rs"
         },
+        {
+          "module": "statistic",
+          "function": "dtw_distance",
+          "file": "statistic.rs"
+        },
+        {
+          "module": "statistic",
+          "function": "dtw_path",
+          "file": "statistic.rs"
+        },
         {
           "module": "statistic",
           "function": "linearreg",
@@ -6485,7 +6537,7 @@
       ]
     },
     "wasm_node": {
-      "export_count": 221,
+      "export_count": 222,
       "exports": [
         "ad",
         "adosc",
@@ -6546,6 +6598,7 @@
         "digital_price",
         "donchian",
         "drawdown_series",
+        "dtw_distance",
         "dx",
         "early_exercise_premium",
         "ema",
@@ -6712,9 +6765,9 @@
     }
   },
   "parity_summary": {
-    "python_indicator_count": 207,
+    "python_indicator_count": 210,
-    "wasm_export_count": 221,
+    "wasm_export_count": 222,
-    "common_python_wasm_count": 91,
+    "common_python_wasm_count": 92,
     "common_python_wasm": [
       "ad",
       "adosc",
@@ -6743,6 +6796,7 @@
       "dema",
       "detect_breaks_cusum",
       "donchian",
+      "dtw_distance",
       "dx",
       "ema",
       "ht_dcperiod",
@@ -6817,6 +6871,7 @@
       "asin",
       "atan",
       "batch_apply",
+      "batch_dtw",
       "beta",
       "cdl2crows",
       "cdl3blackcrows",
@@ -6886,6 +6941,7 @@
       "cosh",
       "div",
       "drawdown",
+      "dtw",
       "exp",
       "feature_matrix",
       "floor",

docs/changelog.rst

@@ -1,7 +1,7 @@
 Release Notes
 =============
 
-These docs track package version ``1.1.3``.
+These docs track package version ``1.1.4``.
 
 1.1.0-audit (2026-03-28)
 ------------------------

docs/guides/dtw.md

@@ -0,0 +1,80 @@
+# Dynamic Time Warping
+
+ferro-ta ships three DTW entry points. Pick the one that matches your
+workload — the distance-only path is measurably faster than the one that
+reconstructs the warping path, and `BATCH_DTW` parallelises over rows.
+
+## Quick reference
+
+| Function | Returns | When to use |
+|---|---|---|
+| `DTW_DISTANCE(a, b, window=None)` | `float` | You only need the distance. Fastest. |
+| `DTW(a, b, window=None)` | `(float, ndarray[N, 2])` | You need the alignment path for plotting or downstream analysis. |
+| `BATCH_DTW(matrix, reference, window=None)` | `ndarray[N]` | You have N candidate series and one reference; uses rayon. |
+
+## Distance convention
+
+ferro-ta's DTW uses squared-Euclidean local cost accumulated along the
+optimal path, with a single `sqrt()` applied at the end. This matches
+`dtaidistance.dtw.distance()` to within floating-point tolerance (parity
+tests assert numerical agreement, not bitwise identity). Example:
+
+```python
+>>> import ferro_ta as fta
+>>> fta.DTW_DISTANCE([0.0, 1.0, 2.0], [1.0, 2.0, 3.0])
+1.4142135623730951  # == sqrt(2), same as dtaidistance
+```
+
+If you are migrating from a library that uses absolute-difference local
+cost without the final sqrt (e.g. `fastdtw`'s default), your numbers will
+not line up. That is a choice ferro-ta made for parity with the
+scientific-Python ecosystem.
+
+## Window constraint (Sakoe-Chiba band)
+
+Passing `window=w` constrains the DP to cells where `|i - j| < w`. This
+turns the O(n·m) cost into O(n·w), which is typically a 5–20× speedup for
+realistic `w`. A narrower band can only *increase* the distance, so
+`window=` is safe to use whenever your series are roughly aligned.
+
+```python
+# Unconstrained
+fta.DTW_DISTANCE(a, b)
+
+# Constrained: warping may shift up to 5 positions
+fta.DTW_DISTANCE(a, b, window=5)
+```
+
+## Batch usage
+
+`BATCH_DTW` compares each row of a 2-D matrix against one reference
+series, in parallel:
+
+```python
+import numpy as np
+import ferro_ta as fta
+
+reference = np.random.random(500)
+candidates = np.random.random((1000, 500))
+
+distances = fta.BATCH_DTW(candidates, reference, window=20)
+nearest = int(np.argmin(distances))
+```
+
+Parallelism is via rayon; no thread-pool configuration is needed on the
+Python side. For the sequence lengths ferro-ta targets (thousands of
+bars, hundreds to low thousands of candidates), batch-parallel classic
+DTW beats FastDTW-style approximations.
+
+## Edge cases
+
+- **Empty input:** raises `FerroTAInputError`.
+- **NaN in input:** propagates to the output (matches IEEE 754). Call
+  `ferro_ta.core.exceptions.check_finite()` first if you want to fail
+  loudly instead.
+- **Different-length series:** fully supported. The path array length
+  is bounded by `max(n, m) <= len(path) <= n + m - 1`.
+
+## See also
+
+- `tests/unit/indicators/test_statistic.py` — parity tests against `dtaidistance`.

docs/guides/simd.md

@@ -0,0 +1,92 @@
+# SIMD acceleration
+
+ferro-ta accelerates hot reductions with **runtime CPU-feature dispatch**
+via the [`multiversion`](https://crates.io/crates/multiversion) crate. Each
+dispatched function is compiled into several variants — baseline, SSE,
+AVX2/FMA, AVX-512 on x86_64; NEON on aarch64 — and the fastest one the
+**current** CPU supports is chosen at load time via CPUID.
+
+## Why dispatch instead of `-C target-cpu`
+
+A static `RUSTFLAGS=-C target-cpu=x86-64-v3` build *requires* AVX2 on the
+running CPU; on an older chip it crashes with an illegal instruction
+(SIGILL). Runtime dispatch instead ships every code path in one binary and
+picks at runtime, so a single artifact:
+
+- runs on **any** CPU of the target architecture (no SIGILL on pre-AVX2
+  hardware), and
+- still uses wide vector units where the hardware has them.
+
+That property is what lets the **same** wheel / Docker image / crate run
+across a heterogeneous fleet.
+
+## When it helps
+
+SIMD helps indicators whose inner loop is a reduction over contiguous
+`f64` data — e.g. the initial window sum that seeds SMA, the `(T, S)` seed
+for WMA, and similar fixed-window reductions. It does **not** help:
+
+- The O(n) streaming recurrences (`window_sum += new - old`): each step
+  depends on the previous one, so they are inherently sequential.
+- Branchy inner loops (SAR, candlestick patterns).
+- Streaming classes (a single-bar update is one or two ops).
+
+The shared primitives live in `crates/ferro_ta_core/src/simd.rs`
+(`sum`, `wma_seed`). They accumulate into independent lanes before a final
+horizontal combine — that lane independence is what allows the optimizer to
+vectorize each CPU-feature variant. A consequence is that results differ
+from a strict left-to-right sum by a few ULPs, well inside every
+indicator's documented tolerance.
+
+## The `simd` feature
+
+Dispatch is gated behind the `simd` Cargo feature, which is **on by
+default**:
+
+```bash
+# default build — runtime dispatch enabled
+cargo build -p ferro_ta_core --release
+
+# pure-scalar build (debugging / baseline benchmarking)
+cargo build -p ferro_ta_core --release --no-default-features
+```
+
+For Python, wheels published to PyPI are built with the default features,
+so `pip install ferro-ta` ships the dispatched fast path with no action on
+your part. To build a pure-scalar extension from source:
+
+```bash
+maturin develop --release --no-default-features
+```
+
+## Measured speedups
+
+The nightly `benchmarks/bench_simd.py` job (see
+`.github/workflows/nightly-bench.yml`) builds the extension twice — once
+with `--no-default-features` (pure scalar) and once with `--features simd`
+(dispatch) — and reports the per-indicator delta. Numbers are regenerated
+on every run and vary with hardware; treat any table in a PR as a snapshot,
+not a contract. The dispatched kernels here target correctness-preserving
+reductions, so gains are modest on the sliding-window indicators and larger
+on full-array reductions.
+
+## Adding a SIMD-optimized indicator
+
+1. Write and test the **scalar** implementation first — it is the ground
+   truth.
+2. If the hot path is a contiguous `f64` reduction, route it through a
+   `crate::simd` primitive, or wrap a new helper in
+   `#[multiversion::multiversion(targets = "simd")]` with the loop body
+   accumulating into independent lanes.
+3. Add a parity test comparing the dispatched result against the strict
+   scalar reference within tolerance (see `simd.rs` tests for the pattern).
+4. Benchmark scalar vs dispatch via `bench_simd.py`. Only keep the SIMD
+   path if it wins — alignment and tail-handling overhead can make a naive
+   vectorization *lose* to scalar.
+
+## See also
+
+- `crates/ferro_ta_core/src/simd.rs` — dispatched primitives and tests.
+- `benches/indicators.rs` — criterion suite.
+- `crates/ferro_ta_core/Cargo.toml` `[features] simd = ["dep:multiversion"]`
+  — the gate (default-on).

docs/support_matrix.rst

@@ -180,7 +180,7 @@ For source builds, packaging details, and platform notes, see
 Release status
 --------------
 
-These docs track package version ``1.1.3``.
+These docs track package version ``1.1.4``.
 
 - Release notes by version: :doc:`changelog`
 - Canonical project changelog: `CHANGELOG.md <https://github.com/pratikbhadane24/ferro-ta/blob/main/CHANGELOG.md>`_

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "ferro-ta"
-version = "1.1.3"
+version = "1.1.4"
 description = "Rust-powered Python technical analysis library with a TA-Lib-compatible API"
 readme = "README.md"
 license = { text = "MIT" }
@@ -84,7 +84,7 @@ omit = ["*/_ferro_ta*", "*/mcp/*"]
 
 [tool.coverage.report]
 show_missing = true
-fail_under = 65
+fail_under = 80
 exclude_lines = [
     "pragma: no cover",
     "if TYPE_CHECKING:",

python/ferro_ta/__init__.py

@@ -110,8 +110,14 @@ __version__ = _detect_version()
 # ---------------------------------------------------------------------------
 from ferro_ta.core.exceptions import (  # noqa: F401
     FerroTAError,
+    FerroTaError,
     FerroTAInputError,
     FerroTAValueError,
+    InsufficientDataError,
+    InvalidInputError,
+    InvalidPeriodError,
+    LengthMismatchError,
+    NumericConvergenceError,
 )
 
 # ---------------------------------------------------------------------------

python/ferro_ta/core/exceptions.py

@@ -131,6 +131,63 @@ class FerroTAInputError(FerroTAError, ValueError):
     code = "FTERR002"
 
 
+# ---------------------------------------------------------------------------
+# Finer-grained exception subclasses (added in 1.2.0).
+#
+# These are drop-in compatible with the base classes: every subclass still
+# inherits from ``FerroTAError`` and ``ValueError``, so existing user code
+# like ``except FerroTAValueError:`` or ``except ValueError:`` keeps working.
+# The subclasses exist so users can catch *specific* failure modes without
+# string-matching on the error message.
+# ---------------------------------------------------------------------------
+
+
+class InvalidPeriodError(FerroTAValueError):
+    """Parameter like ``timeperiod``, ``fastperiod``, ``slowperiod`` is out of range.
+
+    Default error code: ``FTERR001``.
+    """
+
+
+class InsufficientDataError(FerroTAInputError):
+    """Input array is shorter than the minimum required for the indicator.
+
+    Default error code: ``FTERR003``.
+    """
+
+    code = "FTERR003"
+
+
+class LengthMismatchError(FerroTAInputError):
+    """Two or more input arrays (e.g. OHLC) have different lengths.
+
+    Default error code: ``FTERR004``.
+    """
+
+    code = "FTERR004"
+
+
+class NumericConvergenceError(FerroTAValueError):
+    """An iterative calculation failed to converge within tolerance.
+
+    Raised by iterative pricing models (implied volatility root-finding,
+    Newton-Raphson, etc.) when the maximum iteration count is exhausted.
+    """
+
+
+class InvalidInputError(FerroTAInputError):
+    """Input contains NaN/Inf in strict mode, wrong dtype, or wrong shape.
+
+    Default error code: ``FTERR005``.
+    """
+
+    code = "FTERR005"
+
+
+# Public aliases that match the names documented in the README and
+# CHANGELOG [Unreleased] section.
+FerroTaError = FerroTAError  # type: ignore[misc]
+
 # ---------------------------------------------------------------------------
 # Validation helpers (called by Python wrappers)
 # ---------------------------------------------------------------------------
@@ -154,7 +211,7 @@ def check_timeperiod(value: int, name: str = "timeperiod", minimum: int = 1) ->
         If ``value < minimum``.
     """
     if value < minimum:
-        raise FerroTAValueError(
+        raise InvalidPeriodError(
             f"{name} must be >= {minimum}, got {value}",
             suggestion=f"Set {name}={minimum} or higher.",
         )
@@ -193,7 +250,7 @@ def check_equal_length(**arrays: object) -> None:
 
     if len(set(lengths.values())) > 1:
         detail = ", ".join(f"{k}={v}" for k, v in lengths.items())
-        raise FerroTAInputError(
+        raise LengthMismatchError(
             f"All input arrays must have the same length. Got: {detail}",
             code=_CODE_LENGTH_MISMATCH,
             suggestion="Trim or align your arrays so that open, high, low, close, and volume all have the same number of rows.",
@@ -223,7 +280,7 @@ def check_finite(arr: object, name: str = "input") -> None:
 
     a = np.asarray(arr, dtype=np.float64)
     if not np.all(np.isfinite(a)):
-        raise FerroTAInputError(
+        raise InvalidInputError(
             f"{name} contains NaN or Inf values. "
             "ferro_ta propagates NaN by default; call check_finite() only "
             "when you require all-finite inputs.",
@@ -255,7 +312,7 @@ def check_min_length(arr: object, min_len: int, name: str = "input") -> None:
     elif hasattr(arr, "shape"):
         length = arr.shape[0]  # type: ignore[union-attr]
     if length < min_len:
-        raise FerroTAInputError(
+        raise InsufficientDataError(
             f"{name} must have at least {min_len} elements, got {length}",
             code=_CODE_TOO_SHORT,
             suggestion=f"Provide at least {min_len} data points. Current length: {length}.",

python/ferro_ta/indicators/statistic.py

@@ -12,19 +12,33 @@ LINEARREG_ANGLE  — Linear Regression Angle (degrees)
 TSF              — Time Series Forecast
 BETA             — Beta
 CORREL           — Pearson's Correlation Coefficient (r)
+DTW              — Dynamic Time Warping (distance + warping path)
+DTW_DISTANCE     — Dynamic Time Warping distance only (faster)
+BATCH_DTW        — Batch DTW: N series vs 1 reference, in parallel
 """
 
 from __future__ import annotations
 
+from typing import Optional
+
 import numpy as np
 from numpy.typing import ArrayLike
 
+from ferro_ta._ferro_ta import (
+    batch_dtw as _batch_dtw,
+)
 from ferro_ta._ferro_ta import (
     beta as _beta,
 )
 from ferro_ta._ferro_ta import (
     correl as _correl,
 )
+from ferro_ta._ferro_ta import (
+    dtw as _dtw,
+)
+from ferro_ta._ferro_ta import (
+    dtw_distance as _dtw_distance,
+)
 from ferro_ta._ferro_ta import (
     linearreg as _linearreg,
 )
@@ -247,6 +261,98 @@ def CORREL(real0: ArrayLike, real1: ArrayLike, timeperiod: int = 30) -> np.ndarr
         _normalize_rust_error(e)
 
 
+def DTW(
+    series1: ArrayLike,
+    series2: ArrayLike,
+    window: Optional[int] = None,
+) -> tuple[float, np.ndarray]:
+    """Dynamic Time Warping — distance and optimal warping path.
+
+    Parameters
+    ----------
+    series1 : array-like
+        First time series.
+    series2 : array-like
+        Second time series (may differ in length from series1).
+    window : int, optional
+        Sakoe-Chiba band width. ``None`` (default) = unconstrained.
+
+    Returns
+    -------
+    distance : float
+        DTW distance (accumulated Euclidean cost along the optimal path).
+    path : numpy.ndarray, shape (N, 2)
+        Warping path as ``(i, j)`` index pairs from ``(0, 0)`` to
+        ``(len(series1)-1, len(series2)-1)``.
+    """
+    try:
+        return _dtw(_to_f64(series1), _to_f64(series2), window)
+    except ValueError as e:
+        _normalize_rust_error(e)
+
+
+def DTW_DISTANCE(
+    series1: ArrayLike,
+    series2: ArrayLike,
+    window: Optional[int] = None,
+) -> float:
+    """Dynamic Time Warping distance only (faster — no path reconstruction).
+
+    Parameters
+    ----------
+    series1 : array-like
+        First time series.
+    series2 : array-like
+        Second time series (may differ in length from series1).
+    window : int, optional
+        Sakoe-Chiba band width. ``None`` (default) = unconstrained.
+
+    Returns
+    -------
+    float
+        DTW distance (accumulated Euclidean cost along the optimal path).
+    """
+    try:
+        return _dtw_distance(_to_f64(series1), _to_f64(series2), window)
+    except ValueError as e:
+        _normalize_rust_error(e)
+
+
+def BATCH_DTW(
+    matrix: ArrayLike,
+    reference: ArrayLike,
+    window: Optional[int] = None,
+) -> np.ndarray:
+    """Batch Dynamic Time Warping — N series vs 1 reference, computed in parallel.
+
+    Parameters
+    ----------
+    matrix : array-like, shape (N, L)
+        N time series of length L. Each row is compared against ``reference``.
+    reference : array-like, shape (L,)
+        The reference series.
+    window : int, optional
+        Sakoe-Chiba band width. ``None`` (default) = unconstrained.
+
+    Returns
+    -------
+    numpy.ndarray, shape (N,)
+        DTW distance from each row of ``matrix`` to ``reference``.
+    """
+    try:
+        mat = np.ascontiguousarray(matrix, dtype=np.float64)
+        if mat.ndim != 2:
+            from ferro_ta.core.exceptions import FerroTAInputError
+
+            raise FerroTAInputError(
+                f"matrix must be a 2-D array, got {mat.ndim}-D.",
+                suggestion="Pass a 2-D NumPy array of shape (N, L).",
+            )
+        return _batch_dtw(mat, _to_f64(reference), window)
+    except ValueError as e:
+        _normalize_rust_error(e)
+
+
 __all__ = [
     "STDDEV",
     "VAR",
@@ -257,4 +363,7 @@ __all__ = [
     "TSF",
     "BETA",
     "CORREL",
+    "DTW",
+    "DTW_DISTANCE",
+    "BATCH_DTW",
 ]

scripts/pre_push_checks.sh

@@ -1,27 +1,40 @@
 #!/usr/bin/env bash
+# Pre-push CI gate — runs checks in parallel to minimise wall-clock time.
+#
+# Usage:
+#   scripts/pre_push_checks.sh                    # all checks
+#   scripts/pre_push_checks.sh rust_clippy wasm   # selected checks
+#   scripts/pre_push_checks.sh --list
+#   FERRO_FAST=1 scripts/pre_push_checks.sh       # skip docs + wasm bench
 set -euo pipefail
 
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 cd "$ROOT_DIR"
 
 AVAILABLE_CHECKS=(
-  version
+  version changelog manifest
-  changelog
+  rust_fmt rust_clippy rust_core rust_bench
-  rust_fmt
+  python_lint python_typecheck python_test
-  rust_clippy
+  docs wasm
-  rust_core
-  rust_bench
-  python_lint
-  python_typecheck
-  python_test
-  docs
-  wasm
-  manifest
 )
-
 DEFAULT_CHECKS=("${AVAILABLE_CHECKS[@]}")
 
-python_env_ready=0
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+need_cmd() {
+  if ! command -v "$1" >/dev/null 2>&1; then
+    echo "Missing required command: $1" >&2; exit 1
+  fi
+}
+
+run_cmd() {
+  printf '    +'
+  printf ' %q' "$@"
+  printf '\n'
+  "$@"
+}
 
 usage() {
   cat <<'EOF'
@@ -30,93 +43,60 @@ Usage:
   scripts/pre_push_checks.sh <check> [<check> ...]
   scripts/pre_push_checks.sh --list
 
-Runs the repo's basic local CI gate before push. By default it covers:
+Environment:
-  version changelog rust_fmt rust_clippy rust_core rust_bench
+  FERRO_FAST=1   Skip docs and wasm (fastest local feedback loop)
-  python_lint python_typecheck python_test docs wasm manifest
-
-Notes:
-  - This mirrors the required CI categories we can run locally.
-  - It intentionally skips the multi-Python test matrix, audit jobs, perf smoke,
-    and benchmark-regression jobs.
 EOF
 }
 
-list_checks() {
+# ---------------------------------------------------------------------------
-  printf '%s\n' "${AVAILABLE_CHECKS[@]}"
+# Individual check functions
-}
+# ---------------------------------------------------------------------------
-
-need_cmd() {
-  local command_name="$1"
-  if ! command -v "$command_name" >/dev/null 2>&1; then
-    echo "Missing required command: $command_name" >&2
-    exit 1
-  fi
-}
-
-run_cmd() {
-  printf '  +'
-  printf ' %q' "$@"
-  printf '\n'
-  "$@"
-}
-
-ensure_python_env() {
-  if [[ "$python_env_ready" -eq 1 ]]; then
-    return
-  fi
-
-  need_cmd uv
-  run_cmd uv sync --extra dev --extra docs --extra mcp
-  run_cmd uv run --extra dev --extra docs --extra mcp maturin develop --release
-  python_env_ready=1
-}
-
-run_version() {
-  need_cmd python3
-  run_cmd python3 scripts/bump_version.py --check
-}
-
-run_changelog() {
-  need_cmd python3
-  run_cmd python3 scripts/check_changelog.py
-}
-
-run_rust_fmt() {
-  need_cmd cargo
-  run_cmd cargo fmt --all -- --check
-}
-
-run_rust_clippy() {
-  need_cmd cargo
-  run_cmd cargo clippy --release -- -D warnings
-}
-
-run_rust_core() {
-  need_cmd cargo
-  run_cmd cargo build -p ferro_ta_core
-  run_cmd cargo test -p ferro_ta_core
-}
-
-run_rust_bench() {
-  need_cmd cargo
-  run_cmd cargo bench -p ferro_ta_core --no-run
-}
 
+run_version()     { need_cmd python3; run_cmd python3 scripts/bump_version.py --check; }
+run_changelog()   { need_cmd python3; run_cmd python3 scripts/check_changelog.py; }
+run_manifest()    { need_cmd python3; run_cmd python3 scripts/check_api_manifest.py; }
+run_rust_fmt()    { need_cmd cargo;   run_cmd cargo fmt --all -- --check; }
 run_python_lint() {
   need_cmd uv
   run_cmd uv run --with ruff ruff check python/ tests/
   run_cmd uv run --with ruff ruff format --check python/ tests/
 }
 
+run_rust_clippy()      { need_cmd cargo; run_cmd cargo clippy --release -- -D warnings; }
+run_rust_core()        { need_cmd cargo; run_cmd cargo build -p ferro_ta_core && run_cmd cargo test -p ferro_ta_core; }
+run_rust_bench()       { need_cmd cargo; run_cmd cargo bench -p ferro_ta_core --no-run; }
+
 run_python_typecheck() {
   need_cmd uv
-  run_cmd uv run --with mypy --with numpy python -m mypy python/ferro_ta --ignore-missing-imports --no-error-summary
+  run_cmd uv run --with mypy --with numpy python -m mypy python/ferro_ta \
+    --ignore-missing-imports --no-error-summary
   run_cmd uv run --with pyright python -m pyright python/ferro_ta
 }
 
+# python_test and docs both need a compiled extension.
+# Use a flag file so only the first concurrent caller runs maturin develop;
+# subsequent callers (in parallel background jobs) wait and reuse it.
+_MATURIN_LOCK="${TMPDIR:-/tmp}/ferro_ta_maturin_$$.lock"
+_MATURIN_FLAG="${TMPDIR:-/tmp}/ferro_ta_maturin_$$.done"
+
+ensure_python_env() {
+  [[ -f "$_MATURIN_FLAG" ]] && return
+  (
+    flock 9
+    if [[ ! -f "$_MATURIN_FLAG" ]]; then
+      need_cmd uv
+      run_cmd uv sync --extra dev --extra docs --extra mcp
+      run_cmd uv run --extra dev --extra docs --extra mcp maturin develop --release
+      touch "$_MATURIN_FLAG"
+    fi
+  ) 9>"$_MATURIN_LOCK"
+}
+
 run_python_test() {
   ensure_python_env
-  run_cmd uv run --extra dev --extra mcp --with pytest-cov pytest tests/unit/ tests/integration/ -v --cov=ferro_ta --cov-report=term-missing --cov-fail-under=65
+  run_cmd uv run --extra dev --extra mcp --with pytest-cov \
+    pytest tests/unit/ tests/integration/ \
+    -v --cov=ferro_ta --cov-report=term-missing --cov-fail-under=65
 }
 
 run_docs() {
@@ -125,69 +105,161 @@ run_docs() {
 }
 
 run_wasm() {
-  need_cmd node
+  need_cmd node; need_cmd wasm-pack
-  need_cmd wasm-pack
-  local benchmark_json="../.wasm_benchmark.prepush.json"
   (
     cd wasm
-    trap 'rm -f "$benchmark_json"' EXIT
     run_cmd wasm-pack test --node
     run_cmd npm run build
-    run_cmd node bench.js --json "$benchmark_json"
+    if [[ "${FERRO_FAST:-0}" != "1" ]]; then
+      local bj="../.wasm_benchmark.prepush.json"
+      run_cmd node bench.js --json "$bj"
+      rm -f "$bj"
+    fi
   )
 }
 
-run_manifest() {
-  need_cmd python3
-  run_cmd python3 scripts/check_api_manifest.py
-}
-
 run_check() {
-  local check_name="$1"
+  case "$1" in
-  case "$check_name" in
+    version)          run_version ;;
-    version) run_version ;;
+    changelog)        run_changelog ;;
-    changelog) run_changelog ;;
+    manifest)         run_manifest ;;
-    rust_fmt) run_rust_fmt ;;
+    rust_fmt)         run_rust_fmt ;;
-    rust_clippy) run_rust_clippy ;;
+    rust_clippy)      run_rust_clippy ;;
-    rust_core) run_rust_core ;;
+    rust_core)        run_rust_core ;;
-    rust_bench) run_rust_bench ;;
+    rust_bench)       run_rust_bench ;;
-    python_lint) run_python_lint ;;
+    python_lint)      run_python_lint ;;
     python_typecheck) run_python_typecheck ;;
-    python_test) run_python_test ;;
+    python_test)      run_python_test ;;
-    docs) run_docs ;;
+    docs)             run_docs ;;
-    wasm) run_wasm ;;
+    wasm)             run_wasm ;;
-    manifest) run_manifest ;;
+    *) echo "Unknown check: $1 — use --list" >&2; exit 1 ;;
-    *)
-      echo "Unknown check: $check_name" >&2
-      echo "Use --list to see supported checks." >&2
-      exit 1
-      ;;
   esac
 }
 
-if [[ "${1:-}" == "--help" || "${1:-}" == "-h" ]]; then
+# ---------------------------------------------------------------------------
-  usage
+# Parallel runner — starts all checks concurrently, collects results
-  exit 0
+# ---------------------------------------------------------------------------
-fi
 
-if [[ "${1:-}" == "--list" ]]; then
+run_parallel() {
-  list_checks
+  local -a checks=("$@")
-  exit 0
+  [[ "${#checks[@]}" -eq 0 ]] && return 0
-fi
+
+  local -a pids logs names
+  local start
+  start=$(date +%s)
+
+  printf '\nStarting %d checks in parallel: %s\n' "${#checks[@]}" "${checks[*]}"
+
+  for check in "${checks[@]}"; do
+    local log
+    log=$(mktemp /tmp/ferro_prepush_XXXXXX)
+    logs+=("$log")
+    names+=("$check")
+    run_check "$check" >"$log" 2>&1 &
+    pids+=($!)
+  done
+
+  local failed=0
+  local -a failed_names
+  printf '\n'
+  for i in "${!pids[@]}"; do
+    if wait "${pids[$i]}" 2>/dev/null; then
+      printf '  ✓ %s\n' "${names[$i]}"
+    else
+      printf '  ✗ %s\n' "${names[$i]}"
+      failed_names+=("${names[$i]}")
+      failed=1
+    fi
+  done
+
+  # Print logs for failed checks only
+  if [[ "$failed" -eq 1 ]]; then
+    for i in "${!names[@]}"; do
+      local name="${names[$i]}"
+      if [[ " ${failed_names[*]:-} " == *" $name "* ]]; then
+        printf '\n'; printf '━%.0s' {1..60}; printf '\nFAILED: %s\n' "$name"; printf '━%.0s' {1..60}; printf '\n'
+        cat "${logs[$i]}"
+      fi
+    done
+  fi
+
+  for log in "${logs[@]}"; do rm -f "$log"; done
+  rm -f "$_MATURIN_LOCK" "$_MATURIN_FLAG"
+
+  printf '\nElapsed: %ds\n' "$(( $(date +%s) - start ))"
+  return "$failed"
+}
+
+# ---------------------------------------------------------------------------
+# Argument parsing
+# ---------------------------------------------------------------------------
+
+[[ "${1:-}" == "--help" || "${1:-}" == "-h" ]] && { usage; exit 0; }
+[[ "${1:-}" == "--list" ]] && { printf '%s\n' "${AVAILABLE_CHECKS[@]}"; exit 0; }
 
 selected_checks=()
 if [[ "$#" -gt 0 ]]; then
   selected_checks=("$@")
 else
   selected_checks=("${DEFAULT_CHECKS[@]}")
+  if [[ "${FERRO_FAST:-0}" == "1" ]]; then
+    selected_checks=()
+    for c in "${DEFAULT_CHECKS[@]}"; do
+      [[ "$c" == "docs" || "$c" == "wasm" ]] && continue
+      selected_checks+=("$c")
+    done
+    printf 'FERRO_FAST=1: skipping docs + wasm\n'
+  fi
 fi
 
-total_checks="${#selected_checks[@]}"
+# ---------------------------------------------------------------------------
-index=0
+# Execution strategy:
-for check_name in "${selected_checks[@]}"; do
+#   Phase 1 — instant gate (sequential, fail-fast):
-  index=$((index + 1))
+#              version, changelog, manifest, python_lint, rust_fmt
-  printf '\n[%d/%d] %s\n' "$index" "$total_checks" "$check_name"
+#              These are trivial to run and catch the most common mistakes early.
-  run_check "$check_name"
+#              If any fail here we abort immediately without waiting for slow checks.
+#
+#   Phase 2 — everything else in parallel:
+#              rust_clippy, rust_core, rust_bench, python_typecheck,
+#              python_test, docs, wasm
+# ---------------------------------------------------------------------------
+
+FAST_CHECKS=(version changelog manifest python_lint rust_fmt)
+
+phase1=()
+phase2=()
+for c in "${selected_checks[@]}"; do
+  is_fast=0
+  for f in "${FAST_CHECKS[@]}"; do [[ "$c" == "$f" ]] && is_fast=1 && break; done
+  if [[ "$is_fast" -eq 1 ]]; then phase1+=("$c"); else phase2+=("$c"); fi
 done
 
-printf '\nAll selected pre-push checks passed.\n'
+# Phase 1: fast gate
+if [[ "${#phase1[@]}" -gt 0 ]]; then
+  printf 'Phase 1 — fast gate (%d checks)\n' "${#phase1[@]}"
+  start1=$(date +%s)
+  for c in "${phase1[@]}"; do
+    printf '  [%s] ... ' "$c"
+    log=$(mktemp /tmp/ferro_prepush_XXXXXX)
+    if run_check "$c" >"$log" 2>&1; then
+      printf 'ok\n'
+    else
+      printf 'FAILED\n'
+      cat "$log"
+      rm -f "$log"
+      echo "" >&2
+      echo "Fast gate failed on '$c' — aborting before slow checks." >&2
+      exit 1
+    fi
+    rm -f "$log"
+  done
+  printf 'Phase 1 passed (%ds)\n' "$(( $(date +%s) - start1 ))"
+fi
+
+# Phase 2: parallel slow checks
+if [[ "${#phase2[@]}" -gt 0 ]]; then
+  printf '\nPhase 2 — parallel slow checks\n'
+  run_parallel "${phase2[@]}" || exit 1
+fi
+
+printf '\nAll pre-push checks passed.\n'

src/statistic/dtw.rs

@@ -0,0 +1,98 @@
+use ndarray::Array2;
+use numpy::{IntoPyArray, PyArray1, PyArray2, PyReadonlyArray1, PyReadonlyArray2};
+use pyo3::exceptions::PyValueError;
+use pyo3::prelude::*;
+use rayon::prelude::*;
+
+/// Dynamic Time Warping — distance and optimal warping path between two 1-D series.
+///
+/// Returns a tuple `(distance, path)` where `path` is a NumPy array of shape
+/// `(N, 2)` containing `(i, j)` index pairs from `(0, 0)` to `(n-1, m-1)`.
+///
+/// Local cost: `|series1[i] - series2[j]|` (Euclidean, matches `dtaidistance`).
+#[pyfunction]
+#[pyo3(signature = (series1, series2, window = None))]
+pub fn dtw<'py>(
+    py: Python<'py>,
+    series1: PyReadonlyArray1<'py, f64>,
+    series2: PyReadonlyArray1<'py, f64>,
+    window: Option<usize>,
+) -> PyResult<(f64, Bound<'py, PyArray2<usize>>)> {
+    let s1 = series1.as_slice()?;
+    let s2 = series2.as_slice()?;
+    if s1.is_empty() || s2.is_empty() {
+        return Err(PyValueError::new_err(
+            "series1 and series2 must not be empty",
+        ));
+    }
+    let (dist, path) = ferro_ta_core::statistic::dtw_path(s1, s2, window);
+    let n = path.len();
+    let flat: Vec<usize> = path.iter().flat_map(|&(i, j)| [i, j]).collect();
+    let arr =
+        Array2::from_shape_vec((n, 2), flat).map_err(|e| PyValueError::new_err(e.to_string()))?;
+    Ok((dist, arr.into_pyarray(py)))
+}
+
+/// Dynamic Time Warping — distance only (faster, no path reconstruction).
+///
+/// Returns the accumulated Euclidean cost along the optimal warping path.
+/// Use this when you only need the distance, not the alignment path.
+#[pyfunction]
+#[pyo3(signature = (series1, series2, window = None))]
+pub fn dtw_distance<'py>(
+    _py: Python<'py>,
+    series1: PyReadonlyArray1<'py, f64>,
+    series2: PyReadonlyArray1<'py, f64>,
+    window: Option<usize>,
+) -> PyResult<f64> {
+    let s1 = series1.as_slice()?;
+    let s2 = series2.as_slice()?;
+    if s1.is_empty() || s2.is_empty() {
+        return Err(PyValueError::new_err(
+            "series1 and series2 must not be empty",
+        ));
+    }
+    Ok(ferro_ta_core::statistic::dtw_distance(s1, s2, window))
+}
+
+/// Batch Dynamic Time Warping — compute DTW distance from each row of a 2-D matrix
+/// to a single reference series, in parallel.
+///
+/// Parameters
+/// ----------
+/// matrix : np.ndarray, shape (N, L)
+///     N time series of length L. Each row is compared against `reference`.
+/// reference : np.ndarray, shape (L,)
+///     The reference series.
+/// window : int, optional
+///     Sakoe-Chiba band width. `None` = unconstrained.
+///
+/// Returns
+/// -------
+/// np.ndarray, shape (N,)
+///     DTW distances, one per row.
+#[pyfunction]
+#[pyo3(signature = (matrix, reference, window = None))]
+pub fn batch_dtw<'py>(
+    py: Python<'py>,
+    matrix: PyReadonlyArray2<'py, f64>,
+    reference: PyReadonlyArray1<'py, f64>,
+    window: Option<usize>,
+) -> PyResult<Bound<'py, PyArray1<f64>>> {
+    let mat = matrix.as_array();
+    let ref_slice = reference.as_slice()?;
+
+    if ref_slice.is_empty() {
+        return Err(PyValueError::new_err("reference must not be empty"));
+    }
+
+    let (n_rows, _) = mat.dim();
+    let rows: Vec<Vec<f64>> = (0..n_rows).map(|i| mat.row(i).to_vec()).collect();
+
+    let result: Vec<f64> = rows
+        .par_iter()
+        .map(|series| ferro_ta_core::statistic::dtw_distance(series, ref_slice, window))
+        .collect();
+
+    Ok(result.into_pyarray(py))
+}

src/statistic/mod.rs

@@ -4,6 +4,7 @@
 mod beta;
 pub(crate) mod common;
 mod correl;
+mod dtw;
 mod linearreg;
 mod stddev;
 mod var;
@@ -23,5 +24,8 @@ pub fn register(m: &Bound<'_, PyModule>) -> PyResult<()> {
     m.add_function(pyo3::wrap_pyfunction!(self::linearreg::tsf, m)?)?;
     m.add_function(pyo3::wrap_pyfunction!(self::beta::beta, m)?)?;
     m.add_function(pyo3::wrap_pyfunction!(self::correl::correl, m)?)?;
+    m.add_function(pyo3::wrap_pyfunction!(self::dtw::dtw, m)?)?;
+    m.add_function(pyo3::wrap_pyfunction!(self::dtw::dtw_distance, m)?)?;
+    m.add_function(pyo3::wrap_pyfunction!(self::dtw::batch_dtw, m)?)?;
     Ok(())
 }

tests/unit/indicators/test_statistic.py

@@ -1,10 +1,14 @@
 """Unit tests for ferro_ta.indicators.statistic"""
 
 import numpy as np
+import pytest
 
 from ferro_ta.indicators.statistic import (
+    BATCH_DTW,
     BETA,
     CORREL,
+    DTW,
+    DTW_DISTANCE,
     LINEARREG,
     LINEARREG_ANGLE,
     LINEARREG_INTERCEPT,
@@ -308,3 +312,177 @@ class TestTSF:
         expected = _naive_linearreg(_A, timeperiod=14, x_value=14.0)
         result = TSF(_A, timeperiod=14)
         np.testing.assert_allclose(result, expected, equal_nan=True)
+
+
+# ---------------------------------------------------------------------------
+# DTW — Dynamic Time Warping
+# ---------------------------------------------------------------------------
+
+dtai = pytest.importorskip("dtaidistance", reason="dtaidistance not installed")
+
+_DTW_RNG = np.random.default_rng(42)
+
+
+class TestDTW:
+    # --- Validation against dtaidistance (SOTA reference) ---
+
+    def test_distance_matches_dtaidistance_random(self):
+        """Core correctness: our distance == dtaidistance on 20 random pairs."""
+        for _ in range(20):
+            n = int(_DTW_RNG.integers(5, 50))
+            a = _DTW_RNG.random(n)
+            b = _DTW_RNG.random(n)
+            expected = dtai.dtw.distance(a, b)
+            actual = DTW_DISTANCE(a, b)
+            np.testing.assert_allclose(
+                actual, expected, rtol=1e-9, err_msg=f"Mismatch on series length {n}"
+            )
+
+    def test_distance_matches_dtaidistance_unequal_length(self):
+        """Handles unequal-length series correctly."""
+        for _ in range(10):
+            a = _DTW_RNG.random(int(_DTW_RNG.integers(5, 30)))
+            b = _DTW_RNG.random(int(_DTW_RNG.integers(5, 30)))
+            expected = dtai.dtw.distance(a, b)
+            actual = DTW_DISTANCE(a, b)
+            np.testing.assert_allclose(actual, expected, rtol=1e-9)
+
+    def test_path_distance_matches_dtaidistance(self):
+        """DTW() path variant: returned distance matches dtaidistance."""
+        a = _DTW_RNG.random(20)
+        b = _DTW_RNG.random(25)
+        expected = dtai.dtw.distance(a, b)
+        dist, _ = DTW(a, b)
+        np.testing.assert_allclose(dist, expected, rtol=1e-9)
+
+    def test_path_matches_dtaidistance_warping_path(self):
+        """Warping path matches dtaidistance.dtw.warping_path() on same-length series."""
+        for _ in range(10):
+            n = int(_DTW_RNG.integers(5, 20))
+            a = _DTW_RNG.random(n)
+            b = _DTW_RNG.random(n)
+            expected_path = dtai.dtw.warping_path(a, b)
+            _, actual_path = DTW(a, b)
+            actual_pairs = [tuple(int(x) for x in row) for row in actual_path]
+            assert actual_pairs == expected_path, (
+                f"Path mismatch for n={n}:\n  ours={actual_pairs}\n  dtai={expected_path}"
+            )
+
+    def test_window_constrained_matches_dtaidistance(self):
+        """Sakoe-Chiba window matches dtaidistance window parameter."""
+        a = _DTW_RNG.random(30)
+        b = _DTW_RNG.random(30)
+        for w in [3, 8, 15]:
+            expected = dtai.dtw.distance(a, b, window=w)
+            actual = DTW_DISTANCE(a, b, window=w)
+            np.testing.assert_allclose(
+                actual, expected, rtol=1e-9, err_msg=f"Mismatch at window={w}"
+            )
+
+    def test_batch_matches_dtaidistance(self):
+        """BATCH_DTW matches calling dtaidistance per-row."""
+        ref = _DTW_RNG.random(20)
+        matrix = _DTW_RNG.random((8, 20))
+        batch_result = BATCH_DTW(matrix, ref)
+        for i in range(8):
+            expected = dtai.dtw.distance(matrix[i], ref)
+            np.testing.assert_allclose(
+                batch_result[i],
+                expected,
+                rtol=1e-9,
+                err_msg=f"Batch mismatch at row {i}",
+            )
+
+    # --- Mathematical properties ---
+
+    def test_identical_distance_is_zero(self):
+        a = np.array([1.0, 2.0, 3.0, 4.0, 5.0])
+        dist, _ = DTW(a, a)
+        assert dist == pytest.approx(0.0, abs=1e-10)
+
+    def test_symmetry(self):
+        a, b = _DTW_RNG.random(20), _DTW_RNG.random(20)
+        assert DTW_DISTANCE(a, b) == pytest.approx(DTW_DISTANCE(b, a), rel=1e-10)
+
+    def test_triangle_inequality(self):
+        a, b, c = _DTW_RNG.random(15), _DTW_RNG.random(15), _DTW_RNG.random(15)
+        assert DTW_DISTANCE(a, c) <= DTW_DISTANCE(a, b) + DTW_DISTANCE(b, c) + 1e-9
+
+    # --- Known hardcoded values ---
+
+    def test_known_shifted_series(self):
+        # [0,1,2] vs [1,2,3]: optimal path (0,0)→(1,0)→(2,1)→(2,2)
+        # Squared costs: 1+0+0+1=2, sqrt(2). Verified against dtaidistance.
+        a = np.array([0.0, 1.0, 2.0])
+        b = np.array([1.0, 2.0, 3.0])
+        np.testing.assert_allclose(DTW_DISTANCE(a, b), np.sqrt(2.0), rtol=1e-9)
+
+    def test_known_single_element(self):
+        # sqrt((3-7)^2) = sqrt(16) = 4.0
+        np.testing.assert_allclose(
+            DTW_DISTANCE(np.array([3.0]), np.array([7.0])), 4.0, rtol=1e-9
+        )
+
+    def test_known_constant_series(self):
+        assert DTW_DISTANCE(np.full(10, 5.0), np.full(10, 5.0)) == pytest.approx(
+            0.0, abs=1e-12
+        )
+
+    # --- Path structural guarantees ---
+
+    def test_path_starts_at_origin(self):
+        _, path = DTW(_DTW_RNG.random(10), _DTW_RNG.random(10))
+        assert tuple(int(x) for x in path[0]) == (0, 0)
+
+    def test_path_ends_at_corner(self):
+        _, path = DTW(_DTW_RNG.random(7), _DTW_RNG.random(9))
+        assert tuple(int(x) for x in path[-1]) == (6, 8)
+
+    def test_path_is_monotone(self):
+        _, path = DTW(_DTW_RNG.random(20), _DTW_RNG.random(20))
+        for k in range(1, len(path)):
+            assert path[k][0] >= path[k - 1][0]
+            assert path[k][1] >= path[k - 1][1]
+
+    def test_path_steps_unit_size(self):
+        _, path = DTW(_DTW_RNG.random(15), _DTW_RNG.random(12))
+        for k in range(1, len(path)):
+            di = int(path[k][0]) - int(path[k - 1][0])
+            dj = int(path[k][1]) - int(path[k - 1][1])
+            assert di in (0, 1) and dj in (0, 1)
+            assert not (di == 0 and dj == 0)
+
+    # --- DTW_DISTANCE == DTW distance ---
+
+    def test_distance_only_matches_full(self):
+        a, b = _DTW_RNG.random(25), _DTW_RNG.random(25)
+        d_full, _ = DTW(a, b)
+        np.testing.assert_allclose(DTW_DISTANCE(a, b), d_full, rtol=1e-10)
+
+    # --- Batch ---
+
+    def test_batch_single_row(self):
+        ref = np.array([1.0, 2.0, 3.0])
+        result = BATCH_DTW(np.array([[1.0, 2.0, 3.0]]), ref)
+        assert result[0] == pytest.approx(0.0, abs=1e-10)
+
+    def test_batch_matches_single_calls(self):
+        ref = _DTW_RNG.random(20)
+        matrix = _DTW_RNG.random((8, 20))
+        batch = BATCH_DTW(matrix, ref)
+        for i in range(8):
+            np.testing.assert_allclose(
+                batch[i], DTW_DISTANCE(matrix[i], ref), rtol=1e-10
+            )
+
+    # --- Edge cases ---
+
+    def test_empty_series_raises(self):
+        with pytest.raises((ValueError, Exception)):
+            DTW(np.array([]), np.array([1.0, 2.0]))
+
+    def test_window_constrained_ge_unconstrained(self):
+        a, b = _DTW_RNG.random(20), _DTW_RNG.random(20)
+        d_full = DTW_DISTANCE(a, b)
+        d_narrow = DTW_DISTANCE(a, b, window=2)
+        assert d_narrow >= d_full - 1e-9

uv.lock

@@ -950,7 +950,7 @@ wheels = [
 
 [[package]]
 name = "ferro-ta"
-version = "1.1.3"
+version = "1.1.4"
 source = { editable = "." }
 dependencies = [
     { name = "numpy" },
@@ -1272,11 +1272,11 @@ wheels = [
 
 [[package]]
 name = "idna"
-version = "3.11"
+version = "3.18"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/6f/6d/0703ccc57f3a7233505399edb88de3cbd678da106337b9fcde432b65ed60/idna-3.11.tar.gz", hash = "sha256:795dafcc9c04ed0c1fb032c2aa73654d8e8c5023a7df64a53f39190ada629902", size = 194582, upload-time = "2025-10-12T14:55:20.501Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/cd/63/9496c57188a2ee585e0f1db071d75089a11e98aa86eb99d9d7618fc1edce/idna-3.18.tar.gz", hash = "sha256:ffb385a7e039654cef1ab9ef32c6fafe283c0c0467bba1d9029738ce4a14a848", size = 196711, upload-time = "2026-06-02T14:34:07.794Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/0e/61/66938bbb5fc52dbdf84594873d5b51fb1f7c7794e9c0f5bd885f30bc507b/idna-3.11-py3-none-any.whl", hash = "sha256:771a87f49d9defaf64091e6e6fe9c18d4833f140bd19464795bc32d966ca37ea", size = 71008, upload-time = "2025-10-12T14:55:18.883Z" },
+    { url = "https://files.pythonhosted.org/packages/1e/5e/d4e9f1a599fb8e573b7b87160658329fbf28d19eac2718f51fc3def3aa5a/idna-3.18-py3-none-any.whl", hash = "sha256:7f952cbe720b688055e3f87de14f5c3e5fdaa8bc3928985c4077ca689de849a2", size = 65455, upload-time = "2026-06-02T14:34:06.319Z" },
 ]
 
 [[package]]
@@ -2906,7 +2906,7 @@ wheels = [
 
 [[package]]
 name = "pytest"
-version = "9.0.2"
+version = "9.1.1"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "colorama", marker = "sys_platform == 'win32'" },
@@ -2917,9 +2917,9 @@ dependencies = [
     { name = "pygments" },
     { name = "tomli", marker = "python_full_version < '3.11'" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/d1/db/7ef3487e0fb0049ddb5ce41d3a49c235bf9ad299b6a25d5780a89f19230f/pytest-9.0.2.tar.gz", hash = "sha256:75186651a92bd89611d1d9fc20f0b4345fd827c41ccd5c299a868a05d70edf11", size = 1568901, upload-time = "2025-12-06T21:30:51.014Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/e4/47/b9efed96c114afcfa3c9d3fe98a76a1d14c74a9e266d397cf6eb64be5e01/pytest-9.1.1.tar.gz", hash = "sha256:1088fbde8f2b49d95a549a195707afa7a76a3ce9bcadc26b6d71f0ffda5fe313", size = 1636369, upload-time = "2026-06-19T10:58:32.857Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/3b/ab/b3226f0bd7cdcf710fbede2b3548584366da3b19b5021e74f5bde2a8fa3f/pytest-9.0.2-py3-none-any.whl", hash = "sha256:711ffd45bf766d5264d487b917733b453d917afd2b0ad65223959f59089f875b", size = 374801, upload-time = "2025-12-06T21:30:49.154Z" },
+    { url = "https://files.pythonhosted.org/packages/24/25/1de2678b631f5a49215c6c96fff41ba892b0a34df68d6d80292b1b48aa7f/pytest-9.1.1-py3-none-any.whl", hash = "sha256:37a86b45efb9a47a61a36449063e8e18d0cab3161329fc099eb21783169c4f0c", size = 386536, upload-time = "2026-06-19T10:58:31.347Z" },
 ]
 
 [[package]]
@@ -4261,11 +4261,11 @@ wheels = [
 
 [[package]]
 name = "urllib3"
-version = "2.6.3"
+version = "2.7.0"
 source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/c7/24/5f1b3bdffd70275f6661c76461e25f024d5a38a46f04aaca912426a2b1d3/urllib3-2.6.3.tar.gz", hash = "sha256:1b62b6884944a57dbe321509ab94fd4d3b307075e0c2eae991ac71ee15ad38ed", size = 435556, upload-time = "2026-01-07T16:24:43.925Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/53/0c/06f8b233b8fd13b9e5ee11424ef85419ba0d8ba0b3138bf360be2ff56953/urllib3-2.7.0.tar.gz", hash = "sha256:231e0ec3b63ceb14667c67be60f2f2c40a518cb38b03af60abc813da26505f4c", size = 433602, upload-time = "2026-05-07T16:13:18.596Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/39/08/aaaad47bc4e9dc8c725e68f9d04865dbcb2052843ff09c97b08904852d84/urllib3-2.6.3-py3-none-any.whl", hash = "sha256:bf272323e553dfb2e87d9bfd225ca7b0f467b919d7bbd355436d3fd37cb0acd4", size = 131584, upload-time = "2026-01-07T16:24:42.685Z" },
+    { url = "https://files.pythonhosted.org/packages/7f/3e/5db95bcf282c52709639744ca2a8b149baccf648e39c8cc87553df9eae0c/urllib3-2.7.0-py3-none-any.whl", hash = "sha256:9fb4c81ebbb1ce9531cce37674bbc6f1360472bc18ca9a553ede278ef7276897", size = 131087, upload-time = "2026-05-07T16:13:17.151Z" },
 ]
 
 [[package]]

wasm/Cargo.lock

@@ -49,11 +49,11 @@ checksum = "9330f8b2ff13f34540b44e946ef35111825727b38d33286ef986142615121801"
 
 [[package]]
 name = "ferro_ta_core"
-version = "1.1.3"
+version = "1.1.4"
 
 [[package]]
 name = "ferro_ta_wasm"
-version = "1.1.3"
+version = "1.1.4"
 dependencies = [
  "ferro_ta_core",
  "js-sys",

wasm/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "ferro_ta_wasm"
-version = "1.1.3"
+version = "1.1.4"
 edition = "2021"
 description = "WebAssembly bindings for ferro-ta technical analysis indicators"
 license = "MIT"

wasm/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ferro-ta-wasm",
-  "version": "1.1.3",
+  "version": "1.1.4",
   "description": "WebAssembly bindings for ferro-ta technical analysis indicators",
   "main": "node/ferro_ta_wasm.js",
   "module": "web/ferro_ta_wasm.js",

wasm/src/lib.rs

@@ -1653,6 +1653,18 @@ pub fn correl(real0: &Float64Array, real1: &Float64Array, timeperiod: usize) ->
     from_vec(ferro_ta_core::statistic::correl(&to_vec(real0), &to_vec(real1), timeperiod))
 }
 
+/// Dynamic Time Warping distance between two series.
+///
+/// Returns the accumulated Euclidean cost along the optimal warping path.
+/// Pass `window` as `0` for unconstrained (no Sakoe-Chiba band).
+#[wasm_bindgen]
+pub fn dtw_distance(series1: &Float64Array, series2: &Float64Array, window: usize) -> f64 {
+    let s1 = to_vec(series1);
+    let s2 = to_vec(series2);
+    let w = if window == 0 { None } else { Some(window) };
+    ferro_ta_core::statistic::dtw_distance(&s1, &s2, w)
+}
+
 // ===========================================================================
 // Streaming / Stateful API
 // ===========================================================================