fix(ci): retain native dependencies in per-interpreter caches

.github/workflows/ci.yml

@@ -184,6 +184,7 @@ jobs:
       - uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           version: ${{ env.UV_VERSION }}
+          cache-suffix: python-${{ matrix.python-version }}
       - name: Set up Python ${{ matrix.python-version }}
         uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
         with:

.github/workflows/native-stress.yml

@@ -81,6 +81,7 @@ jobs:
       - uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884
         with:
           version: "0.10.9"
+          cache-suffix: python-${{ matrix.python-version }}
       - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97
         with:
           python-version: ${{ matrix.python-version }}

docs/getting-started/installation.md

@@ -47,6 +47,21 @@ Use the [upstream installation guide](https://github.com/TA-Lib/ta-lib-python#in
 when your platform requires a source build. Comparisons must declare matching moving-average
 types: Engineer's APO/PPO default remains SMA (`matype=0`), while TA-Lib 0.8 defaults to EMA.
 
+### Feature store
+
+The `store` extra adds DuckDB and PyArrow; the core quick start does not need it.
+Our Python 3.15 qualification used DuckDB 1.5.6 source builds on Linux, macOS and Windows.
+If no matching wheel is available, install the
+[upstream native build prerequisites](https://duckdb.org/docs/current/dev/building/python)
+before enabling the extra. Source builds need a C++ toolchain and can take tens of minutes.
+
+Within the source checkout, enable the extra and verify both imports:
+
+```bash
+uv sync --no-dev --extra store
+uv run --no-sync python -c "import duckdb; import ml4t.engineer.store; print(duckdb.__version__)"
+```
+
 ### Statistical diagnostics
 
 `find_optimal_d()` and `fdiff_diagnostics()` use Statsmodels for the augmented