Build with uv + vendored choldate; verified pure-NumPy replacement

.gitignore

@@ -3,4 +3,15 @@ bazel-*
 data/factor_covariance.csv
 data/factor_exposure.csv
 data/specific_variance.csv
-data/topic_exposure.csv 
+data/topic_exposure.csv
+
+# Python bytecode / caches
+__pycache__/
+*.py[cod]
+.pytest_cache/
+
+# Packaging artifacts
+*.egg-info/
+
+# OS cruft
+.DS_Store 

.python-version

@@ -0,0 +1 @@
+3.12

AGENTS.md

@@ -0,0 +1,105 @@
+# Directory Map
+
+This is a map of the directory for reference.
+
+```
+.
+├── .DS_Store
+├── .gitignore
+├── .vscode
+│   └── settings.json
+├── AGENTS.md
+├── README.md
+├── data
+│   ├── BFI228.csv
+│   ├── blackrock_covariance.csv
+│   └── eighthr.data
+├── environment.yml
+├── figures
+│   ├── .DS_Store
+│   ├── blackrock_results.pdf
+│   ├── dist_comparison.pdf
+│   ├── model_selec_results_n=100_gaussian.pdf
+│   ├── model_selec_results_n=100_non_gaussian.pdf
+│   ├── model_selec_results_n=200_gaussian.pdf
+│   ├── model_selec_results_n=200_gaussian_fb.pdf
+│   ├── model_selec_results_n=200_non_gaussian.pdf
+│   ├── model_selec_results_n=200_non_gaussian_fb.pdf
+│   ├── model_selec_results_n=300_gaussian.pdf
+│   ├── model_selec_results_n=300_non_gaussian.pdf
+│   ├── model_selec_results_n=400_gaussian.pdf
+│   ├── model_selec_results_n=400_non_gaussian.pdf
+│   ├── model_selec_results_n=500_gaussian.pdf
+│   ├── model_selec_results_n=500_non_gaussian.pdf
+│   ├── ozone_results.pdf
+│   ├── reliability.pdf
+│   └── temp.pdf
+├── notebooks
+│   ├── .DS_Store
+│   ├── .ipynb_checkpoints
+│   │   ├── Development Notebook-checkpoint.ipynb
+│   │   ├── blackrock_css-checkpoint.ipynb
+│   │   ├── missing_data-checkpoint.ipynb
+│   │   ├── model_selection-checkpoint.ipynb
+│   │   ├── testing_data_generation_and_null_distribution-checkpoint.ipynb
+│   │   └── testing_subset_selection-checkpoint.ipynb
+│   ├── blackrock_css.ipynb
+│   ├── missing_data.ipynb
+│   ├── model_selection.ipynb
+│   ├── notebook_utils
+│   │   ├── .DS_Store
+│   │   ├── __pycache__
+│   │   │   ├── data_generation.cpython-310.pyc
+│   │   │   ├── missing_data.cpython-310.pyc
+│   │   │   └── utils.cpython-310.pyc
+│   │   ├── data_generation.py
+│   │   ├── missing_data.py
+│   │   └── utils.py
+│   ├── results
+│   │   ├── gaussian_results.csv
+│   │   └── non_gaussian_results
+│   ├── testing_data_generation_and_null_distribution.ipynb
+│   └── testing_subset_selection.ipynb
+├── pycss
+│   ├── .DS_Store
+│   ├── CSS.py
+│   ├── __init__.py
+│   ├── __pycache__
+│   │   ├── PCSS.cpython-310.pyc
+│   │   ├── __init__.cpython-310.pyc
+│   │   ├── css.cpython-310.pyc
+│   │   ├── hypothesis_testing.cpython-310.pyc
+│   │   ├── model_selection.cpython-310.pyc
+│   │   ├── subset_selection.cpython-310.pyc
+│   │   └── utils.cpython-310.pyc
+│   ├── model_selection.py
+│   ├── subset_selection.py
+│   └── utils.py
+├── pycss.egg-info
+│   ├── PKG-INFO
+│   ├── SOURCES.txt
+│   ├── dependency_links.txt
+│   ├── not-zip-safe
+│   ├── requires.txt
+│   └── top_level.txt
+├── requirements.txt
+├── setup.py
+└── tests
+    ├── .DS_Store
+    ├── __pycache__
+    │   ├── test_CSS.cpython-310-pytest-7.3.1.pyc
+    │   ├── test_PCSS.cpython-310-pytest-7.2.1.pyc
+    │   ├── test_PCSS.cpython-310-pytest-7.3.1.pyc
+    │   ├── test_css.cpython-310-pytest-7.2.1.pyc
+    │   ├── test_subset_selection.cpython-310-pytest-7.2.1.pyc
+    │   ├── test_subset_selection.cpython-310-pytest-7.3.1.pyc
+    │   ├── test_utils.cpython-310-pytest-7.2.1.pyc
+    │   └── test_utils.cpython-310-pytest-7.3.1.pyc
+    ├── test_CSS.py
+    ├── test_subset_selection.py
+    ├── test_utils.py
+    └── utils
+        ├── .DS_Store
+        └── __pycache__
+            └── test_general.cpython-310-pytest-7.2.1.pyc
+```

README.md

@@ -2,17 +2,17 @@
 
 ## Installation
 
-1. If you do not have conda installed already, please install it. There are many ways to get conda. We recommend installing Mambaforge which is a conda installation with mamba installed by default and set to use conda-forge as the default set of package repositories. 
+This fork uses [uv](https://docs.astral.sh/uv/). From the repo root:
 
-2. Clone the git repo:
 ```
-git clone git@github.com:AnavSood/CSS.git
+uv sync
+uv run pytest tests/
 ```
 
-3. Set up your conda environment. The list of packages that will be installed inside your conda environment can be seen in `environment.yml`.
-```
-mamba update -y conda
-# create a development virtual environment with useful tools
-mamba env create
-conda activate CSS
-```
+`pycss` itself is pure Python. The original Cython
+[choldate](https://github.com/modusdatascience/choldate) dependency is
+replaced by a pure-NumPy implementation in `pycss/chol.py`; the original
+package is vendored under `vendor/choldate/` (with the build fixes from
+[choldate PR #8](https://github.com/modusdatascience/choldate/pull/8)) as a
+dev-only dependency so the replacement is tested against it
+(`tests/test_chol.py`). Reference benchmarks live in `benchmarks/`.

benchmarks/RESULTS.md

@@ -0,0 +1,25 @@
+# Reference Benchmarks: cholupdate
+
+**Date:** 2026-07-18
+
+**Machine:**
+- Architecture: arm64
+- OS: macOS 26.5.2
+- Python: 3.12.13
+
+## Results
+
+```
+cholupdate micro-benchmark (mean per call)
+     p   numpy (ms)  choldate (ms)  numpy/choldate
+    10       0.0278         0.0006            49.7
+    50       0.1346         0.0022            62.4
+   100       0.2716         0.0072            37.9
+   250       0.7090         0.0359            19.7
+   500       1.5132         0.1415            10.7
+  1000       3.4433         0.5547             6.2
+
+swapping_css end-to-end (p=100, k=10, single init)
+numpy backend:    0.005 s
+choldate backend: 0.003 s
+```

benchmarks/bench_chol.py

@@ -0,0 +1,66 @@
+"""Reference benchmark: pure-NumPy cholupdate vs vendored Cython choldate.
+
+Run from the repo root: uv run python benchmarks/bench_chol.py
+Numbers are for reference only; optimization is not a project goal.
+"""
+import time
+
+import numpy as np
+
+import pycss.subset_selection as ss
+from pycss.chol import cholupdate as np_cholupdate
+
+try:
+    from choldate import cholupdate as c_cholupdate
+except ImportError:
+    c_cholupdate = None
+
+
+def bench_update(fn, p, reps=200, seed=0):
+    rng = np.random.default_rng(seed)
+    M = rng.standard_normal((p, p))
+    A = M @ M.T + p * np.eye(p)
+    R0 = np.ascontiguousarray(np.linalg.cholesky(A).T)
+    xs = rng.standard_normal((reps, p))
+    start = time.perf_counter()
+    for i in range(reps):
+        fn(R0.copy(), xs[i].copy())
+    return (time.perf_counter() - start) / reps
+
+
+def bench_swapping(backend, p=100, k=10, seed=0):
+    """Time swapping_css with pycss.subset_selection's cholupdate swapped out."""
+    rng = np.random.default_rng(seed)
+    M = rng.standard_normal((2 * p, p))
+    Sigma = M.T @ M / (2 * p)
+    original = ss.cholupdate
+    ss.cholupdate = backend
+    try:
+        start = time.perf_counter()
+        ss.swapping_css(Sigma, k, S_init=np.arange(k))
+        return time.perf_counter() - start
+    finally:
+        ss.cholupdate = original
+
+
+def main():
+    print("cholupdate micro-benchmark (mean per call)")
+    print(f"{'p':>6} {'numpy (ms)':>12} {'choldate (ms)':>14} {'numpy/choldate':>15}")
+    for p in [10, 50, 100, 250, 500, 1000]:
+        t_np = bench_update(np_cholupdate, p) * 1e3
+        if c_cholupdate is not None:
+            t_c = bench_update(c_cholupdate, p) * 1e3
+            print(f"{p:>6} {t_np:>12.4f} {t_c:>14.4f} {t_np / t_c:>15.1f}")
+        else:
+            print(f"{p:>6} {t_np:>12.4f} {'n/a':>14} {'n/a':>15}")
+
+    print("\nswapping_css end-to-end (p=100, k=10, single init)")
+    t_np = bench_swapping(np_cholupdate)
+    print(f"numpy backend:    {t_np:.3f} s")
+    if c_cholupdate is not None:
+        t_c = bench_swapping(c_cholupdate)
+        print(f"choldate backend: {t_c:.3f} s")
+
+
+if __name__ == "__main__":
+    main()