Release 1.1.0: changelog, version bump, and a publish workflow (#75)

precise had no publish workflow, so a GitHub release did nothing. This mirrors
humpday: build, check, test-install the wheel, then upload, with trusted
publishing over OIDC rather than a stored token. PyPI needs a publisher
configured for this repository, the publish.yml workflow and the pypi
environment before an upload will be accepted.

1.1.0 rather than 1.0.1: three new estimators is a feature release. No public
API was removed, though two behaviours change -- suggest() recommends
differently now that the frozen model is not inert, and LedoitWolfCovariance no
longer collapses under heavy tails.

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: microprediction

.github/workflows/publish.yml

@@ -0,0 +1,76 @@
+name: Publish to PyPI
+
+# Fires when a GitHub release is published. Build and test-install happen before anything is
+# uploaded, so a wheel that cannot be installed never reaches PyPI. Authentication is trusted
+# publishing (OIDC) rather than a stored token, which is why the publish job needs id-token: write
+# and an environment; PyPI must have a publisher configured for this repository, workflow and
+# environment name for the upload to be accepted.
+on:
+  release:
+    types: [published]
+  workflow_dispatch:
+    inputs:
+      test_pypi:
+        description: 'Deploy to Test PyPI instead of PyPI'
+        type: boolean
+        default: false
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install build dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install build twine
+      - name: Build package
+        run: python -m build
+      - name: Check package
+        run: twine check dist/*
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  test-install:
+    needs: build
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - name: Test installation from wheel
+        run: |
+          pip install dist/*.whl
+          python -c "import precise; print(f'precise {precise.__version__} installed')"
+          python -c "from precise import all_estimators; print(f'{len(all_estimators())} estimators')"
+
+  publish:
+    needs: [build, test-install]
+    runs-on: ubuntu-latest
+    environment:
+      name: ${{ github.event.inputs.test_pypi == 'true' && 'test-pypi' || 'pypi' }}
+      url: ${{ github.event.inputs.test_pypi == 'true' && 'https://test.pypi.org/p/precise' || 'https://pypi.org/p/precise' }}
+    permissions:
+      id-token: write  # mandatory for trusted publishing
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - name: Publish to Test PyPI
+        if: github.event.inputs.test_pypi == 'true'
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://test.pypi.org/legacy/
+      - name: Publish to PyPI
+        if: github.event.inputs.test_pypi != 'true'
+        uses: pypa/gh-action-pypi-publish@release/v1

CHANGELOG.md

@@ -4,7 +4,7 @@ All notable changes to `precise` are documented here. The format follows
 [Keep a Changelog](https://keepachangelog.com/), and the project aims to follow
 [Semantic Versioning](https://semver.org/).
 
-## [Unreleased]
+## [1.1.0] — 2026-09-13
 
 ### Added
 - `NonlinearShrinkageCovariance`: online analytical **nonlinear** shrinkage of the covariance
@@ -25,6 +25,31 @@ All notable changes to `precise` are documented here. The format follows
   distribution is not an equivalence, and Oriol (arXiv:2410.14420) derives the weighted formulas
   properly — but it has no window boundary, so it does not pay for a shock twice.
   `research/turnover.py` measures the difference.
+- A JOSS paper under `papers/joss/`, and the note *Spectral Calibration Without a Split* under
+  `papers/online_spectral_calibration/`.
+
+### Fixed
+- `LedoitWolfCovariance` collapsed to a scaled identity under heavy tails. `pi_bar` averages a
+  quantity growing like the fourth power of the observation, so one fat-tailed draw could pin the
+  shrinkage intensity at 1 and leave the estimate with no off-diagonals: at t(3) innovations it
+  retained 0.088 of the covariance structure where OAS retained 0.543. Each observation's
+  contribution is now winsorized at ten times the running mean, which leaves Gaussian behaviour
+  unchanged to three decimals and takes t(3) retention to 0.409.
+- The frozen recommender was inert. `sklearn`'s `tree_.value` holds class proportions, and the
+  exporter cast them with `int()`, flooring every value under 1.0 to zero — 46 of 47 nodes carried
+  no weight, so `suggest()` had been ranking on the heuristic ruleset alone. The model now covers
+  19 of 20 estimators (it was 9), and two training runs produce a byte-identical artifact.
+- Training was irreproducible whenever `randomcov` was installed: three generative ensembles in
+  `research/oos.py` ignored the `rng` they were passed and drew from global state.
+- The training grid stopped at `n/p = 3`, so every data-rich low-dimensional problem was
+  extrapolation. It now spans `p` from 5 to 60 and `n/p` from 0.5 to 25.
+- A malformed generated model can no longer break `import precise`.
+
+### Changed
+- `suggest()`'s safe default moves from `LedoitWolfCovariance` to `NonlinearShrinkageCovariance`,
+  which has the best mean rank of any single fixed choice over eleven ensembles and seven `(p, n)`
+  regimes. It is not uniformly best: at `p` close to `n` it ranks 9.41 and the trained model is
+  worth far more there, which is why the model still leads and this only breaks ties.
 
 ## [1.0.0] — 2026-06-05
 

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "precise"
-version = "1.0.0"
+version = "1.1.0"
 description = "Online (incremental) covariance and correlation estimation — the online complement to sklearn.covariance"
 readme = "README.md"
 requires-python = ">=3.9"