PanorAi 3.4.0 release checklist¶

This minor release publishes direct spherical dense stereo as an explicitly Experimental API. Stable geometry-v1, projection, feature and object-workflow contracts remain unchanged.

Candidate identity¶

[ ] The exact release commit contains the reviewed stereo implementation, visualization helpers, generated documentation assets, release metadata and no uncommitted files.
[ ] The changelog uses the intended publication date and every tracked workflow, recovery default, fallback and artifact policy identifies 3.4.0 / v3.4.0. Protected environment selectors remain on the immutable previous release until a separate publication authorization is recorded.
[ ] A fresh review checks the complete v3.3.1..candidate diff and records any limitations without promoting Experimental behavior to Stable.

Experimental stereo contract¶

[ ] The public pose convention is X_b = R_b_from_a @ X_a + t_b_from_a; output is radial range and inherits translation units.
[ ] Every range candidate is pose-constrained, horizontal ERP sampling wraps, invalid pixels are explicit and A→B→A consistency is tested.
[ ] Analytic range, seam, validation, immutability, visualization and frozen export-inventory tests pass with warnings as errors.
[ ] The ten-case study records exactly five Matterport360 and five Stanford2D3D pose-success pairs, selection hashes, per-case panels, raw arrays, accuracy and coverage while stating that reference baseline magnitude supplies metric scale.
[ ] Matterport360/Stanford2D3D media and local study outputs are absent from wheel, sdist and published documentation.

Documentation and media¶

[ ] Strict Sphinx copies and renders the method diagram and the real CC0 photograph/analytic-geometry diagnostic panel without warnings.
[ ] Tutorial signatures, parameters, result fields, example, pose convention, failure behavior, evaluation limitations and future C++ boundary match the published code.
[ ] The original source image provenance and CC0 license remain bundled; the generated PNG has the reviewed SHA-256 enforced by the artifact auditor.
[ ] Pages is deployed from the exact signed release tag, never from a mutable branch or a locally built documentation tree.

Source, wheel and platform validation¶

[ ] Full warnings-as-errors tests pass on Python 3.11–3.14, including NumPy, Torch CPU/autograd, geometry conformance/oracle, multimodal flows and the new stereo suite.
[ ] Strict documentation and executable examples pass from source, installed wheel and extracted sdist where applicable.
[ ] Both existing native extensions build and execute on the complete Linux, macOS and Windows wheel matrix; the Python/OpenCV stereo module is present in every wheel without introducing a new compiled extension.
[ ] The hosted aggregate contains exactly 20 wheels and one normalized sdist; every file passes metadata, license, content, provenance and checksum audits plus twine check.
[ ] Fresh installed-wheel smoke imports and runs panorai.stereo outside the checkout and confirms optional Torch/Open3D remain unloaded.

Publication sequence — separate authorization required¶

Merge the reviewed PR only after all required hosted checks pass.
After explicit publication authorization, read back the three protected environment policies and change only their exact tag selector from v3.3.1 to v3.4.0.
Create a signed annotated v3.4.0 tag on the exact GO-CANDIDATE merge.
Publish one GitHub Release; it is the sole trigger of python-publish.yml.
Independently inventory the workflow-built artifacts and attestations.
Verify TestPyPI serves byte-identical files and installed validation passes.
Allow the workflow to promote those same files to PyPI and deploy Pages.
Install panorai==3.4.0 from production PyPI in a clean environment and repeat public smoke, native-extension and stereo checks.

No manual upload may bypass the release workflow or protected environments.