PanorAi 3.6.0 release checklist

This minor release adds opt-in Experimental spherical deep-learning model-port surfaces and refreshes the package summary. Stable geometry-v1, object workflow, spherical-feature core, projection, blending and every 3.x compatibility name remain unchanged.

Candidate identity

  • [ ] One exact reviewed protected-main commit contains every release input and has no uncommitted or untracked release changes.

  • [ ] CHANGELOG.md, package metadata, release workflow, Pages recovery, manifest, artifact policy and contract tests identify exactly 3.6.0 / v3.6.0.

  • [ ] A fresh review covers the complete v3.5.0..candidate diff and verifies that every new deep-learning surface remains Experimental and opt-in.

  • [ ] Immutable published tags and index files through 3.5.0 remain unchanged.

Public contract and dependency isolation

  • [ ] Stable/Compatibility inventories pass with no removed or re-tiered names.

  • [ ] import panorai, import panorai.geometry and import panorai.experimental do not load Torch, Torchvision or external model code and cause no network/cache side effect.

  • [ ] panorai.experimental.deep_learning is the only opt-in import boundary; Torchvision loads only when a pretrained ImageNet helper is called.

  • [ ] deep-learning and deep-learning-depth remain optional extras; base dependencies and supported Python 3.11–3.14 boundaries are unchanged.

Model portability and safety

  • [ ] Analytic spherical convolution covers exact 1x1 equivalence, parameter identity, seam/longitude equivariance, finite poles, shapes and autograd.

  • [ ] Planar FCN conversion reproduces the original AlexNet, VGG16 and ResNet18 classifier logits within the documented tolerance using official external weights.

  • [ ] The metric-depth loader requires explicit terms acceptance, pins source and checkpoint identities, verifies size/SHA-256, rejects traversal/links and uses torch.load(..., weights_only=True).

  • [ ] The real external model port accounts for all 45 Conv2d, three ConvTranspose2d and four absorbed reflection pads while preserving learned parameter identity and leaving no learned planar spatial layer.

  • [ ] No accuracy, semantic-equivalence, spherical-equivariance or suitability claim is inferred from mechanical model loading and parameter preservation.

Private-data and artifact boundary

  • [ ] No private dataset name, identifier, byte, prediction, metric, result, generated visualization or narrative appears in tracked release content, GitHub source archive, wheel, sdist, documentation output or release notes.

  • [ ] No checkpoint, model archive, extracted third-party source, cache, manifest, generated CAM or prediction appears in the source archive, wheel or sdist.

  • [ ] Legacy vendored model implementations, research training/data helpers, generated API stubs and obsolete root scaffolding are absent from the source tree and source archive.

  • [ ] The unresolved checkpoint-license boundary is explicit: opt-in is not a license grant and PanorAi never redistributes the checkpoint.

  • [ ] Wheel and sdist content/license audits require every new PanorAi-owned source module and reject prohibited model/data extensions.

Source, documentation and artifact validation

  • [ ] Full warnings-as-errors tests pass on Python 3.11–3.14; Torch CPU runs geometry, FCN/CAM, acquisition-contract and metric-depth adapter tests.

  • [ ] Strict Sphinx builds without warnings and executable examples pass.

  • [ ] Both native extensions build and execute on Linux x86_64/aarch64, macOS Intel/arm64 and Windows AMD64 for Python 3.11–3.14.

  • [ ] The hosted aggregate contains exactly 20 wheels and one normalized sdist; every file passes Twine, checksum, metadata, content/license and provenance audits.

  • [ ] A fresh compatible installed wheel outside the checkout passes version, origin, native, import-isolation and --deep-learning smoke tests without downloading a model.

  • [ ] A distinct fresh review reproduces the artifact inventory and installed evidence from the exact candidate.

Publication sequence

  1. Merge only after every hosted PR job passes and repeat the complete matrix on the protected-main merge commit.

  2. Record GO-CANDIDATE only after exact merged-source, aggregate-artifact and installed-wheel review.

  3. Change only the three protected environment selectors from v3.5.0 to v3.6.0; preserve reviewers, administrator enforcement, custom policies and Trusted Publishers.

  4. Create one signed annotated v3.6.0 tag on the exact candidate and publish one GitHub Release, the sole trigger of python-publish.yml.

  5. Verify TestPyPI serves the exact workflow-built files and run installed native, geometry, Torch, deep-learning and documentation checks.

  6. Promote the identical files to PyPI, attach checksums/provenance and deploy Pages; no manual upload may bypass a failed gate.

  7. Install panorai[deep-learning]==3.6.0 from production PyPI in a fresh environment and verify exact artifact identity and installed smoke without acquiring external model files.

Published tags and index files are immutable and may not be moved, deleted, recreated or replaced.