Support metadata-validated POD WinRT vectors

[leileizhang (lei9444)]

Summary

Partially addresses #161: phase-one support for empty and populated checked POD WinRT vectors, not completion of the broader struct-collection issue.

• Validate closed metadata types, exact element identity/IIDs, and recursive natural POD layouts before publication; reject owned or incomplete layouts.
• Add aligned owned POD storage and cache immutable vtable plans using the existing central libffi callback machinery. Cover IVector<T>, live/snapshot IVectorView<T>, IIterable<T>, iterators, bulk operations, and observable/reentrant lifetime paths, including ARM64 HFAs and larger/nested aggregates.
• Preserve word-sized scalar/reference/HSTRING fast paths, aliases, field-wise equality (padding ignored, signed zeros equal, NaNs unequal), ownership, and null distinctions. Map admission, metadata-free unsafe-constructor contracts, owned structs, and top-level scalar limits remain unchanged.

No public facade, codegen implementation, dependency, or release-workflow changes. The existing native_callback implementation is reused unchanged. See docs/architecture/winrt-pod-vectors.md for the design and boundaries.

Local validation and revision attribution

Implementation validated at 0ef52e6. Follow-up bb153fb changes only release wording in four documentation files; git diff --check passed. The original three implementation/coverage commits were not rebased or rewritten, and historical evidence manifests remain unchanged.

Validation at the implementation revision	Result
Native ARM64 cargo test --locked --offline -p dynwinrt	503 unit + 17 metadata-lifetime + 1 WinRT regression + 4 doctests passed; 1 existing WinAppSDK initialization ignore
Focused collection/map/callback tests, ARM64 optimized	63 passed
Same focused tests, Windows x64 / i686 emulation on ARM64	66 / 63 passed; not native x64/i686 hardware qualification
Selected codegen snapshot/collection/observable/preparation tests	15 passed, not the full codegen suite
Matched Python WinRT/runtime regressions	92 passed
Matched JS scoped collection/root tests and generated/native-class regressions	10 + 21 passed; 1 existing x64-only map-test skip
Public JS typechecks and existing TS runner tests	Passed; 2 runner tests

Matched generated JS/Python E2E passed stock InkStrokeBuilder Point inputs and all three Geopath/GeoboundingBox BasicGeoposition overloads. The optional pod_hint_specs.json fixture exercises generated RectInt32 vector preparation only and remains explicitly documented as such. Empty producers remain real non-null collections; native consumers retain their own empty-input semantics.

Additional local ARM64 gallery trial

A later user-confirmed trial used the exact release addon from 0ef52e69 (SHA256 D214D97ACB6BFBB61DA14F6871138C325E93F5754DBC84F4BEB8AF55F7EB0F94) with unchanged f50fd14 codegen, all 132 WinMD hashes, and all 1,313 generated files. Real empty RectInt32 input matched the prior null control (850x566, 481,100 mask bytes; 401,288 opaque / 79,812 transparent). A nonempty rectangle (225, 20, 400, 540) passed field roundtrip and native AI mask execution, and the renderer → IPC → utility → native → canvas route also passed. Missing-image input still rejected with 0x80070002.

This is local evidence for those gallery paths, not general AI/UI or WinAppSDK qualification; it does not change the preparation-only test fixture's scope.

Release admission

Proposed for this release, subject to PR review, final integration, and platform CI; not an already-released capability or a claim that CI is green. A local merge-tree check against main at 6a9bb4ebf57b3a9cd81891decc3f7701ee08e56b was conflict-free, but the merge result has not been requalified by the historical local suites.

Coverage follow-up: 7da277c

The failed coverage run reached the final Python aggregate gate: 7,220 / 10,340 lines = 69.83%, below the unchanged 70% minimum. It was not a build or test-assertion failure. Adding the POD E2E inputs expanded the measured generated projection set; a diagnostic control without those two new specs measured 71.06%.

7da277c changes only the existing Python geolocation E2E case: constructor/factory parity, BasicGeoposition view bulk reads and lookup, shape/spatial/altitude properties, and reconstructed bounding-box checks. No production code, workflows, coverage filters, thresholds, or existing assertions were changed.

The local Python coverage slice used the official codegen artifact for CI merge 5215faa, source-matched native runtime, and local SDK 26100 metadata. The before result matched CI's numerator and denominator exactly; after the assertions it measured 7,263 / 10,348 = 70.19%, passing the unchanged validator. All 44 standard Python E2E specs and 4 + 19 implementation scenarios passed, as did binding tests (14 existing unavailable WinUI-fixture skips), the focused two POD specs, threshold regression tests, Python compilation, and git diff --check.

This reproduction was native ARM64 / CPython 3.13.15 / cached coverage.py 7.16.0; the failed hosted run used x64 / CPython 3.12.10 / coverage.py 7.16.1. No new remote CI success is claimed; the follow-up still requires its normal CI checks.

[github-actions]

Mixed-language test coverage

Workflow status: ✅ Passed

Layer	Lines	Functions	Branches/regions
Rust, including native .pyd/.node	86.41%	81.54%	86.07% regions
Python aggregate	70.19%	n/a	36.95% branches
Python runtime	99.12%	n/a	97.37% branches
Generated Python WinRT projections	68.99%	n/a	27.3% branches
Generated Python WinRT implementations	72.3%	n/a	47.04% branches
JavaScript aggregate	21.98%	24.37%	57.89% branches
JavaScript runtime	44.27%	45.76%	78.99% branches
Generated WinRT projections	22.55%	18.39%	55.34% branches
Generated WinRT implementations	48.66%	59.76%	61.14% branches
Generated Classic COM projections	11.92%	23.97%	54.2% branches

View workflow run and download full HTML/LCOV/XML reports