Fix explicit WinRT delegate selection for JavaScript and Python

[leileizhang (lei9444)]

Problem

Explicitly selecting ThreadPool,WorkItemHandler classified the delegate as an ordinary runtime class. Its placeholder displaced the correct dependency module: JavaScript rejected normal callbacks with Failed to recover WinGUID, and Python rejected them as non-DynWinRTValue inputs before native invocation.

Fix

Classify non-generic WinRT delegate roots from their metadata kind/base, public WinRT flags, typed IID, and Invoke contract before class fallback. Reuse the existing shared delegate/interface dependency parser for both languages and include standalone delegates in namespace selection. Malformed/non-WinRT/open-generic roots fail explicitly; public-interface and Classic COM routing remain unchanged. The delegate fix does not change runtime, ABI, lifetime, renderer, release, or unrelated projection behavior.

Regression coverage includes delegate-only, combined/reversed roots, automatic dependencies, repeat/incremental selection, namespace roots, ThreadPoolTimer's delegate family, malformed contracts, JS CJS/ESM and declarations, and Python with/without stubs. CLI help and codegen documentation now describe delegate selection.

Separate CI compatibility prerequisite: the additive commit copies only the approved tiny Rust 1.99 fixture-format correction from 49c7e6e2e7042b527f1fee9d6655a8df8f537099: allow a trailing comma in the test-only stub! matcher and format its two existing audio invocations. The one-file delta is byte-identical to that source commit. Both invocation argument lists and the emitted extern "system" function signature/E_NOTIMPL body are unchanged. No functional commit or dependent branch is imported; the original main base is retained.

Verification

• 33 focused codegen/delegate/snapshot/raw-signature tests pass; strict TypeScript, mypy and Pyright callable consumers include positive and rejection cases. All eight new regressions pass again with the additive prerequisite.
• Fresh main generator and matched production bindings reproduce both original failures with zero callbacks. Fixed explicitly generated bindings execute and await a real OS ThreadPool callback in JS CJS/ESM and Python on ARM64 and x64 processes under Windows ARM64 emulation. Child processes are timeout-bounded; Python uses RoApartment() and projected_lifetime_scope().
• Sixteen .js/.d.ts/.py/.pyi comparisons for ThreadPool/WorkItemHandler and ThreadPoolTimer/TimerElapsedHandler are byte-identical to fresh-main automatic dependency output. Existing JS/Python delegate runtime selectors pass; local Rust 1.97/1.98.1 formatting is clean.
• Full local codegen run: 1,028 passed, one ignored, one existing long-path mypy import failure. The exact-main and fixed generators produce byte-identical diagnostics for that failure at the same original test path (82 relevant stubs identical); both pass in a short-path control. No unrelated projection fix included.

Independent base: de91e9b3889aafdda44098d5ea1b731f9976a3c6. Metadata: Windows SDK 10.0.26100.0, Windows.winmd SHA-256 11191cbcc796e520d47547b3f986be94a48556fe336d4e63334596d9f31a7806. Local native evidence uses Node 24.15.0 and CPython 3.13.15; hosted x64 CI results are separate.

[github-actions]

Mixed-language test coverage

Workflow status: ✅ Passed

Layer	Lines	Functions	Branches/regions
Rust, including native .pyd/.node	86.95%	82.47%	86.66% regions
Python aggregate	72.34%	n/a	39.42% branches
Python runtime	98.17%	n/a	94.57% branches
Generated Python WinRT projections	71.07%	n/a	33.5% branches
Generated Python WinRT implementations	71.97%	n/a	46.3% branches
JavaScript aggregate	21.91%	25.18%	57.61% branches
JavaScript runtime	44.27%	45.76%	78.99% branches
Generated WinRT projections	22.8%	18.7%	54.84% branches
Generated WinRT implementations	45.99%	59.19%	60.97% branches
Generated Classic COM projections	11.88%	23.97%	53.4% branches

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