Share generated interface member descriptors

[leileizhang (lei9444)]

Summary

• Add the opt-in --shared-interface-members JavaScript codegen flag.
• Reuse canonical required-interface prototype descriptors on compatible concrete classes.
• Keep conflicts, overloads, ambiguous interface identities, and one-off interfaces class-local.
• Preserve the default generation path, raw interface APIs, deep imports, and mixed-flag incremental safety.

Compatibility

The implementation uses one canonical interface identity for source emission and descriptor copying. It falls back to class-local members for ambiguous flat filenames and rejects incompatible incremental overwrites.

When sharing replaces an inline raw-interface declaration, the class deep module preserves the same public JS/TypeScript import through an equivalent canonical re-export. Declaration APIs and import paths are preserved, but flag-on .d.ts files are not byte-identical. Flag-off/default output remains byte-compatible.

The option is JavaScript-only and disabled by default.

Corrected isolated Gallery A/B

Measured with the latest branch, the same 32 Gallery WinMDs and 99 references, and no bundle post-processing:

Metric	Baseline	Shared	Delta
Total generated JS	48,563,986 B	48,335,321 B	-0.47%
Per-type JS	48,287,170 B	48,058,505 B	-0.47%
First-screen closure JS	13,373,627 B	13,568,182 B	+1.45%
First-screen closure modules	882	952	+70
DTS bytes	11,006,778 B	10,880,821 B	-1.14%
Static require edges	27,963	28,257	+294
Cyclic SCCs	39	59	+20

File sets and all 2,781 root exports are preserved. Deep modules lose no baseline exports; two modules gain canonical raw-interface re-exports. TypeScript root/deep import validation passes and unsafe copied sources are zero.

Validation

• Default JS/Python output matches the base branch.
• Uri deep IStringable JS/DTS imports are preserved, including Node 22 ESM named imports.
• Shared-to-unflagged incremental generation preserves existing source markers.
• Duplicate same-name TypeDefs retain distinct IIDs for ambiguity planning.
• Gallery Binding/Router smoke and E2E: 37/37 passed.
• Full cargo test -p dynwinrt-codegen passed.

Recommendation

Correctness and compatibility issues are resolved, but the final measured benefit is negligible and the first-screen dependency graph becomes larger and more cyclic. Based on the current data, this change does not provide enough value to justify merging in its present form.

[github-actions]

Mixed-language test coverage

Workflow status: ✅ Passed

Layer	Lines	Functions	Branches/regions
Rust, including native .pyd/.node	81.05%	77.58%	81.02% regions
Generated Python projections	77.01%	n/a	62.35% branches
JavaScript aggregate	18.21%	19.53%	50.72% branches
JavaScript runtime	19.36%	14.28%	7.14% branches
Generated WinRT projections	22.3%	19.96%	52.8% branches
Generated Classic COM projections	11.25%	18.08%	45.94% branches

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