G# + cs2gs: generalized variadic carriers — ...X[T] ≡ C# params X<T> (ADR-0173, #3627)

[DavidObando]

Fixes #3627 — the params ReadOnlySpan<T> cs2gs gap, the last Cs2Gs.Tests translate wall in the #3501 selfmig nightly. Implements the agreed scope: ...X[T] becomes semantically equivalent to C#13 params X<T>, with no magic span types or new sugar (design for those stays in #3627's discussion).

Semantics (ADR-0173)

The type written after ... is interpreted like C#'s params type:

• Carrier shapes become the parameter's exact CLR type: ...[]T (explicit array), ...List[T], the five IEnumerable-family interfaces (IEnumerable/ICollection/IList/IReadOnlyCollection/IReadOnlyList), ...Span[T], ...ReadOnlySpan[T].
• Anything else keeps the ADR-0101 element meaning: ...int32 ≡ params int[]; ...HashSet[int32] is still a params array of HashSets (HashSet is not a supported carrier).

Call sites mirror C#: expanded trailing args are element-coerced (#1493 rules) and packed — array as-is, interface upcast, new List[T](T[]), span T[] ctor (heap-allocating v1); a single trailing arg already convertible to the carrier passes through unchanged. Parser needed zero changes.

Metadata: array family keeps [ParamArrayAttribute]; every other carrier stamps C#13's [ParamCollectionAttribute], so csc consumers see a genuine params X<T> member (verified by reflection + runtime invoke witness).

New GS0544 when a List/span carrier's element is a same-compilation erased type with no closed CLR construction shape (the array carrier still covers that case).

cs2gs

• params X<T> declarations translate to ...X[T] for supported carriers — spans included, retiring the gap.
• Expanded calls on source-declared callees stay in natural form (gsc packs); referenced/BCL span callees keep the pre-existing lowerings (import-side [ParamCollection] expanded calls are a recorded follow-up; pass-through works).
• Array params with carrier-shaped elements print explicit spellings (params byte[][] → ...[][]uint8, params List<int>[] → ...[]List[int32]) so they don't reinterpret.
• The deliberately untranslatable params HashSet<int> fixture in Cs2Gs.Tests is now compiled in-memory by its test, so the project's own source is fully translatable.

Breaking-change note

Carrier-shaped ...X spellings are reinterpreted (previously: params array of X elements). No in-tree G# used such spellings; the escape hatch for the old meaning is ...[]X. Bare ...T for non-collection T is unchanged.

Coverage

One shared resolver (VariadicCarriers) feeds all eight ADR-0101/0102 declaration sites: free functions, methods, constructors, primary constructors, interface methods, named delegates, lambdas, and function-type clauses.

Verification

• New Issue3627VariadicCarrierTests (13 binder/eval tests) + Issue3627VariadicCarrierEmitTests (attribute reflection, ILVerify, E2E run) — all green.
• Full Core.Tests locally: 8227/8227 passed.
• Cs2Gs.Tests: full run surfaced two pinned-old-behavior tests (Issue1727 params byte[][], Issue3466 params-collection demotion) — first exposed a real translator bug (array-of-array spelling), both fixed and green; Issue1901/Issue3460 suites green.
• New conformance sample samples/VariadicCarriers.gs + golden; PE baseline regenerated (+1 entry).
• Local selfmig proof over the cs2gs subtree: Cs2Gs.Tests translate stage now PASSES (both prior fingerprints — the ReadOnlySpan gap and the fixture — gone). Remaining local compile-stage failures are the known Debug-SDK staleness artifact of skipping the gate script's prerequisite builds, identical fingerprints before/after this change; the nightly gate is the authoritative check.

Docs

ADR-0173 (new), GS0544 in diagnostics docs + website mirror, spec variadic section, coverage-matrix inventory row (regenerated via cs2gs coverage --write).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Nng28yiBdPVdeML7mSZphs