Fix #661/#658 (inferred class method return propagation), #593 (class-instance iterability)

[nickna]

Completes four related generator/iterator type-checker issues.

#661 / #658 — inferred class method return types now reach call sites

CheckClassBody resolves a method's inferred (un-annotated) return type during the body pass and writes it to the mutable class, but the class had already been frozen with <inferred> placeholders before that pass — so new C().m() read <inferred> (~any). Now, when any inferred method return was resolved, the frozen Class/GenericClass (and the TypeMap the compiler reads) are rebuilt from the resolved mutable state and re-published after the body pass. The structural compatibility cache is cleared afterward so results computed against the placeholder (keyed by the stable DeclarationId) aren't reused.

• class C { *m() { yield 1; } } → [...new C().m()] now type-checks (was "must be an iterable type … got ''") — the Type checker: class generator method without an explicit return type is treated as non-iterable (both modes) #661 headline.
• class C { name() { return "hi"; } } → const n: number = new C().name() is now the TS2322 tsc reports — Inferred class method return types leak <inferred>/any at call sites (not propagated) #658, the non-generator face of the same gap.
• Covers static / async / private methods and recursion.

#593 — class-instance iterability

A class is iterable only via an inherited [Symbol.iterator]. extends Array<number> resolved Array to a name-only placeholder that dropped the element type, so SharpTS bound any for a plain instance in for...of (under-rejecting) and threw a spurious TS2488 for an extends Array instance in yield* (over-rejecting).

• extends Array/Set/Map/String/TypedArray records the element type as an @@iterator on the placeholder; for...of / yield* / spread now bind the real element type (number, not any).
• A plain class instance (no inherited [Symbol.iterator]) is now TS2488 — matching tsc and the runtime, which already throws "for...of requires an iterable". The #592-dependent direction (user [Symbol.iterator](), which doesn't parse yet) is out of scope and forward-compatible.

#487 / #532 — already fixed by 533d0cf; regression coverage added/confirmed

Both were resolved by commit 533d0cf (#548/#532/#487/#483). Added explicit Generator/AsyncGenerator 3-arg lib-spelling regression tests (#487); #532's nested-generator tests already exist. Verified both repros pass in both modes.

Issues filed for pre-existing gaps surfaced here

• Array destructuring desugars to index access, so it fails over non-indexable iterables (generators, Set, Map) #685 — array destructuring desugars to index access, so it fails over non-indexable iterables (generators, Set, Map). Surfaced because Type checker: class generator method without an explicit return type is treated as non-iterable (both modes) #661 now produces a real Generator<T>; independent of it (free function* and explicitly-typed generators fail identically).
• Compiled mode: static generator method (static *m()) fails with "Yield not supported in this context" #692 — compiled static generator method (static *m()) fails "Yield not supported in this context" (pre-existing lowering gap; an explicit return type fails identically).

Testing

• Full suite: 12714 / 0 (incl. 24 new regression tests).
• TSConformance: 31 / 31 (no baseline change).
• Test262: unaffected — all changes are type-checker-only and the harness runs .js untyped by default. Corpus isn't initialized in this worktree.
• Updated MethodCompletionValueTests to probe the undefined sentinel through as any, since an off-the-end method now correctly infers void and void + 10 is the TS2365 tsc reports.

Fixes #661
Fixes #658
Fixes #593
Closes #487
Closes #532