Scope loop flags and specialization thisType override to local contexts (A6 phase 2a)

[ASDAlexander77]

Summary

A6 phase 2 (copy-on-override), first field group — and both converted sites turned out to be live examples of the latent-leak bug class the plan predicted.

• Loop flags. DoStatement/WhileStatement/ForStatement now set isLoop/loopLabel on a local loopGenContext passed to their condition/body/incrementor generation (including the for await async-execute body, which now copies from the loop context). Previously they mutated the caller's context through const_cast and never restored — after the first loop in a function, isLoop stayed true for all subsequent code, affecting the mlirGenDisposable LoopScope walk.
• instantiateSpecializedFunction. The arrow/function-expression thisType = nullptr protection now applies to a function-local context copy (the whole body uses instantiateGenContext); previously the null leaked into the caller's context permanently, which could break this access in code following a generic-arrow instantiation inside a method.

Deliberately unchanged: loopLabel is still assigned after label was cleared (always empty — pre-existing oddity, preserved; the label still reaches the loop op via LABEL_ATTR_NAME). Remaining const_cast sites are the mailbox-pattern fields (generatedStatements ×4, typeAliasMap ×2, mlirGenParameters thisType) for the next slice — those intentionally communicate upward and need honest signatures rather than copies.

Test plan

• tslang rebuilt cleanly (vs2026 debug)
• New smoke test covering for + break/continue, while + break, do-while, and nested labeled continue outer — JIT output 12 4 3 3 as expected
• All eight prior smoke tests identical
• MLIRGenTests.exe: 12/12 pass

🤖 Generated with Claude Code