Patch declaration parents in ClassSpecializer

[chaudhary-lakshay]

Fixes #284.

Problem

@Burst on a class whose constructor parameter defaults to burstValues(...) containing lambdas fails to compile:

Platform declaration clash: The following declarations have the same JVM signature
(access$ExampleUnitTest$_init_$lambda$0()Lkotlin/Unit;)

Cause

BurstValuesArgument.expression() deep-copies each value with parameter.parent as the new parent, which for a constructor parameter is the original constructor:

https://github.com/cashapp/burst/blob/trunk/burst-kotlin-plugin/src/main/kotlin/app/cash/burst/kotlin/Argument.kt#L112

ClassSpecializer then inserts those copies into the generated subclass constructors and the generated no-args constructor, but never repairs parents afterward. FunctionSpecializer does call patchDeclarationParents() on its generated functions, which is why the same burstValues with lambdas works fine on a function parameter and only breaks on a constructor parameter.

The result is that every copied lambda's local IrSimpleFunction still claims TestClass.<init> as its parent, so lowering mints the same synthetic accessor name once per specialization.

Why it appeared in 2.11.0

The bug is latent, not new. Between 2.10.2 and 2.11.0, ClassSpecializer.kt is unchanged and the Argument.kt diff is only an import move plus formatting. The one substantive change is kotlin = "2.2.20" → "2.3.0"; Kotlin 2.3's lambda naming stopped tolerating the wrong parent.

On current trunk (Kotlin 2.4.10) the IR validator catches it one phase earlier, so the same test data fails with:

error: the compiler plugin 'app.cash.burst.kotlin.BurstIrGenerationExtension' generated invalid IR.
Declaration with wrong parent:

emitted three times — matching two generated subclasses plus the no-args constructor for a three-value burstValues.

Fix

Call patchDeclarationParents() on both generated constructors, mirroring what FunctionSpecializer already does. Per-copy name disambiguation would be the wrong layer: the IR is structurally invalid, and repairing parents lets Kotlin's existing local-declaration naming do the right thing.

Test

New box test BurstValuesWithInlineFunctionsInConstructor.kt, the constructor-parameter counterpart to the existing BurstValuesWithInlineFunctions.kt. It fails on unpatched trunk with the IR validation error above and passes with the fix. Full :burst-kotlin-plugin-tests:test suite is green.

It uses loadClassInstance<CoffeeTest>("CoffeeTest") rather than CoffeeTest() for the default specialization, since box() is compiled in the same unit and the frontend binds the call to the default argument that the plugin later strips.

Not verified

I reproduced this on trunk's Kotlin 2.4.10, not on the reporter's exact 2.3.0 + AGP 9 setup, and I did not run the attached BurstLambdas sample against a patched build.