Harden the generated FFM runtime and copy value types verbatim

[kdroidFilter]

Summary

Enums and data classes are now generated on the JVM as verbatim copies of their native declarations. While verifying this end to end, torture tests and seeded monkey tests uncovered a series of runtime bugs in the generated FFM layer, all fixed here.

Value types

• JVM enum class / data class keep default values, body members, entry bodies, companion objects, KDoc and annotations (e.g. a custom equals/hashCode on a data class with a ByteArray field was previously dropped).
• Imports of native-only packages (kotlinx.cinterop, platform.*, kotlin.native) and of module packages are removed.

Runtime bugs fixed

Bug	Symptom before	Fix
Exception thrown in a JVM callback	JVM terminated (uncaught exception in upcall)	Upcalls catch, native caller unwinds, original exception rethrown
Global native error state	Errors lost or raised on the wrong thread	Thread-local error
Collections > 4096 elements / 8 KB strings	IndexOutOfBoundsException or crash (returns, sets, maps, suspend, DC fields)	Oversized result parked natively, read back with a grown buffer
String / ByteArray > 8 KB	Native function called twice (side effects duplicated)	Read the parked result instead
Data class String field > 8 KB	Silently truncated	Marker + StableRef read back
Flow with slow collector	19 935 / 20 000 elements dropped	Credit-based backpressure
Suspend cancelled before native start	close() spins forever	CoroutineStart.ATOMIC, completion always signalled
Cancel racing completion	Use-after-free of the job StableRef	Job handle owned and serialized by the JVM
Call after close(), concurrent close(), closed object as argument	Use-after-free	Per-object lease: IllegalStateException, deferred disposal
Cleaner action captured the proxy	Unclosed objects never reclaimed	Action only references the lease
Per-call upcall stubs in long-lived arenas	Unbounded memory growth → OutOfMemoryError	Static coroutine stubs + per-call callback arenas
Abstract/sealed hierarchies	Native objects never disposed	Generic StableRef disposal
List<Short> / List<Byte>, non-String map keys	Corrupted values / ClassCastException	Consistent element encoding
NUL in String arguments	Silently truncated	IllegalArgumentException (returns decoded by length)

Behavior changes

• Callbacks are valid for the duration of the call only (like JNI). Native code must not store a JVM callback; event streams should be exposed as Flow<T> (as systeminfo already does). README updated.
• Calls on a closed object throw IllegalStateException.
• String arguments containing \u0000 throw IllegalArgumentException.

Tests

• ValueTypeParityTest, ValueTypeTortureTest: value types across FFM, overflow parity, class-init races in fresh classloaders, re-entrant callbacks.
• RuntimeTortureTest, BufferTortureTest: every buffer past its limit, error isolation across threads, lifecycle races, cleaner reclamation, flows and cancellations.
• MonkeyTest: seeded model-based monkey (each native result checked against a JVM model) and concurrent chaos monkey (8 threads, random closes, throwing callbacks, cancellations, GC). Replay with MONKEY_SEED=<seed>, widen with MONKEY_SEEDS=<n> (add --rerun).
• systeminfo: screen capture test accepts the PNG produced on macOS.

Results on macOS arm64: calculator 1785 tests green, monkeys green on 500 seeds each; benchmark and systeminfo green; plugin unit tests green.