Overview

Fixes #661.

The shared swiftHeapObjectFinalizationRegistry (added in #648) crashes intermittently in CI when its callback fires during Node.js process shutdown. V8 GC collects leftover SwiftHeapObject wrappers, the finalizer calls state.deinit(state.pointer) (a Wasm export), but by that point the Wasm linear memory or function table is already torn down. Result: RuntimeError: memory access out of bounds or table index is out of bounds, roughly 15% of CI runs.

The hasReleased flag prevents double-release but doesn't help here - the object hasn't been released, the Wasm instance is just gone.

Fix

Wrap state.deinit(state.pointer) in try/catch in both the FinalizationRegistry callback and the release() method. If Wasm is already torn down, we swallow the error. The process is exiting anyway.

An instanceAlive flag was considered but requires hooking into teardown lifecycle for no real benefit over a simple try/catch. The finalizer is best-effort cleanup by design.

Test

testReleaseHandlesThrowingDeinit tests the release() path: JS replaces the wrapper's deinit to throw WebAssembly.RuntimeError after calling the real one, then calls release(). Without the try/catch guard, the error propagates and crashes the process.

The FinalizationRegistry callback path isn't tested directly because the import-side wrapper uses passUnretained (no extra retain), so triggering GC on it would over-release and corrupt the Swift heap - a test artifact, not a production scenario. The template change for both paths is identical (try { state.deinit(state.pointer); } catch {}), so the release() test covers the logic.