GeezerVison

Makes ObjC calls on objects the GC already released fail in a way an app can trace, and in debug builds names what released them. Builds on the "Known hazard" section of docs/knowledge/v8-resurrecting-finalizers.md.

The problem

When native code retains an instance of a JS-derived class, the runtime keeps it alive by refusing its finalizer (GcProtect), not by holding a strong handle. So a GC still finalizes everything reachable only through that instance's JS properties. An ObjC object owned only by JS and stored in such a property gets released, and the instance survives holding a husk.

Real case: a Firebase App Check provider written in JS. Firebase retains the provider, and the provider held its GACAppAttestProvider in a property. After a GC, every token request hung.

A call on that husk reached ArgConverter::Invoke's wrapper == nullptr branch:

• Debug builds threw Cannot call method '…' on a disposed native object (class: <unknown>, …). This can happen during HMR or fast view churn., which points away from the cause.
• Release builds hit tns::Assert(false), so the process aborted with no JS context.

What changes

Calls on a released ObjC object always throw a ReferenceError. This covers methods and properties, in every build, whatever releasedObjectPolicy says:

• It's an ordinary JS exception, routed through NativeScriptException and ReThrowToV8. Callers can catch it, and uncaught it reaches the app's error handlers and crash reporting with a JS stack.
• Calls deliberately ignore "report", unlike the struct, Pointer and Reference touches, which keep using the policy. Skipping a call breaks its caller's contract: a completion handler that never runs is exactly the hang above. With no listener and no console in release builds, a skipped call would also leave no trace at all.

The message names what was released (tns::ReleasedObjectMessage, now shared with GetValueOrReport). Example:

The native NSMutableArray object has been released (method 'count' (count)). It was released by the garbage collector while still referenced from 'OwnerOfChild.child': 'OwnerOfChild' is kept alive only because native code retains it, which does not keep the objects its JS properties reference alive. Keep the owner or the property's value reachable from a JS root

The class and property in that message come from debug-only tracking in the finalizer drain:

• When FinalizerCallback keeps a protected ObjC object, it walks that object's own enumerable data properties one level deep. Accessors are skipped, so no JS runs inside the drain.
• If a property is already a husk (its finalizer ran first in the drain), it is labelled immediately with its class (taken from its constructor) and Owner.property.
• If a property is still a releasable ObjC object, it is recorded in Caches::KeptObjectProperties, so its own finalizer labels it if it runs later in the same drain.
• Each husk is reported once, together with a logged warning.
• A GC prologue callback (installed in debug only) clears the record, and DisposeAllRegistered releases it while the isolate is still alive.

Cost

• Release builds: FinalizerCallback runs exactly the previous DisposeValue call after one RuntimeConfig.IsDebug check, and no GC prologue callback is installed. The message is built and the error thrown only on the husk path, never on a normal call.
• Debug builds: finalized objects aren't tagged or recorded. A release costs one empty-vector check, unless a kept owner recorded that object in the same drain. The real work is the one-level property walk of natively retained objects the GC keeps, which scales with how many such objects are no longer reachable from JS. If that ever shows up in profiles, it can go behind an ns:runtime setting.

The default stays as it is. Making GcProtect a strong root would turn every JS↔ObjC cycle into a leak, and from the JS heap a cycle looks exactly like this case (see the doc). This PR makes the trade-off traceable instead of hanging or aborting. #509 proposes an explicit per-instance opt-in for "strong while natively retained".

Tests

• New spec in GCFinalizerTests.js: names the property an ObjC object was released through when its natively retained owner was kept. An __extends owner is retained only from an NSMutableArray and holds an NSMutableArray child only in a property; then two forced GCs. Under the default "report" policy the spec checks:
  • a call on the husk throws a ReferenceError naming the class and '…​.child';
  • a second call (addObject) throws too;
  • no releasednativeaccess event fires.
• Full suite on a simulator: 1772 passed, 0 failed. The debug warning fired for the OwnerOfChild.child husks.
• ASan (-a): 1767 passed, 0 failed, no sanitizer reports.
• Docs: docs/knowledge/v8-resurrecting-finalizers.md (calls always throw, the debug naming, and the existing regression spec's corrected name) and docs/ns-builtin-modules.md (the releasedObjectPolicy exception for calls, and the debug messages).