refactor(runtime): consolidate engine backends and harden ownership

[DjDeveloperr]

Summary

• consolidate runtime teardown and ownership handling across direct-engine and Node-API backends
• keep direct-engine marshalling concrete and hot-path-specific while sharing lifecycle infrastructure
• remove duplicate Android QuickJS/mimalloc, JSC, and V8 sources plus generated binaries, archives, samples, and vendored allocator collateral
• generate one shared signature-dispatch table for every backend
• harden circular conversion, weak references, finalizers, worker shutdown, JNI refs, timers, nested FFI layouts, callback lifetimes, and per-runtime caches
• package the complete shared Objective-C bridge for React Native and add package tests

The committed range is 396 files, 4,104 insertions, and 86,265 deletions: net -82,161 lines.

Architecture

The Node-API backend remains the ownership/lifecycle reference. Direct backends now use a small shared cleanup registry and shared bridge organization, but engine-native value marshalling stays in the concrete V8, JSC, QuickJS, and Hermes implementations to preserve their faster call paths. Hermes native structs use JSI NativeState instead of a hidden JavaScript property.

Memory validation

• macOS CLI builds: V8, QuickJS, JSC, Hermes x direct/Node-API
• semantic ownership matrix: 288/288 across all eight configurations with GSD enabled and disabled
• post-CI ownership/identity/initializer regression matrix green across all eight configurations with GSD enabled and disabled
• QuickJS direct post-fix semantics rerun: 15/15 GSD on and 15/15 GSD off; the former teardown crash path also passed 10 consecutive runs in each mode
• full RSS stress reports green across all eight configurations
• four-wave Objective-C wrapper plateau test green across all eight configurations
• worker lifecycle, worker error, main-env shutdown, JS-function finalization, selector-group finalization, nested FFI, pointer buffer, unmanaged transfer, and circular JS/native tests included
• React Native package tests pass
• FFI boundary audit passes

JSC Node-API previously allocated a new JSClassRef for every weak reference/wrapper. Sharing one class per info type reduced representative RSS from 278.3 MB to 62.4 MB for wrapper churn and from 162.4 MB to 68.9 MB for mixed stress. The final focused plateau rerun was 49.0 MB.

Window-bearing stress tests are excluded by default. They are opt-in with MEMTEST_WINDOW_TESTS=1, and the AppKit cases construct borderless windows offscreen.

Performance

Median microseconds per call, five alternating GSD-on/off runs:

Runtime	GSD	NSDate.date	NSDate.double	NSDate.primitive	NSString.compare	CFAbsoluteTime	CGPoint.get	CGPoint.set	CGPoint.construct
V8 direct	on	9.489	6.241	0.070	0.117	0.073	0.043	0.113	0.377
V8 direct	off	9.446	6.220	0.084	0.212	0.072	0.042	0.107	0.377
V8 N-API	on	69.843	0.229	0.254	1.014	0.034	0.268	0.228	0.442
V8 N-API	off	69.475	0.316	0.279	1.254	0.061	0.270	0.226	0.443
QuickJS direct	on	5.277	5.380	0.102	0.174	0.098	0.044	0.073	0.279
QuickJS direct	off	5.415	5.470	0.110	0.256	0.099	0.044	0.073	0.272
QuickJS N-API	on	68.735	0.174	0.238	0.774	0.055	0.600	1.179	11.088
QuickJS N-API	off	68.446	0.274	0.261	1.018	0.086	0.602	1.224	11.480
JSC direct	on	8.991	8.911	0.178	0.302	0.130	0.087	0.172	0.462
JSC direct	off	9.084	8.881	0.183	0.429	0.125	0.092	0.186	0.465
JSC N-API	on	74.139	0.527	0.707	4.523	0.268	0.480	0.976	1.715
JSC N-API	off	105.815	0.738	0.785	5.100	0.279	0.450	0.907	1.682
Hermes direct	on	4.276	4.020	0.186	0.263	0.158	0.099	0.196	2.504
Hermes direct	off	4.369	4.107	0.187	0.337	0.157	0.098	0.193	2.504
Hermes N-API	on	134.547	0.337	0.265	1.025	0.118	0.136	0.241	0.755
Hermes N-API	off	99.975	0.440	0.287	1.241	0.145	0.135	0.239	0.758

NSDate.double is dateWithTimeIntervalSince1970, NSDate.primitive is timeIntervalSince1970, and NSString.compare is compareOptions.

Direct-engine medians remain comparable to or better than the pre-change baselines. Allocation-heavy NSDate.date is noisy in Node-API configurations because finalizer and GC work is included. JSC Node-API's bounded-memory fix intentionally replaces the old leaky allocation behavior; isolated reruns put that case around 111-114 us/call.

Android and packaging

• native C++ targets built for Debug and RelWithDebInfo across four ABIs for QuickJS, QuickJS-NG, JSC, Hermes, PrimJS, V8 10, V8 11, and V8 13
• JSC native targets were rebuilt across all four ABIs after the final shared-class RAII change
• React Native no-pack staging contains the full host_objects bridge and both generated dispatch tables
• metadata generator builds and its normalized source hash matches the built generator

The aggregate Android Gradle invocation still reaches unrelated Java/Kotlin tasks and fails under the machine's Java 25 installation. Native CMake tasks are green; the repository documents JDK 17 as the supported toolchain.

Artifact audit

• no added jars, archives, native binaries, node_modules, build directories, or generated package staging
• no absolute worktree or temporary paths in committed files
• zero Git LFS objects to push; all changed LFS entries are deletions

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 2963b8a8-e1aa-4927-9721-36d8b545c3d0

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}