chore: bump deno_core 0.368.0 → 0.412.0 (v8 142.1.0 → 150.4.0)

[imfing]

Summary

Two related changes:

1. Bump deno_core across 44 releases (0.368.0 → 0.412.0), moving V8 from 142.1.0 to 150.4.0.
2. Add the Build V8 archive workflow (workflow_dispatch-only) that prebuilds rusty_v8 from source and publishes it as release assets — so after this merges, the archive for v150.4.0 can be built once and reused by all future wheel builds.

deno_core API migrations

• op2 macro: the bare #[op2(async)] flag was removed upstream (async(...) now requires an explicit lazy/deferred/fake mode). Our async ops return Result<impl Future>, which the macro now auto-detects as eager async — declared with plain #[op2].
• PollEventLoopOptions: pump_v8_message_loop field removed (pumping is unconditional now).
• InspectorSessionProxy: tx/rx moved into the new InspectorSessionChannels::Regular variant.

The dependency graph now goes through the new deno_v8 wrapper crate, but the underlying v8 (rusty_v8) crate and its build script are unchanged.

Behavioral fix (caught by tests)

deno_core now captures a tokio runtime handle at isolate creation to honor delayed V8 foreground tasks (e.g. GC memory-reducer tasks) and aborts the process if such a task is posted from an isolate created outside a tokio context. tests/test_runtime.py::test_sync_eval_triggers_heap_termination reproduced the abort reliably under heap pressure.

Fixes:

• spawn_runtime_thread now creates RuntimeCoreState (and thus the JsRuntime) inside a scoped tokio_rt.enter() guard.
• SnapshotBuilder gets its own current-thread tokio runtime, entered during JsRuntimeForSnapshot creation and kept alive alongside the isolate.

V8 archive workflow

.github/workflows/build-v8-archive.yml + .github/scripts/build-v8-archive.sh:

• Manual dispatch only; runs on free GitHub runners (both jobs on ubuntu-latest: x86_64 builds natively, aarch64 cross-compiles via the existing Dockerfile.aarch64).
• Reuses the existing manylinux_2_28 builder containers and the v8 git-tag patch (version read dynamically from Cargo.lock, so it stays in sync with bumps like this one).
• Publishes librusty_v8_release_{target}.a.gz + src_binding_release_{target}.rs to a prerelease tagged librusty_v8-v{version} (deliberately not matching the v* pattern that triggers release CI).
• Frees ~25–30GB of runner disk first (V8 needs ~25GB; free runners start with ~14GB).

Why from source at all: denoland's prebuilt archives require glibc ≥ 2.30 (pthread_cond_clockwait), while manylinux_2_28 wheels must link against glibc 2.28.

A follow-up PR will switch wheel CI to consume these archives via RUSTY_V8_ARCHIVE/RUSTY_V8_SRC_BINDING_PATH, removing the per-release V8 compile and the GCP spot-VM provisioning.

Testing

Full Python suite passes locally: 291 tests. cargo fmt and clippy clean (aside from the pre-existing result_large_err warning). The archive workflow itself needs a trial dispatch after merge — expected ~2.5–4h per arch on free runners; if V8 150.x needs a newer clang than the containers' clang-19, that run will surface it.

Suggested sequencing

1. Merge this PR.
2. Dispatch Build V8 archive from main → publishes librusty_v8-v150.4.0.
3. Follow-up PR reworks wheel CI to consume the archives on free runners (retiring the GCP jobs).