sandbox Phase 3 — precompile the WASM module once, instantiate per invocation

[os-zhuang]

Part 3 of #3275 — see its Phase 3 section for the full spec. Blocked by #3296 (Phase 2).

Verified API (quickjs-emscripten@0.32.0): newVariant(baseVariant, { wasmModule }) where wasmModule?: OrLoader<WebAssembly.Module> — "Emscripten will instantiate the WebAssembly.Instance from this existing WebAssembly.Module." The compiled WebAssembly.Module is stateless bytecode; each newQuickJSWASMModule(variant) still allocates a fresh linear memory (isolation preserved) as long as wasmMemory is NOT also passed (that would share memory = the rejected model).

Change (packages/runtime/src/sandbox/quickjs-runner.ts), module-level once per process:

const wasmModule = await WebAssembly.compile(baseWasmBytes);      // compile ONCE
const FAST_VARIANT = newVariant(RELEASE_SYNC, { wasmModule });    // reuse across instantiations

then Phase 2's newQuickJSWASMModule() → newQuickJSWASMModule(FAST_VARIANT).

• Confirm how to obtain baseWasmBytes once for the RELEASE_SYNC variant (variant wasmLocation; CustomizeVariantOptions.wasmBinary). Fallback: the Emscripten instantiateWasm(imports, onSuccess) hook via emscriptenModule — new WebAssembly.Instance(cachedModule, imports) per call.
• If clean bytes access proves awkward in 0.32, use the instantiateWasm fallback; if neither is clean, this phase is deferrable — Phase 2 already removes the asyncify cost.

Acceptance criteria

• Compiled module cached once; per-invocation instantiate only.
• Isolation invariant asserted (RSS soak): fresh instance ⇒ fresh linear memory; RSS returns on dispose across N invocations.
• Bench shows the compile-cost removal vs Phase 2.
• Sandbox suite green unchanged.

Do NOT take the shared-singleton-module route (see #3275 §REJECTED — cross-tenant blast radius + RSS ratchet).

Refs: #3275 (parent) · #3296 (Phase 2).

🤖 Generated with Claude Code

[os-zhuang]

Deferred (PR #3318, ADR-0102 D3). After verifying the quickjs-emscripten@0.32.0 API against the installed sync variant, precompile-once does not clear D3's own bar:

• No clean, portable, edge-safe way to obtain the .wasm bytes. @jitl/quickjs-wasmfile-release-sync loads its wasm lazily via the variant's importModuleLoader (node/browser/workerd conditions). Precompiling needs the raw bytes first — sourcing them ourselves means a Node fs/fetch of emscripten-module.wasm (breaks the Workers/edge-safe design that motivated quickjs-emscripten over isolated-vm) or replicating emscripten's per-runtime locateFile (fragile). The instantiateWasm hook has the same problem.
• Uncertain payoff. Engines already code-cache wasm compilation for identical bytes, so newQuickJSWASMModule() largely amortizes the compile already; D3 wouldn't remove the inherent per-invocation instantiate. Phase 2 (perf(runtime): drop asyncify — sandbox on the sync QuickJS variant (ADR-0102 D2, #3296) #3309) already shipped the dominant win — dropping asyncify.

Revisit only if profiling shows per-invocation compile is a real hotspot and quickjs-emscripten exposes a portable precompiled-module handle. Until then, D1 (#3301) + D2 (#3309) stand on their own. Closed via #3318.

---

Generated by Claude Code