Fix Safari: strip dead worker-entry export that breaks shell boot

[brandonpayton]

Problem

The deployed browser demos (https://automattic.github.io/kandelo/) never finish booting in Safari/WebKit — the shell freezes at instantiating kernel... with no console error. Chromium and Firefox are unaffected. It only reproduces in the production build, so it's invisible on the Vite dev server (which is why it couldn't be reproduced locally at first).

Root cause

Rolldown (Vite 8's bundler) synthesizes a dead export { … } on the kernel/process worker entry chunks. The source has no such export and nothing imports it — but its presence makes WebKit evaluate the type:"module" worker twice:

1. Evaluation feat: centralized multi-process kernel #1 handles init, sets initReady = true, posts ready.
2. Evaluation fix: dispatch_channel_syscall build errors #2 re-runs the module top level and reinstalls self.onmessage, bound to a fresh module state whose initReady is false — shadowing feat: centralized multi-process kernel #1's handler.

kernel.init() then sends register_lazy_files / register_lazy_archives and waits for an ack. Those are now handled by the uninitialized handler, which parks them in its !initReady pending queue and never replies. kernel.init() never resolves → the machine never boots.

Fix

A worker-build Vite plugin (dropWorkerEntryExports) strips the dead trailing export { … } from worker entry chunks at renderChunk, so the worker is a single-evaluation module on every engine.

The "proper" lever — preserveEntrySignatures: false — has no effect here: rolldown-vite (Vite 8 / rolldown 1.0.3) doesn't thread that option into the worker build (byte-identical output). There's an inline, dated note to revisit and drop this plugin once it does.

Verification

Reproduced and verified on real Safari 26.5 (via safaridriver) and Playwright WebKit, against a production build served exactly like GitHub Pages (service worker injects COOP/COEP, no direct headers):

Build	real Safari	WebKit (run echo)	Chromium
Baseline (unfixed)	❌ hang at "instantiating kernel..."	—	✅
Fixed	✅ boots	✅ boots + runs command	✅ boots + runs command

Confirmed the rebuilt worker chunks (kernel worker and nested process worker) no longer contain any export, and the command-run test proves the nested process worker still spawns.

Scope

One file: apps/browser-demos/vite.config.ts. No runtime/kernel/ABI changes.

🤖 Generated with Claude Code

[github-actions]

prepare-merge: runtime/materialization tests passed against the synthetic PR merge; package staging and durable package publishing were skipped. merge-gate=success posted on PR HEAD and squash auto-merge enabled.