fix(cloudflare/workers): keep the previous dev worker instance serving until its replacement starts

[sam-goodwin]

Fixes #859.

Under alchemy dev, a reconcile that replaces a Worker instance ("Changes detected, interrupting existing instance") tore down the running workerd — and deleted its dev-registry entry — before the replacement completed runtime.start. For DO/container hosts, whose start includes docker image prep, that dark window spans seconds to minutes and re-opens on every change. Cross-script DO bindings resolve through the registry, so consumers 503 with exactly the reported error for the whole window:

Durable Object "EvalHost" defined in worker "eval-host-dev" not found.

I reproduced the issue's exact forensic state live (host worker present only as its stable-port proxy, no --debug-port instance, no registry entry, "Changes detected" logged with no subsequent "Started") — it is the window between an instance interrupt and the replacement's serve completing, held open by churn or a slow container start.

The serve path is now make-before-break:

  const serveWith = (worker, bundle, proxy) =>
    Semaphore.withPermits(serveLock(worker.id), 1)(
-     // close previous workerd, THEN start the replacement (dark window)
-     Scope.close(previous) → runtime.start(...)
+     // start the replacement while the previous instance keeps serving
+     // (and stays registered), then cut the proxy over and close it
+     runtime.start(...) → proxy.set(url) → Scope.close(previous)
    )

• workerd scopes fork from the provider root scope instead of the instance scope, so tearing down an instance's bundle watcher no longer kills its running workerd — the last good instance keeps serving until the replacement's first serve completes (reconcile replacement, sibling-triggered restartWorker, and rebuilds all included).
• delete / dev: external teardown closes the workerd scope explicitly.
• Both instances share one registry key; safe old-after-new closing needs owner-aware entry removal in cloudflare-runtime (fix(runtime): owner-aware registry entry removal cloudflare-tools#72 — submodule bumped here), which is what forced fix(cloudflare): preserve dev worker registration on reload #812's close-old-first ordering.

The regression test drives the real dev topology (persistent provider sidecar via RpcSpawner + RpcProviderProxy, like alchemy dev): deploy a DO-hosting Effect Worker, redeploy with a changed env var, and assert the old instance still answers mid-handoff and the registry entry never vanishes (stable inode). On main it fails with the issue's failure mode (502 ProxyError: Network connection lost); with this change it passes.

[alchemy-version-bot]

Install the packages built from this commit:

alchemy

bun add alchemy@https://pkg.ing/alchemy/4e3eb4e

@alchemy.run/better-auth

bun add @alchemy.run/better-auth@https://pkg.ing/@alchemy.run/better-auth/4e3eb4e

@alchemy.run/pr-package

bun add @alchemy.run/pr-package@https://pkg.ing/@alchemy.run/pr-package/4e3eb4e

[john-royal]

Makes sense, LGTM