Worlds: move World-scoped state onto the World instance so bundling can't duplicate it

[pranaygp]

Follow-up to #3728, which fixed the immediate regression but left a discipline where an invariant belongs.

Background

@workflow/world-local and @workflow/world-vercel are bundled into the host application's server build. A bundler keys module identity on (resource, layer), and Next.js alone compiles instrument, app-route, ssr and edge as separate module graphs — so one process holds one copy of every module in these packages per layer. Any mutable module-scope binding is therefore per-copy state rather than the process singleton it reads as. #3493 introduced this exposure for world-vercel by removing it from serverExternalPackages; the visible casualty was the events WebSocket transport falling back to HTTP.

#3728 routes that state through globalSingleton() (globalThis + Symbol.for) and enforces it with scripts/lint/module-scope-state.mjs, run over every published packages/world-*.

The problem with stopping there

Most of the state that was converted is conceptually per-World, not per-process:

• ws-transport.ts — the channel registry
• http-client.ts — the undici keep-alive pools and the events dispatcher recycler
• runs.ts — the long-poll-unsupported negative cache

On the World instance, none of it can be duplicated by a bundler at all — it rides the World's existing singleton for free, and no rule is needed to keep it correct. globalThis would then be left holding only the genuinely process-wide things: ID generators whose sequence must not fork (create-run-id.ts, and world-local's evnt_/chnk_ ULID factories) and log-once latches. That is a surface small enough to reason about instead of police.

Direction

1. Move per-World state into createWorld()'s closure / the returned World object in world-vercel and world-local.
2. Keep globalSingleton() for the process-wide remainder.
3. Leave the lint rule in place as the backstop — it should get quieter, not disappear.

Sequencing note: this pairs with #3665. Once request-time handlers and runtime execution share one World, instance state is automatically shared across every layer that reaches it — the two changes together turn "don't rely on module scope" from a rule into a property of the design.

Interaction to watch: setWorld() and the CLI/test paths construct Worlds directly, so anything moved onto the instance must be safe to have more than one of (pools already are; ID generators are not, which is why they stay global).

[pranaygp]

Partly overtaken by #3728, which went wider than this issue assumed.

The sweep now covers every package that lands in the host application's server build, not just the worlds: @workflow/core (22 findings → 0), @workflow/world, @workflow/ai and @workflow/nest, alongside the four world packages. Build-time packages, the CLI, the o11y UI and the test runner stay out — a single module graph makes the hazard impossible there.

Two things learned while doing it, both relevant to whatever picks this up:

One instance of core is not reachable. Layers cannot share a module, and core cannot be made external because core is workflow code — runtime/start.ts:253 and nine methods in runtime/run.ts are 'use step' functions that must go through the SWC loader, which an external package never does. packages/next/src/index.ts:217-221 already strips workflow-bearing packages back out of serverExternalPackages for exactly this reason. Getting to one copy would mean splitting core into a transformed shim plus an external, directive-free engine. That is a real option, but it is an architecture project and it trades against the cold-start win #3493 was chasing.

step-single-flight.ts was the one with teeth. A per-copy map is not single-flight: two invocations arriving through different layers would each believe they were alone in the process and both run the step body, degrading in-process dedup to the cross-process residual its own doc scopes out to the ownership lease.

What remains under this issue is the original idea and still worth doing: move World-scoped state onto the World instance — the ws registry, the undici pools, the long-poll cache — so it rides the World's singleton and cannot be duplicated at all, leaving globalThis holding only the genuinely process-wide things (ID generators whose sequence must not fork, log-once latches). That pairs with #3665, now also fixed in #3728.