TUI hangs on a blank screen when the server worker fails to bootstrap; two unit tests flake in CI

[mohsensadr]

Description

Two related problems, both in the TUI worker/RPC path and its test coverage.

1. Worker bootstrap failure hangs the TUI

#5 fixed early RPC requests being dropped by having the client hold requests
until the worker posts rpc.ready. That closes the drop, but not the hang: if
the worker dies before it reaches Rpc.listen(), rpc.ready never arrives and
every call sits in the client outbox forever. The user sees the same blank
screen #5 set out to fix, just from a different cause.

worker.onerror only logs, and neither the transport setup nor
validateSession has a timeout, so there is no path that ever rejects.

Two smaller issues in the same file:

• listen() looks methods up with a bare rpc[method], which resolves through
  Object.prototype. A request for toString/constructor invokes it and
  replies with a bogus rpc.result instead of an unknown-method error.
• The queue() comment claims it must be "the first statement" in the worker.
  That is not achievable: ESM evaluates imports before the module body, and
  worker.ts transitively imports core/global, which has a top-level await, so
  execution yields to the event loop before queue() runs. queue() narrows
  the window; the rpc.ready handshake is what actually guarantees delivery.
  Worth correcting so nobody later relies on queue() for a new pre-ready
  message type.

2. Unit tests failing in CI

unit (linux) and unit (windows) have both been intermittently red on main,
independent of the above.

• linux: mcp.McpOAuthCallback > ensureRunning starts a server that isRunning() reflects. The callback server is a module-level singleton and bun shares
  module state across test files in one process, so a server left listening by
  an earlier file makes the test's opening
  expect(isRunning()).toBe(false) fail in ~1ms. The leak is
  oauth-auto-connect.test.ts: authenticate() starts the server and the file
  never stops it.

• windows: rotating timeout flakes rather than real bugs. Three red runs hit
  three different tests, each a hardcoded per-test budget that a loaded runner
  exceeds (e.g. cancel with queued callers resolves all cleanly timing out at
  3s, in a file where sibling tests already take 2-8.5s on windows).

Plugins

No response

PhysicsCode version

No response

Steps to reproduce

Worker hang:

1. Make the TUI worker fail during bootstrap (e.g. force Log.init to throw,
   or make an XDG directory unwritable so core/global's mkdir fails).
2. Run physicscode tui.
3. The TUI sits on a blank screen indefinitely. No error, no timeout.

oauth singleton leak:

1. bun test test/mcp/oauth-auto-connect.test.ts
2. Observe McpOAuthCallback.isRunning() is still true afterwards.
3. Run any file ordered after it that asserts the server is stopped, e.g.
   test/mcp/oauth-callback.test.ts, and its first assertion fails
   (Expected: false / Received: true).

Screenshot and/or share link

No response

Operating System

Ubuntu 22.04 + Windows 11 (CI runners)

Terminal

No response