feat(managed-agent): a second concurrent Session on the same Workspace mount should queue, not fail the Turn terminally

What happened?

Reproduced deterministically by the new --second-workspace-session mode of
scripts/run-managed-agent-server-e2e.ts (full packaged stack, macOS arm64).

Two Workspace-bound Sessions share one mount; both Turns are admitted and run
concurrently. The second Session's Turn fails:

qwen serve: Hosted Harness turn … failed: Error: Runtime Broker returned
HTTP 409 (workspace_busy).

{
  "completedTurns": 1,
  "failedTurns": 1,
  "busyEvidence": true,
  "crossTalk": false
}

The Broker correctly serializes the mount — but the failure surfaces as a
terminal turn.failed (hosted_turn_failed), not a queued admission or a
classified, retryable rejection. A second user (or a second Tab/agent) on the
same Workspace simply loses the Turn; the client cannot tell "busy, try
later" from a genuine execution failure, and the coordinator does not
redrive.

What did you expect to happen?

One of:

• the second Turn queues until the mount frees (admission stays 202, the Turn
  starts after the first finishes); or
• admission/settlement is refused with a classified workspace_busy
  code the client can retry — not the generic hosted_turn_failed.

Positives in the current behavior, pinned by the mode: no cross-Session text
leakage, and the winner's tool executes exactly once.

Client information

Packaged-stack E2E runner (--second-workspace-session), repo HEAD
1cf80d46d0, macOS arm64. The mode pins the current fail-terminal behavior
and fails loudly if both Sessions ever complete — i.e. when queueing or
serialization lands.

Anything else we need to know?

Parent: #12380. This is arguably a product decision rather than a defect;
filing so the fail-terminal shape is a conscious choice with a named code.