fix(core): background subagents resume after a server restart instead of showing cancelled - #58261

Background subagents now pick up where they left off after a server restart, instead of showing "Subagent cancelled" and losing their work.
A background subagent is a child agent a session starts and leaves running; when it finishes, its result lands in the parent session. Stopping a subagent on purpose still cancels it, as before.

🔍 Origin: seen in real use: after a clean server restart, every background subagent that was still working showed "Subagent cancelled" and none resumed.

❌ Before

sequenceDiagram
    participant P as Parent session
    participant S1 as Server (shutting down)
    participant DB as Database
    participant S2 as Server (next start)
    Note over S1: Clean shutdown, child still working
    S1->>P: ❌ "Subagent cancelled"
    S1->>DB: ❌ clear the child's "running" marker
    S2->>DB: find "running" markers
    DB-->>S2: none, child is never resumed

✅ After

sequenceDiagram
    participant P as Parent session
    participant S1 as Server (shutting down)
    participant DB as Database
    participant S2 as Server (next start)
    participant C as Background child
    rect rgb(220,245,220)
    Note over S1: Clean shutdown counts as a pause
    Note over DB: "running" marker kept, no result sent
    S2->>DB: find "running" markers
    S2->>C: resume the child once
    C-->>P: real result, delivered once
    end

Also, background subagent handling is now set up once for the whole server instead of once per project, so reloading a project no longer drops the watchers that deliver results.

  on shutdown, for a background job:
-   mark it "cancelled" in the database and tell the parent "Subagent cancelled"
+   tag it "cancelled for shutdown", keep the "running" marker, tell the parent nothing
  on next start:
    resume every job still marked "running", deliver its real result

📊 Proof

With the source changes reverted and the tests kept, the two-process test and the permission-prompt child case fail. 77 core tests pass.

👀 Read: packages/core/src/job.ts, the shutdown cancel path. While the old server is shutting down it sees the job as "cancelled" but the database still says "running", so any new code reacting to a cancellation must check its reason (subagent-completion.ts does). Not checked live in the desktop app. The rest is routine: shared result reading in subagent-completion.ts, a one-line call change in execution.ts, registering the service app-wide in packages/server/src/routes.ts, and tests.

_{Stacked on #58260: merge that first, then this retargets to dev · no schema or saved-format changes}