feat(jobs): graceful restart handling for push/resume (milestone 4 / A)

[shoom1]

Summary

Milestone 4 (Option A — graceful restart handling) of phase-2 push/resume. Builds on the coordinator (#87). Makes the "job finished but its conversation is gone" case correct instead of broken — most relevant after a CLI restart, since ADK's default InMemorySessionService means the conversation (and the pending call the resume must match) is gone on the next run.

The problem

The job and its resume association survive a restart (subprocess sentinel + persisted JobRecord), but the ADK conversation does not. Before this, draining such a job called process_resume, which posted "↻ resuming…" and then streamed nothing (the session/call were gone) — an orphaned, confusing turn.

What it adds

• BaseWorkflowManager.can_resume(record) — default False (no resume support). GoogleADKWorkflowManager overrides it to require the resume ids and a live session that still holds the pending call.
• MessageProcessor.process_resume now gates on hasattr(resume_with_job_result) and await workflow.can_resume(record):
  • resumable → resume turn, exactly as before;
  • not resumable → posts ✗/✓ Background job 'x' finished (…) while its conversation was unavailable — fetch the result with /jobs <id> and returns. The coordinator already marks it resumed, so it notifies once; the result stays reachable via /jobs.

Restart UX (no new startup turn)

Deliberately no spontaneous startup resume — running a full turn's output over the live prompt is the same display-unsafe situation that kept the monitor to the status bar. Instead: awaiting_resume is persisted, the monitor's ↻N to resume cue shows it after a restart, and the next turn boundary / explicit /resume drains it — resuming when the conversation survived, posting the notice when it didn't.

Why not durable cross-restart resume (Option B)?

That requires persistent ADK sessions (--session-id + validating/fixing _inject_session_messages for pending long-running calls) — a session-persistence project, highest effort/lowest certainty. Durable resume is more natural on LangGraph's checkpointer (Option C / milestone 5).

Testing

• tests/workflow/test_adk_job_resume.py — can_resume true when the session is present, false when missing or when ids are absent.
• tests/cli/test_process_resume.py — resume-vs-notice branching, incl. a backend with no resume support (LangGraph today) → notice.
• Offline suite: 1588 passed, 26 xfailed. The live full-loop test still passes through the new can_resume gate.