v0.7: agent can spawn sub-agents for parallel sub-tasks

[proofmancer]

An agent should be able to decompose a task and spin up child agents to run pieces in parallel, automatically or on user request. Same pattern Claude Code itself uses with its Agent/Task tools, exposed inside Argus.

Scope

• Agent definition gets a "can spawn" toggle, off by default.
• When enabled, the system prompt gains a documented tool / convention for calling `argus.spawn({ agentId, prompt })`.
• Each spawned run:
  • inherits the workspace cwd (or a per-agent override once v0.2: multi-directory workspaces (one project, many cwds) #2 ships)
  • inherits the shared MEMORY.md (v0.2: shared MEMORY.md per workspace #3) as system context
  • streams into the parent run's output, tagged with the child agent name
  • shows up in run history (v0.2: persistent run history per agent #4) the same as a manually started run
• Stop the parent run, the child runs get killed too (process group).

Out of scope (until we see usage)

• Resource limits / concurrency caps
• Depth > 1 (sub-agents spawning sub-agents)

Why

Lets one prompt fan out across a backlog without the user juggling tabs. This is what makes Argus more than "Claude CLI with a UI".