CL-6942: split spawn_agent / wait_agents out of task()

[TheGreatAxios]

Summary

• Adds spawn_agent (starts a worker, returns immediately with { agent_id, status: "running" }, never awaits completion) and wait_agents (blocks until any given/live target reaches a terminal state, or a clamped timeout elapses without touching the workers) as two new fleet verbs.
• wait_agents blocking is driven by the existing SubAgentSessionStore mailbox (subscribe) raced against a real timer — no polling.
• task() is unchanged; it remains the fused spawn+block primitive for the single-worker case.
• Registered spawn_agent/wait_agents in FLEET_VERBS (already present from CL-6941) and gated at the same tier-mount point in run.ts as task/search_agents, additively.

Report retention fix: the session store's finished-session cap (maxCompleted, default 20) is a TUI display cap that evicts the oldest finished session — report included — once more than that many have finished. task() never hit this because it awaits its own result before returning. spawn_agent/wait_agents can fan out well past that cap in one turn, so a report could be silently lost before wait_agents ever collected it. Fixed with fleetRecords, a small never-capped map (agent id -> terminal status/report/error) written before the session store's complete()/fail() (so the store's synchronous subscribe notification always sees the up-to-date record); an entry is removed only once wait_agents actually delivers it to a caller.

Worktree isolation: spawn_agent's whole point is running several workers at once, so two implement-intent workers sharing one cwd with no isolation could corrupt each other's edits. I chose the "refuse" option over wiring real worktree isolation into spawn_agent: duplicating task()'s worktree machinery (isolation, cleanup, stash baseline, preserve-on-dirty) inside spawn_agent's independent dispatch path was a larger, riskier surface than this PR's scope justified. spawn_agent now refuses a second concurrent implement-intent (director build) spawn against the same cwd with an actionable error naming the conflicting agent id, and releases the lane once that worker finishes. Explore/plan/review-intent spawns don't write and are unaffected — they may still run fully concurrently. A caller that wants concurrent implementation work today should use task(useWorktree: true), which already has real isolation.

Documented follow-ups (not in scope here): spawn_agent only supports closed-director dispatch (agent= a director id, or intent=) — no custom AgentProfile lookup, no worktree isolation, no nested orchestration, no re-dispatch ledger. These remain task()-only for now.

• CHANGELOG entry under ## [Unreleased].

Test plan

• bun run check green (lint, typecheck, build, full test suite: 5417 pass)
• Tests in src/subagent/agent-fleet.test.ts:
  • spawn_agent returns immediately with a running agent_id without waiting for the worker
  • wait_agents on one target returns once it completes while sibling workers stay running
  • wait_agents with a short timeout on a still-running agent returns timed_out: true without cancelling it, and a second wait_agents call still works
  • wait_agents with no targets waits on all currently running spawned agents
  • reports survive well past the session store's 20-entry display cap until wait_agents collects them
  • refuses a second concurrent implement-intent spawn against the same cwd
  • does not refuse a second concurrent explore-intent spawn against the same cwd
  • releases the write lane once the implement worker finishes, allowing another
• Existing task() / subagent tests unaffected (bun test src/subagent/: pass)
• Rebased cleanly onto latest origin/main (no conflicts in run.ts)

[linear-code]

CL-6942