ops: ephemeral task-scoped dev agents with worktree isolation

[baudbot-agent]

What changed

Major architecture change to how dev agents work.

Before

• One persistent dev-agent session, long-lived, receives all tasks serially
• Dev agent creates its own worktrees and cleans them up
• Single agent for all repos

After

• Ephemeral, task-scoped dev agents — spun up per task, cleaned up when done
• Named by repo + todo: dev-agent-<repo>-<todo-short> (e.g. dev-agent-modem-a8b7b331)
• Start in the worktree: control agent creates the worktree first, then launches pi with CWD = worktree root. This means pi auto-discovers project context from the repo.
• Worktree isolation: multiple agents can work on the same repo simultaneously (separate branches/worktrees)
• Max 4 concurrent dev agents
• Multi-repo tasks: one agent per repo, coordinated sequentially or in parallel by the control agent
• Control agent owns lifecycle: creates worktree → spawns agent → collects results → kills session → removes worktree
• No persistent dev-agent at startup — all on-demand
• Sentry agent unchanged (stays persistent)

Why

• Better project context loading (agent starts in the repo, reads CODEX.md naturally)
• Parallel work on multiple tasks
• Clean isolation between tasks (no stale context carryover)
• Two agents can work on the same repo without conflicts
• Simpler mental model: task in, agent up, work done, agent gone