swarm: agents run headless; no interactive sessions, one runner per invocation

Two related constraints in thrawn swarm, both inherited from the runner abstraction (argv templates with a {prompt} placeholder):

1. Headless agents. Each pane runs e.g. claude -p <prompt> --output-format stream-json on the issue — you watch and steer at the git level, but can't type into the session. An interactive mode (claude REPL started in the worktree, issue text pre-seeded, perhaps via herdr agent prompt) would suit "user as orchestrator" even better. Would likely need an optional per-runner interactive_argv in runners.toml.

2. One runner per swarm invocation. --runner applies to every issue. Mixing runners currently means running swarm twice. Could support COS-2101:codex COS-2102:opus syntax if wanted.

From the swarm mode PR (#4).

🤖 Generated with Claude Code

Start by reading the swarm mode PR (#4) and the runner abstraction behind thrawn swarm, including the runners.toml argv templates. Compare the proposed interactive_argv/herdr agent prompt flow with the current --runner behavior and decide whether per-issue runner selection is also in scope. Done should be defined as an agreed, testable design for interactive sessions and/or mixed runners.