Adversarial multi-agent pipeline with 4 modes (full, small-fix,
brainstorm, deep-research). Each agent reviews the previous agent's
work against actual code, not descriptions.

- tasks skill: mode suggestion + --mode flag, Mode field in task.md
- maw skill: mode-aware pipeline, /maw NNN to run specific task,
  --worktree / --no-worktree flags, mode-specific wrap-up for
  plan-only modes
- README: modes documentation, comparison table, usage examples,
  artifact maps