docs(docs): state the delegate-first execution model in AGENTS.md and the campaign-lead skill

[LittleCoinCoin]

Why

The lead-plus-implementers execution model ("delegate first, but not stubborn about it") was stated only in the opening prompts of development cycles 1–4 and nowhere in the repository. Once those prompts stopped carrying it (cycle 5 onward) every cycle ran with zero dispatches, and three architecture reports cited the campaign-lead rule "do lead-sized leaves yourself" as the reason to run a whole cycle single-handed. The exception had swallowed the rule because the rule was never written down.

What

• AGENTS.md gains an Execution model section, harness- and vendor-free: the roles, the two currencies it trades in (the lead's context tokens on the most capable tier, re-sent every turn, versus an implementer's once-spent tokens on a cheaper tier; parallel wall-clock versus the 4–15 minute dispatch overhead KT-C/KT-D measured), the delegate-first default, the per-leaf exception, and the sequential fallback when a harness offers no parallel workers. It ends by routing to the campaign-lead dev skill before deciding a task is too small. The skill's trigger row moves that decision inside the skill.
• campaign-lead states The Default: Delegate First ahead of its order of operations, so the lead-sized rule reads as the exception it is (per leaf, never per cycle). Its two Claude Code specifics (the Agent tool's worktree base, the end of a subagent's turn) become harness-neutral invariants pointing at the instance record.
• stack-traps receives those two Claude Code instances as symptom / cause / what-to-do entries (#agent-worktree, #worker-turn), plus symptom-index rows. No skill enumerates harnesses; the harness instantiates the contract.
• The dispatch template says tier is the lead's choice and asks implementers to report gate attempts, the escalation signal for that leaf.
• The trigger eval prompt loses its model name ("three implementers").

Gates

• uv run pytest green (AGENTS.md at 125 lines against the 130-line drift cap)
• uv run ruff check / ruff format --check green
• claude plugin validate . and ./dev pass
• cz check --rev-range main..HEAD passes
• Not run: claude plugin eval for the campaign-lead trigger (needs an authenticated claude -p session); the prompt change is a two-word deletion with an unchanged grader.

🤖 Generated with Claude Code