v0.10.1 — hyper-plan-loop: planner writes the plan file directly

Scoped optimization of the experimental hyper-plan-loop skill. Non-breaking — hyper-plan-loop usage is unchanged; stock hyper-plan / hyper-plan-review behavior is identical.

What changed

• planner agent gains a caller-directed write-file mode: when a skill explicitly directs it to a path, it Writes the plan itself and replies WROTE: <path> only. The default return-body mode is unchanged — stock hyper-plan still has the skill own the Write.
• hyper-plan-loop now has the planner teammate write the plan file directly instead of round-tripping the full plan body through the lead each revise iteration:
  • Eliminates per-iteration plan-body token cost.
  • Adds an idle/no-resend prompt contract + a lead-side unsolicited-message protocol → kills the RESEND: / duplicate-body churn.
  • Filesystem cp backup/restore (not Read-into-context) preserves the token saving across malformed revisions.
  • Anchored WROTE: reply gate, cmp -s no-op-revise detection (normal and malformed-redo paths re-enter the full validation pipeline), and a try/catch structure check that routes a missing/clobbered file through .bak restore.

Notes

• Requires the experimental agent-teams feature (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) — unchanged from v0.10.0.
• Manual dogfooding (agent-teams enabled) remains the acceptance path for the experimental loop; it is not unit-testable.
• No migration steps.