Fix the Agent Skill recursive v2 contract

[fazpu]

Why

loopy-loop 0.7.1 implements the recursive session protocol, but the installable Agent Skill still described several retired v1 behaviors. In particular, it taught flat one-level child requests, a mutable Markdown user inbox, schema-v1 eval and terminal control, session-local harness output paths, and a top-level-only stop caveat.

That mismatch is operationally dangerous because coordinators and their spawned agents are explicitly directed to read this skill. They can receive a correct frozen v2 assignment and then be told contradictory behavior by the installed skill.

What changes

• Rewrites the Agent Skill around the frozen assignment and workflow contract.
• Separates coordinator-owned engine state, workflow-owned semantic state and evidence, and gitignored raw attempt traces.
• Documents direct spawned agents and nested harness coordinators as dynamic delegates inside the current loop layer.
• Documents append-only user inputs and acknowledgements.
• Replaces the flat v1 child request with the recursive schema-v2 pending request, immutable dispatch inputs, factual child outcomes, and separate parent acceptance.
• Replaces the schema-v1 eval and stop examples with canonical eval receipts, schema-v2 goal-check projection, and identity-bound terminal control.
• Corrects stop semantics to tree-wide projection at safe coordinator boundaries.
• Aligns model-tier, delivery-receipt, crash-recovery, monitoring, and custom-worker guidance with the binding design.
• Adds a regression test that requires current v2 concepts and rejects the exact obsolete claims.

Scope and compatibility

This is a documentation-contract patch release. Runtime behavior and dependency floors are unchanged from 0.7.1. Existing v1 sessions remain supported and are mentioned only as explicit legacy compatibility.

Verification

• 386 tests passed.
• Ruff formatting and lint passed.
• Pyright passed.
• uv lock check passed.
• Git whitespace check passed.
• Internal contract audit passed against README, D2-D12, the recursive binding design, and session layout.
• Clean-checkout source and wheel builds passed strict Twine metadata checks; the source archive contains no local ignored docs dependencies.
• Claude Code found an initially inert historical-stop regression marker; commit d15a92d corrected it, and Claude re-reviewed the exact fix as PASS.
• Antigravity independently reviewed the final committed diff, ran all 386 tests, and returned PASS with no blockers.

Release and rollout

After required CI and both external reviews pass, merge and publish v0.7.2. Because the Agent Skill is installed from the GitHub repository rather than the wheel, refresh the global Codex, Claude Code, and Antigravity skill installation after the tag, then resume the existing durable UGM session.