Make loop engineering resumable across agent hosts

[ideogram-fredtran]

What

• Keep active loop invocations running until completion or a real blocker.
• Add SHA-bound successor state for intervention recovery while preserving cumulative budget.
• Document honest recurrence behavior for Codex, Claude Code, and Cursor.
• Distinguish malformed CLI use (exit 2) from valid commands rejected by the state contract (exit 3).
• Clarify terminal-cycle accounting with finish --consume.

Karpathy checklist

• Think Before — host/runtime assumptions and intervention boundaries are explicit.
• Simplicity First — state transitions remain script-first; host skills are thin wrappers.
• Surgical Changes — seven loop-engineering/test files only.
• Goal-Driven — red-path fixtures cover invalid transitions and non-mutation.

Verification evidence

• Skill Creator validation: loop-engineering and evidence-gate valid.
• Focused unit tests: 15 passed.
• Static repository checks: 10 passed.
• Full repository checks: 65 passed.
• Ruff formatting and lint checks passed.
• git diff --check passed.
• Independent Claude and OpenCode cold retests recommend publishing faed420fdc9ed149dd7ab4a43917f3cefecdc7b1 as-is.

Risks / follow-ups

A non-blocking session-path collision observation for post-terminal annotate is recorded separately; it does not change this release. Retest harnesses now allocate collision-proof temporary directories.