Add dogfood proof guards and adaptive pilot evals

[uncfreak1255-code]

Summary

• add red-capable debugging and bounded temporary-role development evals
• fail dogfood recording closed when public or authenticated private proof references cannot be verified
• add recorder coverage to npm test
• document selected local invocation paths and the source-pinned Matt Pocock donor audit
• record all ten real dogfood tasks and the final promotion decision
• keep skills/loopspine/SKILL.md byte-identical to v0.2.0

Proof

• npm test: passed
• dogfood: 10/10 verified, 0.0% Sawyer intervention, 4.22 min median time to proof, 30.0% incorrect-stop rate
• pilot diagnostic: +18.75 weighted points, 0 boundary violations, one sample only
• executable trajectories: 6/6 passed
• Claude plugin validation, installed-plugin probe, and trigger/non-trigger smoke: passed
• simplify checkpoint and Codex autoreview: clean

Promotion decision

Global/default promotion is rejected for this round. The pilot recorded three incorrect stops, including unresolved debugging and completion-consistency misses. The three-sample frozen-baseline sealed comparison also remained red at current 0.9412 versus v0.2.0 0.9673 (delta -0.0261), with zero safety violations. Both variants used the same skill SHA, so no behavior-changing candidate is being promoted.

This PR preserves useful eval and proof infrastructure without changing LoopSpine behavior. No hooks, global install, or permanent agents are included.