Simplify loop-engineering for Astra, Fable and Sol orchestration

[cheshirecode]

Loop-engineering mixed model behavior with host capabilities and repeated conflicting instructions across its references. The docs now have one portable workflow with explicit Astra/Fable guidance, Sol orchestration, and an evidence-preserving fallback for other models. Model identity grants neither tool access nor write isolation.

Reviewed all 11 existing instruction documents and added one focused model reference. Reduced total documentation from 14,466 to 6,910 words. Removed historical anecdotes, mandatory questionnaires, hidden-reasoning requests and repeated workflow prose. Corrected scheduler status, changed-goal recovery, Codex-only skill discovery, and successor driver configuration so radar and queue probes survive resume. A scoped wave can finish without archiving unrelated project work.

Replaced 20 wording-only state tests, one obsolete sibling wording assertion and a duplicated static prose checklist with executable documentation examples and reference graph checks. All 26 retained state tests are AST-identical; runtime sources and packaged code are unchanged. The examples now verify completion evidence, bound successors with both probes, and verdict exit handling. Added a Codex-only-root resolver fixture. Retained the existing executable whole-project completion recipe and its real-vault tests for success, orphan refusal and rejected pushes.

Validation:

• macOS: 14 static checks and 94 skill tests passed; final correction passed 18 evidence-gate tests and 3 Worklog completion tests; skill validator and package parity passed.
• Sibling sandbox Linux ARM64, running as dev: the full suite passed, 158 checks and zero failures; package parity passed.
• All five CI jobs passed at 95c249f, including the Node MCP end-to-end test: https://github.com/cheshirecode/dotfiles/actions/runs/34707870047.
• Final source archive SHA256: 3edc11ef0ca7fc94ac58ad0344c5011260e900fc7eede31e7bb8c24d4124bf33.
• Independent review found no remaining defects. Actual GPT-5.6 Sol forward review correctly handled stale/duplicate returns, shared ownership, unavailable Fable, scoped completion, missing scheduler and changed goal. Fable guidance is based on official documentation; no Fable model execution or native Windows validation is claimed.

Worklog: generic-loop-engineering / loop-model-contracts, loop-docs-optimization, loop-model-portability. Current model references are linked from the maintained model guide. Word-count reduction is measured, not a claimed token or quality benchmark.