feat: add /hyper-plan skill — recursive codebase improvement with convergence scoring

[ShaheerKhawaja]

Summary

Adds /hyper-plan, a new skill that chains /plan-ceo-review → /plan-eng-review → execute fixes → /qa into an iterative loop with LLM-as-Judge convergence control. Treats codebase quality like gradient descent — each iteration targets the 2 weakest dimensions until the overall grade hits the target (default 8.0/10).

What PR #144 got wrong

Submitted a raw SKILL.md with no .tmpl template and no registration in the build pipeline. Feedback: "Single raw SKILL.md, no .tmpl template. Closing."

What's fixed

• Proper .tmpl template with {{PREAMBLE}}, {{BASE_BRANCH_DETECT}}, and {{REVIEW_DASHBOARD}} placeholders
• Registered in gen-skill-docs.ts findTemplates() candidates
• Generated SKILL.md committed alongside template — --dry-run passes
• Bisected commits (template → registration → generated output)

Features

• 10-dimension quality rubric with specific evidence requirements per score level (file:line citations required)
• Oscillation detection — if a dimension changes direction 3+ times (up→down→up), it's locked to prevent thrashing
• Degradation tolerance — ≤0.5 score decrease is normal re-scoring variance, only HALT on >0.5 drops
• Focus narrowing — iterations 2+ review only the 2 lowest-scoring dimensions, with graduation at 8.0+
• Validation gate before every commit (lint + types + tests must pass)
• Artifact trail — every iteration's findings and scores saved to .hyper-plan/

How it works

Baseline → Review (all 10 dims) → Fix P0+P1 → Verify via /qa → Re-score → Convergence check
                                                                              ↓
                                                              SUCCESS (≥ target) → stop
                                                              DEGRADED (>0.5 drop) → halt
                                                              OSCILLATING (3+ flips) → lock dim
                                                              CONVERGED (<0.2 delta × 2) → stop
                                                              MAX_REACHED → stop
                                                              CONTINUE → focus 2 weakest → loop

Test plan

• bun run gen:skill-docs — produces hyper-plan/SKILL.md
• bun run gen:skill-docs --dry-run — all FRESH
• bun test — Tier 1 passes
• bun run skill:check — freshness check passes
• Only hyper-plan/ and scripts/gen-skill-docs.ts changed vs upstream/main