feat: add ce:plan-beta and deepen-plan-beta skills

[tmchow]

Summary

This PR rethinks how the planning stack produces plans. The core insight: planning should spend effort on decisions the execution agent cannot safely infer on its own, and stop spending effort on speculative implementation details the agent will derive better from the live codebase.

In practice that means:

• Plans capture requirements trace, boundaries, sequencing, file targets, patterns, risks, and verification
• Plans stop pre-writing code, exact shell commands, or micro-step choreography
• Execution is left to ce:work, where the model has direct access to the repo and can adapt

This is not lighter planning. It is higher-leverage planning: less stale pseudo-implementation, more durable guidance for capable downstream agents.

Shipped as beta skills

The new planning workflow ships as separate beta skills (ce:plan-beta and deepen-plan-beta) alongside the unchanged stable versions. This lets users trial the new approach without disrupting existing workflows.

skills/
├── ce-plan/SKILL.md           # Stable (unchanged)
├── ce-plan-beta/SKILL.md      # New decision-first planning
├── deepen-plan/SKILL.md       # Stable (unchanged)
└── deepen-plan-beta/SKILL.md  # New selective stress test

Usage: Invoke /ce:plan-beta or /deepen-plan-beta directly. Beta plan files use a -beta-plan.md suffix so both stable and beta outputs coexist in docs/plans/ for side-by-side comparison.

Tradeoffs: Beta skills are standalone — ce:brainstorm won't auto-handoff to them, and lfg/slfg orchestration continues using stable skills. This is intentional: simplicity over seamless integration during the trial period. Users compare outputs by running both versions on the same input.

Promotion path: When validated, replace stable skill content with beta, remove -beta naming, delete beta directories. See docs/solutions/skill-design/beta-skills-framework.md and the AGENTS.md promotion checklist.

What changed

ce:plan-beta

• Rewrote the workflow around planning phases, requirements trace, implementation units, verification, and explicit defer-to-implementation boundaries
• Kept the strong parts: origin-doc carry-forward, repo/learnings research, conditional external research, Proof/tracker handoff
• Removed issue-template boilerplate, embedded code examples, exact shell recipes, and over-prescriptive implementation choreography

deepen-plan-beta

• Rewrote from "run lots of agents and add implementation detail" to "identify weak sections, run targeted research, and strengthen only those sections"
• Made deepening conditional and confidence-driven
• Aligned boundaries with ce:plan-beta so it no longer regresses into code blocks or generic padding

ce:work (updated, not beta)

• Updated execution guidance to consume the new plan format directly
• Added awareness of deferred questions, scope boundaries, per-unit patterns, and verification
• Tightened final validation around requirements trace and plan coverage

Workflow cleanup

• lfg/slfg now run deepen-plan only when the plan is risky, deep, or under-grounded
• Question handling names platform-specific blocking question tools
• Handoff wording cleaned up to reduce Claude-specific assumptions

Documentation

• Added beta skills framework pattern to docs/solutions/skill-design/
• Added Beta Skills section and promotion checklist to AGENTS.md
• Added Beta Skills section to README.md

Test plan

• Run /ce:plan — verify stable behavior unchanged
• Run /ce:plan-beta — verify new decision-first planning workflow
• Run /ce:plan-beta then /deepen-plan-beta on the output
• Run /ce:work on a plan from /ce:plan-beta — verify compatibility
• Compare /ce:plan and /ce:plan-beta output on the same input
• Verify lfg/slfg still use stable skills and work normally