chore(ai-research-workflows): v0.2.0 prep — command triggers, retreat paths, private skill

[lsetiawan]

Summary

Prep work for the v0.2.0 redesign of ai-research-workflows. Pure documentation and frontmatter changes — zero behavior change. Blocks #87 (orchestrator refactor depends on A2 extraction having already moved trigger content into commands).

Part of the v0.2.0 release tracked by #87, #88, #89, #90, #91, #92. See spec (local): .agents/spec-ai-research-workflows-v0.2.0.md (PR1 section).

Scope (recommendations items A2, A3, A5)

A2 — CSO ## When to use per command

Extract each "When to Suggest …" subsection from agents/research-workflow-orchestrator.md (lines ~128–220) and inline a trimmed version at the top of each of the 7 command files (research, plan, iterate-plan, experiment, implement, validate, handoff). Keep the command description: frontmatter CSO-style (triggering conditions only, no workflow summary).

A3 — ## Retreat Paths per failure-capable command

Add a ## Retreat Paths section to commands/research.md, commands/experiment.md, commands/implement.md, commands/validate.md. Use the failure-mode → next-command mapping from the recommendations doc §7 (e.g., validate empirical fail → /iterate-plan; experiment fail → /iterate-plan or /plan).

A5 — Flip skill to private

Set user-invocable: false on skills/research-workflow-management/SKILL.md frontmatter. The orchestrator still loads it via its own skills: frontmatter.

Acceptance criteria

• Each of the 7 commands has a ## When to use section at the top
• /research, /experiment, /implement, /validate each have a ## Retreat Paths section
• Orchestrator shrinks by ~100 lines after A2 extraction
• research-workflow-management skill has user-invocable: false and no longer appears in user-invocable skill listings
• Zero behavior change — every command produces the same output as before for the same input

Blocks

• feat(ai-research-workflows): core infrastructure — workflow model, profiles, retreat paths #87 (orchestrator refactor depends on trigger content already having been extracted into commands)

Out of scope

• /prototype, /optimize, worktree isolation, empirical validate, --profile flag, orchestrator refactor — those each live in their own issue.