refactor(aidd-dev): lean the plan skill for small models #1075

Problem

aidd-dev:01-plan is 9 files and 1750 words: four actions (gather, explore, wireframe, plan) with their own Test sections, a status reference only the plan itself cites, and templates that restate rules their action already carries. Under opencode with Mistral small and medium models the router is not followed: medium reads the templates without reading 04-plan, writes the plan files inside the skill folder, and skips the review. The plan also carries no explicit assessment of what the feature relies on, so a coherent plan can still be non-executable.

Scope

• Same method as the brainstorm skill: a mermaid router with action paths, three actions named capture, explore, conclude, a short process of recipe steps, one Output per action, rules only where the action owns them.
• Fold wireframe into conclude (no own input or output) and move the status lifecycle into the plan template.
• explore gathers findings from a lean table reference, with a loop that returns to reading while a finding lacks its evidence, and asks the user only for a blocking decision.
• Plan template: Impact (runtime, regression, architecture, documentation), Rules, External sources, Decisions. This is the plan-side part of feat(aidd-dev/plan): validate feature runtime prerequisites before implementation #955, feat(aidd-dev/plan): conditional Phase 0 "regression safety net" with regression risk assessment #896, feat(aidd-dev): make architecture impact explicit during planning #889 and feat(aidd-dev): carry documentation impact through feature delivery #927.
• Phase template: Test Scope as a table, User Journey removed, placeholders and comments deleted from the written files.
• conclude scores the plan against a short checks table and asks to approve, correct or explore more.

Acceptance criteria

• The skill goes to about 1300 words with the pre-commit gates passing.
• Claude haiku and sonnet: the plan is written to aidd_docs/tasks/<yyyy_mm>/<yyyy_mm_dd>_<slug>/ on the nine scenarios (no source, feature, interface, refactor, trivial fix, correction, explore more, source in a task folder, auto mode), with no file written elsewhere.
• mistral-medium-2604: the plan files are written on the feature scenarios and the skill never edits application code. Not met yet on the trivial-fix scenario (0 of 3).
• Original and lean skill replayed on the same scenarios, plan quality on par or better.
• The score names what was not dug into, left unverified or assumed.

Prior art in this repo

• Brainstorm: refactor(aidd-refine): lean the brainstorm skill for small models #1071 and refactor(aidd-refine): lean the brainstorm skill for small models #1072.
• feat(aidd-dev/plan): validate feature runtime prerequisites before implementation #955, feat(aidd-dev/plan): conditional Phase 0 "regression safety net" with regression risk assessment #896, feat(aidd-dev): make architecture impact explicit during planning #889, feat(aidd-dev): carry documentation impact through feature delivery #927 (plan-side only), refactor(framework): replace skill test sections with scoped rules #985 (Rules naming).
• plugins/aidd-dev/skills/02-implement/actions/02-execute.md reads phase-<n>.md; 05-review reads ## Architecture projection.

Out of scope

• The implement and review sides of feat(aidd-dev): carry documentation impact through feature delivery #927 and feat(aidd-dev/plan): conditional Phase 0 "regression safety net" with regression risk assessment #896.
• auto mode of aidd-orchestrator:01-sdlc.
• A script that creates the plan files: tried, no gain measured on medium.