feat(learn-mentor): add learning-mentor skill

[LeaYeh]

What

Adds learn-mentor, a new skill whose posture is inverted from the rest of the ecosystem: it refuses to just-complete the task and instead drives the user to truly master a topic via the right papers/articles.

Given a learning goal + task, it decomposes the goal into stages and runs each through a four-phase loop, withholding exactly the skill the goal targets so the user reaches mastery by their own hand.

The per-stage loop

Phase	Role
A	Surface gaps (Socratic) — back-cast prerequisites from the stage goal; gaps emerge from where the user stumbles
C	Guided reading (center of gravity) — live-sourced, vetted, laddered (explainer → primary source), read section-by-section with inline jargon/math decoding
D	Implement to the stage goal — conditional withholding: syntax → user writes code; theory → user writes pseudo-code, mentor supplies plumbing
B	Teach-back — lean diagnostic; failure routes back to C

Learning sediments into docs/learning/<slug>.md (per-goal Markdown + embedded Mermaid + glossary) and resumes across sessions.

Design provenance

Designed via a grill-graph session — 11 resolved decision nodes captured in decision-graph.md, converged into spec.md. Five of my initial assumptions were corrected during grilling:

1. A before C — entry is goal+task, not a known concept (you don't know what you lack at the start)
2. Conditional withholding — whether the agent may write code depends on the learning goal
3. Theory bar = pseudo-code — the production target for theory stages
4. B is a lean diagnostic — C is the recovery hub, not a heavyweight exam
5. Persist into repo docs/ — learning settles with the project it serves

Commits

• docs(learn-mentor) — design decision-graph + spec
• feat(learn-mentor) — the skill (plugin.json + SKILL.md)
• feat(settings) — marketplace registration + enable

🤖 Generated with Claude Code