Bind per-agent skills into openclaw.json (step 2)

[yikkuro]

Summary

Step 2 of the agent-skills binding feature: MicroClaw now writes each agent's skill allowlist into config.agents.list[*].skills when it generates/reconciles openclaw.json. This is the OpenClaw-native way to bind skills per agent — at runtime resolveEffectiveAgentSkillFilter reads agents.list[<id>].skills and, when present, uses it to filter the agent's skills (replacing agents.defaults.skills).

Skills are read per-agent from the catalog (agent-catalog.ts skills field / getAgentSkills) — never hardcoded — so later steps can vary skills per agent without touching this plumbing. Since step 1 defaults every agent to all 23 skills, every entry currently receives the full list.

Changes (desktop/src/agent-personas.ts)

• Added skills: readonly string[] to AgentPersona, populated from the catalog in AGENT_PERSONAS.
• createAgentEntry writes entry.skills = [...persona.skills] (fresh array copy) into every newly generated entry.
• ensureAgentPersonasConfig enforces skills authoritatively and idempotently for catalog-known agents (overwrites stale values with a fresh copy); custom/non-catalog agent ids are left untouched. Existing changed detection still correctly reports true when skills are newly added.
• All existing reconcile behavior (id normalization, list/entries merge, default handling, workspaces) preserved.

Tests (desktop/src/agent-personas.test.ts)

Added coverage: generated entries' skills deep-equal catalog; distinct array instances per agent; stale skills overwritten; custom agent left without skills; changed reported true when skills newly added.

Validation

• npm test → 676 passed
• npx tsc --noEmit → clean
• eslint on changed files → clean (pre-existing unrelated main.ts error left as-is)

No UI/renderer/gateway changes.

Stacking

This PR is stacked on and depends on #140 (step 1, still open). Base branch is sunt-microsoft-feat-agent-skills-binding-step1, not main.