feat: skills — user-authored on-demand instruction bundles

[DoyleDev]

Summary

Adds first-class skills support using Anthropic's progressive-disclosure pattern, but executed locally through the Databricks AI Gateway so it works across every model (Claude / GPT / Gemini / Llama / Qwen / custom endpoints) — no beta headers or sandboxed containers required.

How it works

1. At chat time, an `<available_skills>` block is injected into the system prompt listing each enabled skill's slug + one-line description.
2. The model sees the manifest. When relevant, it calls the new built-in `load_skill(slug)` tool.
3. The renderer reads the corresponding `SKILL.md` from disk and returns its body as the tool result. The model now has the full instructions for its next turn.

Discovery

• `~/.mason/skills//SKILL.md` — user-authored, edited in Settings
• `~/.ai-dev-kit/repo/databricks-skills//SKILL.md` — auto-discovered, read-only

Forward-compatible with ai-dev-kit's planned skills-only releases; existing MCP path continues to work in parallel.

UI

• Settings → Skills: list discovered skills with source / description / enable toggle. `+ New Skill` opens an editor with name + description + markdown body. Edit/delete on user skills only. Auto-load toggle mirrors the MCP tools pattern.
• `+` icon → Skills: modal mirroring the Tools modal — quick per-skill enable/disable with grouped source + counter.

Persistence

• `~/.mason/config/skills.json` — `{ disabledSkills: [], autoLoadSkills: true }`
• Anthropic-compatible format (YAML frontmatter `name` / `description` + markdown body), so user-authored skills are portable.

Bumps to 1.4.0 — meaningful new surface area, not just a patch.

Test plan

• Create a skill via Settings → Skills → + New Skill. Verify it lands at `~/.mason/skills//SKILL.md` with frontmatter.
• Toggle the skill off, verify the LLM no longer sees it (check `<available_skills>` block isn't there).
• Send a message that matches the skill's description, verify the model calls `load_skill` and the tool result contains the body.
• Edit an existing user skill, save, verify changes persist.
• Delete a user skill, verify the directory is removed.
• With ai-dev-kit installed, verify its ~30 skills show up under "ai-dev-kit" source and are read-only.
• `+` → Skills modal: quick toggle works, counter reflects.
• Escape closes Skills modal and editor.
• Try with Sonnet, Opus, GPT, Gemini — same behavior.

This pull request and its description were written by Isaac.