[skills] Add an interactive schema picker to ug skill add --mcp

[xsh310]

What

ug skill add --mcp with no --location errored asking for a typed schema. On an
interactive terminal it now opens a picker of the workspace's schemas that contain
skills, streamed in as discovery finds them, and adds the selected schemas to the
skills MCP connection's scope.

Why

The download path already has an interactive picker; the MCP scope path was the only
add-side flow still requiring a typed --location. This makes the two symmetric.

How

• The picker (in mcp.py, beside the other skill MCP-scope commands) reuses the download
  picker's machinery: the workspace walk (list_all_skills) and the streaming checkbox
  widget. list_all_skills probes one schema per call, so each streamed batch is one
  schema's skill set; the loader renders one row per schema with its exact skill count.
• Already-scoped schemas are flagged and stay selectable; re-selecting is a no-op, since
  adding is additive (removal is ug skill remove --mcp).
• To let both mcp.py (picker) and skills_download.py (download) use the workspace walk
  without a circular import, the UC skills read layer (SkillRef, the API client, and
  list_all_skills) moves into a new skills_api.py that depends only on databricks.py.
• add_skills_command's apply body is extracted into add_skill_locations_to_mcp, shared
  by the command and the picker so each runs a single client setup.
• Non-interactive (no TTY) still errors asking for --location, so scripts fail loudly.

Testing

uv run ruff check, ruff format --check, ty check src, and uv run pytest all pass
(the pre-existing test_e2e_user_agent / test_claude_smart_routing_v2 failures also fail
on the base). The read-layer tests move to test_skills_api.py; added tests for the schema
picker, its command, add_skill_locations_to_mcp / configured_skill_locations, and the CLI
dispatch.

Stack

Part of gh stack #576, on top of #586 (skills-download-workspace-id).

This pull request and its description were written by Isaac.

Screen.Recording.2026-09-12.at.11.35.02.PM.mov

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[sunishsheth2009]

P1: list_all_skills(...) returns (refs, reason) and yields partial results once its 30s wall-clock deadline is hit (or a schema probe errors), but the loader discards the return. When the walk times out on a large/slow workspace, the picker simply stops adding rows and the "Loading more…" footer disappears — indistinguishable from a complete load. The user then scopes only the schemas that happened to arrive in time, believing they've seen everything.

Suggest capturing reason and surfacing it — e.g. append a disabled sentinel row (⚠ schema discovery timed out — showing partial results) or print_warning after the picker closes. Same pattern exists in the download picker's _skills_download_background_loader (#579).