Add Cursor as an MCP client

[sunishsheth2009]

Summary

Adds Cursor as an MCP client in ucode, so ucode configure mcp can register Databricks managed MCP servers into ~/.cursor/mcp.json, plus a ucode cursor launcher.

Cursor is MCP-only: cursor-agent runs models on the user's own Cursor account and exposes no gateway base URL, so ucode configures no models for it (it stays out of agents.__init__._MODULES). What ucode does is register Databricks MCP servers in Cursor's config via the existing ucode configure mcp flow.

Auth: the uniform ucode mcp-proxy bridge (no per-client logic)

This PR is reworked on top of the ucode mcp-proxy foundation (#201). Cursor uses the exact same mechanism as every other client: each Databricks MCP server is registered as a local stdio server that runs ucode mcp-proxy, which bridges to the Databricks streamable-HTTP endpoint and mints a fresh OAuth token per request. So there is no literal token in ~/.cursor/mcp.json, no ${env:...} expansion, and no background refresh thread — Cursor just spawns the proxy like any stdio server, and auth never expires mid-session.

This also sidesteps the earlier DCR failure: with no bearer in the config, cursor-agent doesn't attempt the dynamic-client-registration flow that Databricks managed MCP doesn't support.

Depends on #201 ("Route all MCP clients through a uniform ucode mcp-proxy stdio bridge"). This branch is built on top of that commit — merge #201 first. The proxy commit shows in this PR's diff until #201 lands.

Changes

• src/ucode/agents/cursor.py (new): build_mcp_server_entry(argv) → {command, args} stdio entry; write_/remove_mcp_server_config merge surgically into the shared ~/.cursor/mcp.json (preserving the user's own entries); launch() is a thin exec of cursor-agent (no token wiring — the proxy authenticates itself).
• src/ucode/mcp.py: register cursor in MCP_CLIENTS; configure/remove branches dispatch the proxy argv to cursor; MCP_ONLY_CLIENTS so cursor is eligible whenever cursor-agent is installed (it never lands in available_tools).
• src/ucode/cli.py: ucode cursor launcher; accept cursor in --agents (partitioned out of the model-agent path, workspace-only configure).
• src/ucode/agents/__init__.py: comment documenting why cursor is intentionally not a model-routing agent.
• Tests (tests/test_agent_cursor.py, cursor coverage in tests/test_mcp.py) + README.

Verification

Verified end-to-end against a live workspace (uc-dogfood, e2-dogfood UC):

• ucode mcp-proxy stdio handshake against .../api/2.0/mcp/functions/system/ai: initialize → DatabricksMCPServer, tools/list → 4 live tools (system__ai__ai_classify, ai_extract, ai_parse_document, python_exec).
• configure_client_mcp_server("cursor", ...) through the real code path writes the correct {command, args} proxy entry into ~/.cursor/mcp.json (merging, not clobbering).
• cursor-agent is installed and reads ~/.cursor/mcp.json; the registered entry is the proxy argv, so auth is handled by ucode mcp-proxy per request.

uv run pytest → 876 passed (2 pre-existing test_e2e_user_agent network failures unrelated to this change); uv run ruff format --check and uv run ty check src/ clean.

This pull request and its description were written by Isaac.

[sunishsheth2009]

Reworked on top of the uniform ucode mcp-proxy bridge.

Dependency: this PR is stacked on #201 — merge that first. Until #201 lands, its proxy commit appears in this PR's diff.

What changed from the earlier revision:

• Dropped the literal-token-in-config + background-refresh-thread approach.
• Cursor now registers each Databricks MCP server as a local stdio server running ucode mcp-proxy (same as every other client), so ~/.cursor/mcp.json carries no bearer token, there's no ${env:...} expansion, and auth refreshes per request with no per-client logic.

Re-verified end-to-end against uc-dogfood (proxy handshake → 4 live system.ai tools; real registration writes the correct stdio entry).