feat(claude-mcp): disable MCP servers per workspace

[pjdoland]

Summary

Adds a per-row Enable/Disable for Workspace toggle on the Claude MCP settings panel. The toggle writes to the same projects.<cwd>.disabledMcpServers[] array in ~/.claude.json that Claude Code itself reads, so a server can be opted out of one Jupyter workspace without removing it from the user scope or affecting other workspaces.

Solution

The disable list is workspace-wide (a flat array of server names keyed by the current working directory), not per-scope, so a user-scope server can still be opted out of a single workspace.

Backend:

• ClaudeMCPServer gains a disabled_for_workspace: bool field, populated during list_servers() from the ~/.claude.json projects.<cwd>.disabledMcpServers[] entry for the current Jupyter root.
• New ClaudeMCPManager.set_server_disabled(name, disabled) reads ~/.claude.json, mutates the disabled list under the current working dir, and writes back via config._atomic_write_json so the on-disk edit shares the same symlink-preserving, mode-preserving, parent-dir-fsync durability contract as the rest of NBI's config writes.
• New PATCH /notebook-intelligence/claude-mcp/<scope>/<name> accepts {"disabled_for_workspace": <bool>}. The handler validates scope and rejects non-bool payloads (stringly-typed "false" no longer flips the row).

Frontend:

• NBIAPI.setClaudeMCPServerDisabled(name, scope, disabled) mirrors the wire contract.
• The settings panel renders disabled rows with a dashed border, italic name, and a "Disabled for workspace" badge. The toggle button is forced visible on disabled rows so users don't have to hover to re-enable.

Testing

• pytest tests/ (718 passed) including 9 new tests in TestWorkspaceDisable: list-time flag stamping, per-cwd scoping, write idempotency, sibling-key preservation on partial updates, missing-config-file creation, and corrupt-config rejection.
• jlpm tsc --noEmit, jlpm lint:check, jlpm jest (131 passed) all green.
• Manually verified the file-on-disk shape: toggling preserves unrelated top-level keys (telemetry, etc.) and sibling project blocks.

Risks and follow-ups

• Concurrency with Claude CLI: a Claude CLI write between our json.load and os.replace would be clobbered. The window is milliseconds and the toggle is user-initiated. The same hazard exists for claude mcp add, which the existing manager already shells out to.
• The PATCH URL captures scope for symmetry with GET/DELETE; the value is validated but ignored on the write path since the disable list is not per-scope. A future per-scope disable list would be a non-breaking extension.
• If the toggled server has no canonical definition visible from the current cwd, the response synthesizes a minimal record (the panel's refresh() then re-reads the truth).

Closes #283

[mbektas]

validate_scope import is missing

[pjdoland]

Good catch, fixed in 05828e8. Added two regression tests so this class of import-miss can't slip past CI again:

1. TestPatchEndpointDispatches (tests/test_claude_mcp_manager.py) drives ClaudeMCPDetailHandler through tornado.testing with auth stubbed. Confirmed it fails with the exact NameError when the import is removed.
2. ui-tests/tests/claude-mcp-patch.spec.ts hits the running JupyterLab server's PATCH route from a browser context via Galata, so routing, auth middleware, and the prepare() chain are all exercised end-to-end.

Both pass with the fix and fail without it.