MCP UI: writes/role-access switches, health + APP_URL warning, server-backed tool/context estimates

[nicdavidson]

Wire the newer MCP backend into Kevin's editor: settings switches, health / APP_URL warning, real tool and context numbers.

What

1. Settings tab (Kevin's df-mcp-settings, Serving or Authentication section as fits): two switches — Allow writes (allow_writes; off = server is read-only, write verbs and writing custom tools never registered) and Require role access (require_role_access; on = only roles granted access to this MCP service can connect, admins always pass). Short helper text for each. Hidden for system_mcp where not applicable. Saved via Kevin's existing whole-service PUT. Kevin's "read-only" computation in mcp-effective.ts / Tools side panel must account for allow_writes=false (write tools not served), and his Create flow should keep defaults consistent with the backend (new services: require_role_access true, allow_writes per his read-only/read-write choice or true).
2. Health chip + APP_URL warning in Kevin's details header (df-mcp-details): call /_internal/ai/mcp-health, show a chip (ok / warn / error) with the first non-ok check's message; the app_url check is the APP_URL warning. Also: the endpoint URL Kevin shows (currently window.location.origin) should prefer APP_URL-based URL when the health report provides it, or at least warn when they differ. Port from Nic's df-mcp-exposure header (healthMessage / healthVariant).
3. Tools / context estimation: Kevin's preview and side panel estimate tokens client-side (flat 81 tokens/tool, lazy assumed >8000 tokens). Replace with Nic's model: call /_internal/ai/mcp-catalog (as admin; for role preview the other agent adds role selection — expose a way to pass role_id/app_id) for the real tool count, bytes, lazy decision; fall back to Kevin's client-side simulation when the endpoint fails or the form is dirty (unsaved). Use the daemon's byte threshold (32 KiB, LAZY_THRESHOLD_BYTES) and Nic's tokensPerTurn / shapeCatalog / shapeFixedCatalog for estimates, including the fixed system_mcp catalog. Show the lazy-loading explanation (full vs facade size; search_tools → describe_tool → call_tool → fetch_more) where Kevin has his "on-demand delivery" line/tooltip.

Context (read first)

• Base branch: feat/mcp-ui = Kevin's MCP redesign (origin/feat/mcp-config-redesign). Kevin's editor lives in src/app/adf-mcp/ (details shell with Connect / Tools / Settings tabs, mcp-store.ts, mcp-effective.ts, mcp-catalog.ts). Keep his layout, tab structure, style and conventions (plain English strings like his components, not Transloco; his component/spec patterns).
• Source to port from: Nic's earlier implementation on local branch feat/mcp-exposure of this repo, under src/app/adf-ai/mcp/ (df-mcp-api.service.ts, mcp-model.ts + spec, df-mcp-access*, df-mcp-exposure*). Read it with git show feat/mcp-exposure:<path>. Port behaviour, adapt to Kevin's structure; do not bring over Nic's page shell, tabs or routing.
• Backend is merged to df-mcp-server develop (PR Adds files tests #77) and is what df-dev runs: GET /_internal/ai/mcp-access?service=&period=, GET /_internal/ai/mcp-catalog?service=&role_id=|app_id=[&client=&lazy_mode=], GET /_internal/ai/mcp-health, config fields allow_writes (bool, default true) and require_role_access (bool; new services default true, existing false). Admin-only endpoints; call them with the admin session like Nic's df-mcp-api.service.ts does.
• Live instance: df-dev http://localhost:8081, admin@dreamfactory.com / passwordpassword. MCP service logistics_mcp (id 15), role logistics_analyst — do not modify those; create your own throwaway services/roles prefixed e2e_mcp_ and delete them after.
• Tests: npx jest src/app/adf-mcp (node_modules: symlink or copy from /data/projects/dreamfactory/df-development/packages/df-admin-interface/node_modules). src/app/shared/utilities/route.spec.ts has one pre-existing failure on develop — ignore it. Add specs for new logic.
• Do NOT rebuild or commit dist/ (another agent works in parallel; dist gets rebuilt once at integration). No docker compose / container / config edits.
• Deliver: commit on your branch, push, open a PR into feat/mcp-ui (not develop) with a "## Behaviour changes" section (what changed, before/after, upgrade impact, how verified). Then stop and report.