Add Fase 7 UI screens (Group A) and integration plan

[danzeroum]

Summary

This PR adds the design and integration plan for Fase 7 Group A — seven new admin/user-facing screens that expose Fase 6 backend functionality (MCP console, A/B experiments, squad memory/RAG, rate limits, model usage, LSP servers, sandbox & third-party skills) and a comprehensive 5-wave frontend integration roadmap to wire all 13 existing screens to real backend APIs.

Changes

Design & UI (docs/design_handoff_forge_telas/design/Forge Telas.dc.html)

Added 7 new screen sections (≈250 lines of component markup):

• U8 · Mapa de Memória (RAG) — two-column layout: agent memory cards (count, top decision, decay) + semantic search panel with TF-IDF recall (idle/loading/results states, lexical-only per ADR 0013)
• A6 · Experimentos A/B — experiment report cards with dual rate bars (A/B), honest verdict (Significant/Inconclusive/InsufficientData), p-value vs α=0.05, seeded-data warning banner
• A7 · Console MCP — two-column: connected servers (status, transport, tool count) + exposed tools table (namespaced, input_schema, per-call policy)
• A5 · Uso por Modelo — table: model name, tier, call volume (bar), cache-hit %, grouped from props.model
• A4 · Rate Limits — three tier cards (Small/Medium/Large, 600s window) with used/cap bars, cost-salvaguard caveat banner
• A7 · Language Servers (LSP) — two-column: server status (indexing state, diagnostics count) + exposed queries as tools (0-indexed positions, scope-gated)
• A6 · Sandbox & Skills de Terceiro — confinement profile (Docker, bollard, fail-closed) + third-party skill lifecycle (install→vet→enable, states: enabled/analyzing/blocked)

All screens follow existing component patterns (sc-if, sc-for, AsyncStatus-like states, color-coded badges, monospace labels per design system).

Integration Plan (docs/PLANO-INTEGRACAO-FRONTEND.md)

New 241-line document defining the 5-wave roadmap to wire frontend to backend:

• Onda 1 — cheap reads (ledger, experiments, model usage, MCP, providers): 5 GET routes, 1 new LedgerStore method, 1 new admin screen
• Onda 2 — simple mutations (prompts CRUD, verify as job, prompt render via sidecar, Origin/Host guard): first POSTs, permission matrix read-only
• Onda 3 — real session with SSE (DurableSession, AgentLoop on server, ScriptedGenerator for e2e without key): transforms dashboard into app server
• Onda 4 — squad live + HITL permission bridge (oneshot-based request queue, SquadEvent→SSE, consensus real): largest design risk, prototyped early
• Onda 5 — memory/RAG, designer→ledger, rate-limit getter, A/B instrumentation, sandbox/doctor/LSP: polish and optional admin surfaces

Principles: copy telemetry.ts pattern (axum handler + in-crate test + e2e spec), no silent mock fallback, AppState growth via Arc<Mutex<…>>, SSE for streaming, consistent JSON error format, schema contracts in schemas/, fail-closed on missing backend.

Component Logic (Forge Telas.dc.html script section)

Added memQuery, memStatus, memResults state + onMemInput, onMemKey, runRecall() methods to simulate TF-IDF recall (client-side in prototype, backend CoreService.Recall in real implementation). Recall filters by term inclusion and scores 0.55–0.95 based on hit ratio.

Notable Details

• **Honesty-

https://claude.ai/code/session_01Ssk6F2txpfNpSLNLH6eV3o