Switch embeddings to Gemini

Quality: 9.2/10 — tests, Qdrant feedback, token budgeting, adaptive evaluator

Quality: evaluator-optimizer tour validation, terse docs, richer tool descriptions

Upgrade agent + Q&A system prompts with Devin/Cursor patterns

feat: persistent agent memory + README generator

design: premium UI polish + agent reliability fixes

rebrand: update all display name references to Cartographer

Fix thought extraction: unwrap XML tags, keep inner content

Fix thought tag leakage and diagram size

Strip <thought> tags from final answer token stream

Add model selector: dropdown UI + /agent/models endpoint

feat: use Cerebras qwen-3-235b for agent tool calling (tested, confirmed working)

fix: remove Cerebras from AgentService — llama3.1-8b too small for tool use

fix: use correct Cerebras model name (llama3.1-8b)

fix: correct Cerebras model name (llama3.3-70b not llama-3.3-70b)

feat: add Cerebras as primary LLM provider (llama-3.3-70b, 2600 tok/s)

feat: Mermaid diagrams, agent fixes, deployment prep

Major feature: interactive diagrams, MCP integration, and bug fixes

checkpoint: pre-UI-overhaul state

Fix 7 architect-identified issues, add eval harness — targeting 9.5+

Phase 7: MCP server + client — agent now discovers tools via protocol

Switch agent to Groq (free) with Anthropic fallback; fix tool_use_failed error

Phase 5: MCP server + Agentic RAG (ReAct loop)