Premium 429 retry-with-backoff so prebake self-paces under tier caps

Premium-mode cap overrides for richer prebaked artifacts

Phase 2: premium tier opt-in via Anthropic Claude

Fix tour layout, synthesis reliability, and cache safety

fix(synthesis): block Cerebras 8B from synthesis fallback chain

fix: treat Gemini 500 internal error as fallback trigger

fix(tour): always use strongest model for synthesis

improve(tour): signal-checklist DONE criterion, gaps field, clean domain terms

fix(sambanova): switch to DeepSeek-V3.1 (Meta-Llama-3.1-405B deprecated)

fix: skip Gemma 4 inside the fallback chain, not before it

fix: skip thinking models for forced DONE and Phase 3 synthesis

fix: cascade order, Cerebras model, forced DONE char limits

fix(cascade): prioritize SambaNova 405B over Gemma4 31B and Cerebras

fix: Cerebras model slug, DONE template, trace gap display

fix(generation): surface clean error when all providers exhausted

fix: Cerebras model slug rename + SambaNova cascade ordering

fix: treat 503 high-demand as exhausted, trigger provider fallback

fix: stop SDK retries, add timeout fallback, increase DONE budget to 3000

fix: increase token budget and timeout for Gemma 4 31B

fix: add 30s timeout to all OpenAI client instantiations

feat: add Gemma 4 31B as fallback after Gemini 2.5 Flash

fix: provider exhaustion window prevents parallel cascade to Groq

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

fix: cap max_tokens per provider to avoid 400 errors on fallback

fix: skip response_format=json_object for Gemini (causes ~100-token truncation)

fix: preserve max_tokens in fallback, route sambanova/mistral correctly

fix: quality-first provider order + prompt audit fixes

feat: add SambaNova (405B) and Mistral to free-tier fallback chain

feat: upgrade primary LLM from Gemma 4 31B to Gemini 2.5 Flash

fix: keep timeout only on OpenRouter client, remove from Gemini/Cerebras

fix: treat 404 model-not-found as exhausted, trigger fallback chain

fix: 45s timeout on all LLM clients + slow contextual star beat

fix: three bugs from re-index logs

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

Keep generation free-tier only — revert quality split

Fix repo disappearing during re-index + upgrade enrichment models

rebrand: update all display name references to Cartographer

fix: increase RAG max_tokens from 1024/1536 to 4096 (match agent mode)

fix: properly handle all Gemma 4 thought-block output patterns

fix: add _strip_thought_tokens to strip Gemma 4 reasoning from RAG answers

fix: _reset_to_primary was resetting to OpenRouter instead of Gemma 4

feat: Gemma 4 as primary for RAG generation (reverts back)

fix: revert RAG to Cerebras primary, Gemma 4 as fallback

feat: parent-document retrieval + Gemma 4 as RAG primary

Fix stale Cerebras model name in generation.py log message

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 provider for GenerationService (diagrams/tour)

fix: commit missing modified files (diagram_service, generation, DiagramView, ExploreView)

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