feat: Slice 1 polish - MiMo V2.5 model, visible reasoning, image attachments, mobile keyboard fix

- Switch conversation agent to xiaomi/mimo-v2.5 (multimodal: text + image)
- Render native reasoning parts as live 'Thinking trace' (streaming open,
  collapsed after completion); inline <think> extraction for streaming models
- Image attachments: picker (up to 4, 10MB each), base64 to Flue
  AgentPromptImage, authenticated blob-URL replay for historical images
- Mobile keyboard viewport fix: visual-viewport hook, fixed shell,
  interactive-widget=resizes-content, header pinned, composer follows keyboard
- Conversation to Signal to proposed Work: Convex persistence, Effect
  validation in @code/work-os, Work cards with exact source provenance
- Streamdown markdown + Mermaid chart rendering in chat messages
- Flue tool turns hidden, reasoning-containing turns remain visible
- Frontend regression tests: keyboard viewport, responsive shell,
  attachment overflow, authenticated images, reasoning traces, transforms
- .env.example updated to xiaomi/mimo-v2.5 config

.env.example

@@ -25,12 +25,12 @@ DAEMON_COMMAND_LEASE_MS=60000
 FLUE_DB_TOKEN=replace-with-a-long-random-token
 
 # Agent model provider
-AGENT_MODEL_PROVIDER=cheaptricks
-AGENT_MODEL_NAME=glm-5.2
+AGENT_MODEL_PROVIDER=xiaomi
+AGENT_MODEL_NAME=mimo-v2.5
 AGENT_MODEL_API=openai-completions
 AGENT_MODEL_BASE_URL=https://ai.example.com/v1
 AGENT_MODEL_API_KEY=replace-with-provider-api-key
-AGENT_MODEL_CONTEXT_WINDOW=262000
+AGENT_MODEL_CONTEXT_WINDOW=1048576
 AGENT_MODEL_MAX_TOKENS=131072
 
 # Self-hosted git (Gitea) — canonical repository host and API token