Android Dev Build #233.1 (439d713)

Automated Android build artifacts for commit 439d7131aa1d49aeb391f519bf73f2dbe67a4469.
Workflow run: https://github.com/LettuceAI/app/actions/runs/28481221536

Changes since previous Android build android-dev-232-1-f69443a

Compare: android-dev-232-1-f69443a...439d713

• 439d713 feat(llama): persist and display per-message MTP speculative-decoding stats
• bda1c10 feat(llama): localize model editor multi-GPU controls with bottom-menu pickers
• d0835d8 feat(llama): explain multi-GPU distribution and KV modes with bottom-menu pickers
• bda0c7d fix(llama): reserve KV cache VRAM per GPU when distributing layers
• 30486c4 fix(memory): fix re-embedding loop; dedupe lorebook scan and embedding pref reads
• 6d71bc7 feat(llama): multi-GPU layer distribution + KV cache placement (#69)
• 8fc50c3 fix(discovery): build card image URLs for new ct-cards storage host with query params
• e9a86e8 fix(macos): keep native decorations and opaque window so traffic lights and rounded corners render
• 26be727 fix(llama): check abort before loading engine so immediate aborts skip model load
• e3c0623 fix(memory): stamp embedding signature from actual vector to stop spurious vector migrations
• 8ef05ca fix(embedding): resolve version-specific tokenizer file instead of hardcoded tokenizer.json
• 7be1c4b feat(metrics): local LLM performance telemetry, graphs, and per-message view (#68)
• f407b5d feat(openrouter): send roleplay category and current app-attribution headers
• c897374 fix(memory): honor companion character mode when resolving shared-memory owner
• 839e085 ci(desktop): bundle ONNX Runtime in linux and windows dev builds
• 432c799 fix(tts): bundle ONNX Runtime on Windows and route Kokoro through shared runtime init
• fce4d5d fix(chat-appearance): stop per-conversation override from injecting defaults that clobber global settings
• e8dad1c ci(desktop-release): sync Windows CUDA/Vulkan setup with dev workflow
• 7b3889f ci(desktop-release): pin Windows CUDA build to windows-2022