Upgrade default model to MiniCPM-V 4.6; compact output panel into tabs

Deploy student-first exam coach

Docs+engine: drop VoxCPM2/read-aloud to match shipped app (README card + study_engine)

Perf: resident MiniCPM-V + boot warm-up + live status + honest fallback reasons (app.py & study_engine.py together)

Llama Champion: selectable llama.cpp engine (small MiniCPM4 0.5B GGUF via llama-cpp-python, CPU)

MiniCPM-V-4.5 is now the primary engine (text + photo); drop 4.1-8B and 0.5B/1B; Nemotron-4B alternate; README to match

One text model at a time (24GB safety) + voice gating + pinned model metadata

Prefetch model weights on CPU before the GPU call so a cold secondary model no longer times out the ZeroGPU budget

Add drill answer keys (validated on staging)

Add VoxCPM2 read-aloud (5th model; validated on staging)

Add download/print packet button (validated on staging)

Add model picker: MiniCPM-8B / NVIDIA Nemotron-4B / MiniCPM5-1B (validated on staging)

Add MiniCPM-V syllabus-snap (vision) + torchvision (validated on staging)

Live coach timer: ship coach_state + UI together (validated on staging)

Add 'what to skip' triage line

Isolate GPU call (graceful fallback on worker failure) + skip ZeroGPU preload

Resilient generate: never hard-crash on a model-path error

Sync from GitHub: triage time-budget fix, MiniCPM-driven drills, UI polish, sub-4B model option, LICENSE

[Codex] Add optional Nemotron runway and public trace

[Codex] Preserve generated markdown in Space

[Codex] Gate sample generation for final Space

Sanitize MiniCPM reasoning output with Codex

Make Space product-first and ZeroGPU-ready with Codex

Focus MiniCPM demo and readable UI with Codex

Polish UI and hardware honesty with Codex

Update llama runtime evidence from Codex 7fb3aaf

Deploy Exam Panic Rescue staging Space