fix: hide TTS/Image toggles in HF Space environment, avoid torchaudio/torch mismatch errors

fix: defer TTS/Image engine preloading for HF Space environment compatibility

refactor: pre-load all models at module level for zeroGPU support

fix: reload CUDA context between translation cards instead of no-op reset

fix: reset KV cache between translations to prevent CUDA state corruption

fix: reload model on CUDA decode failure for SWA models

refactor: extract TTSEngine and ImageGenEngine from engine.py

feat: cards work

fix: more adjustments for images

fix: image prompt and size

feat: reduce image generation size to 450x300

feat: accept instruct parameter in TTSEngine.synthesize()

feat: download correct model and creates the audio

feat: add voice design mode to TTSEngine with female speaking instruction

fix: target language and update docs

feat: update to use local langaguage and added drop down for more languages

fix: dynamic cefr levels

fix: translation does work

feat: add retry loop to LlamaCppTextEngine for translation validation

refactor: delegate MiniCPMTextEngine.generate to core.text_gen.generate_sentences

feat: adjusting for type safety

feat: wrap MiniCPMTextEngine.generate() in retry loop for exact sentence count enforcement

feat: add _validate_lines and _build_retry_prompt helpers to MiniCPMTextEngine

fix: using cuda for llama cpp python

feat: replace TextEngine with MiniCPMTextEngine (llama-cpp-python, chat template)

feat: better feedback for errors

feat: cli params for moe

feat: swap tildeopen for tiny aya water for translation

feat: add EnginePool singleton — manages mutual exclusion between GPU engines

feat: add ImageGenEngine — diffusers Flux2KleinPipeline wrapper with lazy load/unload

feat: add TTSEngine — OmniVoice Python package wrapper with lazy load/unload

feat: add TextEngine — llama-cli subprocess wrapper for Nemotron and TildeOpen

feat: scaffold repo structure with empty files and headers