Fix model hallucinations by bypassing generation for feedback+question

Fix quiz feedback flow: show next question immediately after feedback

Fix vocab quiz feedback loop and question format

Fix numbered option detection and navigation intent handling

Update lesson welcome to list all vocabulary words

Separate feedback from next question to fix model confusion

Simplify vocab quiz to arabic_to_english only

Fix vocab quiz grading by detecting actual question type

Add comprehensive logging to quiz grading and feedback

Integrate ContentAgent for grammar quiz generation with RAG fallback

Implement grammar quiz with grading and feedback

Fix vocab quiz score counting bug

Add few-shot examples to teaching prompts for better format adherence

Add progress context to prompts and improve numbered options enforcement

Default to vocab on affirmative at lesson_start

Fix: Don't leak quiz answers in questions

Fix: Simplify lesson_start template for better compliance

Fix grading agent: Use chat template + proper generation config

Fix: Don't include navigation input in quiz prompts

Implement vocab quiz flow with answer grading

Refactor to state machine with extensive logging

Add mandatory numbered options to GRAMMAR_EXPLANATION + implement routing

Add vocab quiz question template routing (WIP)

Fix flashcard visibility + strengthen VOCAB_BATCH_INTRO prompt

Make numbered options mandatory in LESSON_WELCOME prompt

Add prompt template logging to orchestrator

Implement flashcard visibility control based on session state

Update lesson 1 to have only 1 grammar topic

Implement flippable flashcards + batch size 3

Use templates! Route to correct template based on session stage

Add comprehensive logging to process_message and update template

Disable automatic quiz pre-generation for now

Pre-generate quizzes via ContentAgent when lesson section starts

Fix vocabulary and logging in orchestrator

Fix syntax error in templates.py

Route all chat messages through orchestrator

Restore ContentAgent (Agent 3) with RAG integration

Add merge_and_unload() and comprehensive model logging

Fix PromptTemplate usage in orchestrator

Implement lazy model loading for ZeroGPU compatibility

Force SentenceTransformer to CPU for ZeroGPU compatibility

Fix model generation with stable sampling and speed optimizations

Fix LoRA adapter loading and add comprehensive logging

Use RAG/Pinecone for lesson cache initialization

Add TeachingAgent and integrate with Orchestrator (TDD)

Rag clean up + missing specs

Address Sourcery feedback: fix CUDA memory check and remove unused quantization config

Use HF_TOKEN from environment for authenticated downloads

Add src/models module for HuggingFace model loading

Address comments