feat: optimize latency, implement global MutationObserver for instant audio play, and add user interrupt handlers for click and spacebar

fix: change stream_to_file to write_to_file for new Groq SDK, sanitize all PDF text for Helvetica font

debug: add verbose step-by-step logging and fallback for plain text LLM responses

fix: revert LLM backend to use the new active NVIDIA NIM API key

fix: switch to gemini-2.0-flash LLM model as primary backend to resolve unauthorized NVIDIA credentials

Switch agent model execution to NVIDIA NIM API running diffusiongemma-26b-a4b-it

Revert agent to use Gemini API model execution to fix Transformers import crash and resource limitations on HF

Integrate Gemma 3 (AutoModelForMultimodalLM & Processor) into agent pipeline

SHL Assessment Recommender - Full implementation with hybrid retrieval, slot-based conversation analysis, safety guards, and evaluation harness