Fix AI insight leaking thinking content + make insights on-demand instead of auto-generated

Fix: detect and strip Nemotron reasoning narration (model talking to itself) from AI output

Fix AI output: aggressively strip thinking tags, markdown, field prefixes from all LLM responses

UX overhaul + Gradio migration + remove llama.cpp

fix: show CLOUD GPU status instead of offline, strip orphan </think> tags, increase Modal timeout to 180s

Remove llama-cpp-python and lighten Dockerfile; Modal handles inference

Add Modal GPU inference support for faster LLM responses

fix: restore llama.cpp with source build, use fine-tuned GGUF model

fix: switch to microsoft/Phi-3-mini-4k-instruct (Gemma 2B down, LoRA not on serverless API)

feat: add HF Inference API with LoRA model, deterministic fallback

refactor: remove llama-cpp entirely, use deterministic mock-only mode

fix: load LLM in background thread so Space stays healthy during 2.84 GB cold start

fix: correct LLM model path default, fix chat 'error: format only' leak, surface load errors

feat: browser-local engine, Zerodha dashboard, historical events, chatbot, per-user isolation

fix: mentor returns real roast from numeric summary when LLM output is empty/malformed

fix: harden LLM calls, cast numpy to native floats, pin llama-cpp-python wheel

fix: accept percent value (0-100) for trades, add MOCK_LLM flag, full Playwright E2E

Phase 2/3: Gradio Server backend, CRT frontend, engine, agents, mentor, tests, CI/CD