bug fixes

What's Changed

• Clean Execution Runtime & Monkeypatch Removal: Eliminated ad-hoc runtime overrides on pd.DataFrame and scikit-learn in execution.py, restoring standard, predictable Python library semantics with AST-based security controls.
• Universal Data Engineering Standards in Prompting: Enforced core data science best practices across code generation prompts (numeric isolation for aggregations/correlations, NaN inspection and imputation, linear model coefficient array flattening, and strict schema adherence).
• Context-Enriched Self-Correction: Augmented execution error prompts to include the executed <failed_code>, live <runtime_diagnostics> (actual DataFrame shape, columns, precise dtypes, and null counts), enabling models to self-heal dynamically across any dataset.
• Remote GPU Acceleration Tuning: Optimized Ollama context budgeting (LLM_NUM_CTX=4096) ensuring 100% GPU VRAM residence (0% CPU offload) on 8 GB cards (e.g. RTX 2060 SUPER) with ~85 t/s on qwen3:4b and ~55 t/s on qwen2.5-coder:7b over Tailscale.
• KaTeX Math Rendering: Enhanced LaTeX equation delimiters ($$...$$, \[...\], \(...\)), normalized escaped delimiters, and fixed paragraph DOM nesting in frontend markdown rendering.