Streaming-First
Optimized for token-by-token LLM output. No flickering, no layout jumps. Partial math and diagrams are deferred until complete.
Math Protection
Distinguishes $100 (currency) from $x^2$ (LaTeX) using delimiter analysis and currency heuristics. No false positives.
Self-Correcting
Pluggable callback detects broken Mermaid or KaTeX from LLM output and can regenerate the block automatically.