Markdown is the lingua franca of large language models. A typed Markdown ADT makes document structure visible to the type system, enabling contracts that verify structural properties of agent output.

Two mutually defined ADTs: MdBlock for block-level elements and MdInline for inline-level content. The two-level design makes illegal states unrepresentable at the type level (a heading cannot contain another heading).

Key advantage over Json: does not depend on Map<K,V> -- Markdown has no key-value structure, only arrays and strings. This means it could potentially be implemented before Json.

Parse/render functions (md_parse, md_render) are pure. Accessor functions (md_has_heading, md_has_code_block, md_extract_code_blocks) enable contracts on Markdown structure. Document conversion (PDF, Word, HTML to Markdown) uses IO effect + host bindings to external tools like MarkItDown.

Refinement type examples:

type HasTitle = { @MdBlock | md_has_heading(@MdBlock.0, 1) };
type HasVeraCode = { @MdBlock | md_has_code_block(@MdBlock.0, "vera") };

Dependencies: #52 (dynamic strings), #134 (string ops), #132 (compound arrays).

Spec reference: Chapter 9, Section 9.7.3.