feat: Add API docs renderer

[milesj]

Summary

Adds ApiDocsRenderer, behind a new renderer_api_docs feature, which renders a markdown API documentation page per type from a SchemaGenerator. It works like the JSON schema renderer: the last type added is the page rendered, and every other type in the generator is a link target.

A page has frontmatter, the type's description, an ## Index table, then a section per property or variant with tags, its comment, and a table of type information (type, default, values, constraints, aliases, env var). Types referenced from the page are linked inline and listed under ## References. Structs render properties, unit enums and unions render variants, headed by the variant name.

Options

• render_tags, render_description, render_link, render_anchor: functions that own how tags, descriptions, type links, and index fragments are rendered, for sites with their own components or heading ids.
• frontmatter: a map of extra key: value entries.
• enum_format: render a unit enum's variants as one table instead of sections.
• include_index, index_page, exclude_aliases, mark_struct_fields_required.
• ApiDocsRenderer::generate_all writes a page for every type into a directory plus an index page listing them.

Changes outside the renderer

• UnionType.variants_names (new, optional). A union derived from an enum records each variant's name by position, so a variant can be labeled. A variant's Schema.name could not carry this: the generator treats any named schema as a standalone reference, so List(Vec<String>) would have become a top-level List definition in JSON schema and TypeScript output. Mirrors how EnumType::from_schemas moves names out of variant schemas. Hand-built unions carry None, and serialization skips it.
• SchemaGenerator::add moves a re-added type to the end. Without this, adding a type that an earlier type had nested left the old root in place, so a per-type loop silently rendered the wrong page. This reorders two TypeScript snapshots where a fixture re-adds a nested type; content is unchanged.

Reviewer notes

• Composite types render as a single code span with a separate References row, because a markdown link cannot sit inside a code span. Bare references link directly.
• Nullable and flattened fields are not tagged Required, unlike the JSON schema renderer's required list, since serde accepts a missing Option and a flattened field has no key.
• Tags in the enum table mode are plain labels, not routed through render_tags, as its block output cannot sit in a table cell.
• Every named type in the generator gets a page, including aliased scalars, since every named type is a link target.

Book page at book/src/schema/generator/api-docs.md fills the existing "API documentation" placeholder in the summary. Changelog has an Unreleased entry.

Test plan

• cargo test --workspace
• cargo clippy --workspace --all-targets clean
• cargo build -p schematic --no-default-features --features renderer_api_docs
• cargo test -p schematic_types on its own
• Regenerated core and partialize snapshots read, not just accepted
• Tried against a real config crate via [patch.crates-io]

🤖 Generated with Claude Code