Follow-up to #103.

## Problem

The mdx lowering serves its original purpose well — prose as *metadata about* units: each section's prose (heading included) is merged into one `` string wrapping that section's JSX. Two properties make it the wrong shape for consumers that want to *render* the document (e.g. a static site builder where `.op.mdx` files are pages):

1. **Interleaving is lost.** Prose parts and JSX children are collected separately per section, so `paragraph, , paragraph` can't be reconstructed — a renderer can only emit "all prose, then the components".
2. **The markdown stays a string**, forcing the consumer to bundle a second markdown parser (in JS, inside QuickJS) even though the lowering already has the full mdast.

## Proposal: a document-mode lowering

Lower prose all the way to intrinsic JSX nodes, in document order — the same thing mdx-js does:

```mdx
# Benchmarks

Here is the setup.

And here is what it means.
```

context mode (today):

```tsx

```

document mode (proposed):

```tsx
<>

Benchmarks

Here is the setup.

And here is what it means.

```

Since `h()` passes string tags through as inert `{type, props, children}` data, the consumer supplies per-tag mapper functions to serialize into its sink (HTML, or anything else) — prose and components become the same substance, and interleaving stops being a "mode" at all: document order is just child order. Inline constructs (emphasis, links, code spans) come from the mdast the parser already produces; no second markdown parser anywhere.

Context mode stays as-is for grounding/prompt consumers — this is an additional lowering target over the same parse (ESM hoisting, frontmatter, positions all shared), selectable by the caller (and eventually by the #103 loader).

## Motivating consumer

Static site builder as an esto script: posts are `.op.mdx` modules imported by the build script; the default export's tree is walked into HTML by per-tag mappers, only for pages whose input signature moved. Context mode is usable for this today (the value is verbatim markdown), but costs a vendored JS markdown parser and restricts component placement to section boundaries — document mode removes both.

No source file or test is named. Start by tracing the existing MDX context-mode lowering and the shared mdast parse path, including how ESM hoisting, frontmatter, and positions are preserved. Done means a caller-selectable document mode emits intrinsic JSX nodes in document order while context mode remains unchanged.

Written by the indexing model from the issue text.