Fix .md export losing method names, signatures and section labels

[mishig25]

Problem

The .md (llms / plain-text) twin of every [[autodoc]] class page lost all of its method headings. E.g. hf_api.md has zero #### headings: the docstring bodies are there, but names, anchors and signatures are gone, and **Parameters:** / **Returns:** are missing, so the return type gets glued onto the last parameter line:

To disable authentication, pass `False`.[DiscussionStatusChange](…)the status change event

Cause

strip_html_from_markdown matched the legacy autodoc markup:

r'<div[^>]*class="docstring[^"]*"[^>]*>.*?<docstring>.*?</docstring>.*?</div>'

but since #797 python emits the svelte component with metadata as props instead — <Docstring name={"…"} anchor={"…"} parameters={[…]}> — and no <docstring> tag exists anymore. So no block ever matched process_docstring_block; everything fell through to strip_remaining_html, which deleted the <Docstring …> opening tag wholesale (name + anchor + signature) and the section tags along with it.

Fix

• extract_docstring_info parses the current form: props from the opening tag, sections from the body. Prop values are decoded with json.JSONDecoder().raw_decode, so } / > inside type annotations and defaults can't truncate the parse.
• strip_html_from_markdown matches <Docstring …> … </Docstring> (the opening tag ends at the last > on its line — the same single-line-JSON contract kit/preprocessors/docstring.js relies on) and keeps a blank line before the description that follows it.
• process_docstring_block renders the call signature from the parameters prop, and now also emits extra parameter groups, **Yields:** and **Raises:**, which were dropped entirely before. Types are backtick-wrapped only when they carry no markup — wrapping a resolved [Name](url) doc link would break it. Properties (isGetSetDescriptor) get no signature; a no-arg method still gets foo().

Section order (signature → parameters → returns → description) matches how Docstring.svelte renders the page.

Result

Verified end-to-end with a real [[autodoc]] HfApi build:

#### merge_pull_request[[huggingface_hub.HfApi.merge_pull_request]]

```python
merge_pull_request(repo_id: str, discussion_num: int, token: bool | str | None = None, comment: str | None = None, repo_type: str | None = None)
```

[Source](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/hf_api.py#L7921)

**Parameters:**

repo_id (`str`) : A namespace (user or an organization) and a repo name separated by a `/`.
…

**Returns:** `DiscussionStatusChange`

the status change event

Merges a Pull Request.

Four regression tests added (class-level block with a method, getset descriptor, > inside props, no-arg signature). Full suite passes.

Not fixed here

• get_signature_component_markdown (the embeddings/meilisearch path) also has <name> commented out and only emits Docstring for: {anchor} — but that predates Simplify docstring pipeline: python emits the <Docstring> component directly #797 and doesn't affect the .md export, so it's left alone.
• Markdown autolinks in prose (<https://huggingface.co>) are eaten by the generic tag regex in strip_remaining_html (hence "Defaults to ." in parameter descriptions). Separate pre-existing bug.

🤖 Generated with Claude Code