Simplify docstring pipeline: python emits the <Docstring> component directly

[mishig25]

What

Docstring rendering used to go through two regex transformation layers: autodoc.py serialized metadata into custom tags (<docstring><name>…</name><anchor>…</anchor><parameters>…</parameters>…<paramgroups>N</paramgroups>), and kit/preprocessors/docstring.js regex-parsed them all back out to build a <Docstring …/> component.

Now python emits the component directly: the <Docstring> open tag carries all metadata props as final values (name/anchor/source/parameters/isGetSetDescriptor, with MDX escaping undone since JS string/JSON props need none), and the component body carries only the markdown-bearing sections (parameter descriptions, return/yield/raise types & descriptions). The kit preprocessor's only remaining job is rendering those sections with mdsvex into the remaining props — it no longer parses names, anchors, sources, signatures, or group counts.

Why the JS stage still exists at all

The prop values are rendered markdown — they must go through the same mdsvex pipeline (remark + custom hljs highlighter + KaTeX) as the rest of the page, and that renderer lives in JS. Likewise the rendered-<ul> parsing stays (now documented in a dedicated helper): parameter lists must be rendered as one markdown list to preserve exact tight/loose list semantics, then parsed back into per-parameter structured data for the tooltips/anchors.

Also fixes a latent bug

Python emits <yielddesc>, but the old regex matched <yieldesc> — yield descriptions have been silently dropped forever. The tag names now match, so docstrings with Yields: sections will (correctly) start showing their description. This doesn't affect accelerate (no rendered yields), but other libraries may gain previously-missing content.

Compatibility note

The tag format is an internal contract between the python package and kit/; the doc-build workflows always build both from the same repo checkout, so the change is atomic. Mixing an old PyPI hf-doc-builder with a newer kit checkout (or vice versa) would break docstring rendering — same caveat as any change to this format.

Verification

• Byte-identical output: full accelerate e2e build compared against the pre-change build — 48/48 pages byte-identical, normalizing only hashed asset paths, the kit build timestamp, and nondeterministic python <object object at 0x…> addresses. (accelerate is docstring-dense: 15 package_reference pages.)
• pytest tests/ passing (expectations updated for the new format; the one test_style_doc failure is pre-existing on main in my env and unrelated).
• eslint/prettier clean, including fixing the no-shadow violations the old file carried.

Second commit: a one-hunk svelte.config.js change (state_referenced_locally warning suppression) that was verified as part of #795 but accidentally left out of that PR (git add kit/src missed the config at kit/ root).

🤖 Generated with Claude Code