which-heading-do-i-need: nested sections are not heading context; heading-less boundaries are transparent

[NullVoxPopuli-ai-agent]

Fixes two ways getSectionHeadingLevel (which-heading-do-i-need) disagreed with its own documented algorithm — "We'll need to check the subtrees between these elements, stopping if we encounter other boundary elements" — surfaced while fixing an axe heading-order failure via universal-ember/kolay#318 (see the review discussion there).

1. Fallback traversal recursed into sectioning content

findHeadingIn's previous-sibling scan skips boundary siblings, but its fallback child-traversal recursed into every child, including nested <section>s. Headings inside a nested/sibling section belong to that section and are not context for ours, but the fallback found them anyway. Consequences:

• sibling headings behind non-boundary wrappers resolved one level deeper than the first heading, instead of equal;
• sibling <section>s cascaded one level per section (h3, h4, h5, ...).

The fallback now skips sectioning content (section, article, aside, nav) — consistent with the sibling scan. header / footer / main are still traversed: a <header> in particular typically contains its section's heading, which the existing "wrapped headings" rendering tests cover (a blanket boundary skip fails 4 of them; this split keeps them green).

2. Heading-less boundaries isolated their contents to <h1>

levelOf returned 1 as soon as the upward walk reached the next boundary, even when that boundary had no heading of its own. A heading-less wrapper <section> therefore cut everything inside it off from the document — kolay's api-docs signatures rendered <h1>s mid-page (a 2+ level heading-order jump on consuming docs pages, flagged by axe-core ≥ 4.12). A boundary without a heading has no level to nest under, so it is now transparent: the walk keeps searching outward for context.

<h2>API Reference</h2>
<section>          <!-- no heading of its own -->
  <section>
    <Heading />    <!-- was: h1; now: h3 -->
  </section>
</section>

Verification

• 2 new unit tests (heading-less sections are transparent; sibling-section headings are not context)
• full test-app suite: 388 pass / 0 heading failures — including all existing wrapped headings, extraneous sections, shadow-root, and mixed section/aside/article cases (4 <InViewport /> failures in my local run are pre-existing: identical on an unmodified checkout, IntersectionObserver-timing related, untouched by this change)
• package lint: eslint, tsc, prettier all green
• downstream check: with this fix packed into kolay#318 and that packed into nvp.ui (floating deps, axe-core 4.12.1), all api-docs signature headings resolve as equal-level siblings one below the heading that precedes the block (e.g. h2 API Reference → h3 Element / h3 Arguments / h3 Blocks), union variants no longer cascade, and nvp.ui's full suite passes 83/83

🤖 Generated with Claude Code

[bolt-new-by-stackblitz]

Run & review this pull request in StackBlitz Codeflow.

[github-actions]

Project	Preview URL
Docs	https://be23b687.ember-primitives.pages.dev

Logs

[NullVoxPopuli-ai-agent]

Re: the Docs' Tests timeout on the first run (86843900699) — investigated as a possible traversal hang; it isn't one:

• The job ran exactly one test (ok 1 - Docs: Scroller at 14:30:40), then produced nothing for 14 minutes until the 15-minute job cap killed it, with Chrome still alive at cleanup — a wedged browser, not a slow creep (main's Docs' Tests baseline is ~2.5m).
• Could not reproduce: the full docs suite passes locally in ~1.5m total, with the all-pages a11y test at 42.4s — and 43.5s when pinned to 2 CPUs (taskset -c 0,1) to match the CI runner class.
• Termination argument for the changed walk: levelOf only ever moves current strictly upward (parentNode, or shadow-root → host), and the stopAt extension for heading-less boundaries comes from nearestAncestor, i.e. is always a strict ancestor — no revisiting is possible; findHeadingIn is structural recursion over children/previousSibling on a finite tree. The extension also doesn't change asymptotics meaningfully: the pages test went 42s → 43s vs the unmodified baseline locally.

The retriggered run confirms: Docs' Tests passed in 2m15s, and every other job is green.