v0.0.173

A page the parser cannot read no longer reports missing titles and headings, content/mojibake and links/anchor-mismatch handle more real text correctly, ax/ai-crawlers reports search-engine bots, the macOS Intel binary runs on older CPUs, and the REST API docs cover rate limits and unknown request fields.

Added

• The REST API docs and the OpenAPI file state that POST /v1/agent-runs from an API key answers an unknown top-level field with a 400 VALIDATION_ERROR naming it, so a typo such as conifg fails instead of running a default audit, and list the accepted top-level fields.
• ax/ai-crawlers reports Googlebot, Bingbot and Applebot with a path-aware verdict. A new info check, search-engine-access, evaluates each search-engine bot against the audited URL with the shared robots.txt matcher and says whether it is allowed, blocked from the site root, blocked for that path only, or blocked at the root but re-permitted for that path by an Allow rule. Applebot with no group of its own follows the Googlebot group, as Apple documents. The 24 AI tokens and their checks are unchanged.
• The REST API docs state the per-key request limit: 60 requests per minute on Free, 300 on Starter, Team and Enterprise, for POST /v1/agent-runs and POST /v1/reports (the read endpoints are not metered), with the RateLimit-* and Retry-After headers and the plain string 429 body. The OpenAPI file no longer lists the limiter's 429 on the read endpoints and shows the real body.

Fixed

• A page whose markup could not be parsed at all (read as an empty document, with the html-parse warning on crawl/html-size) no longer reports what it was never seen to have. The page rules that read the markup (title, H1, meta description, lang, viewport, thin content, no <main> and the rest) are skipped as not evaluated with the reason unread-page: it is listed under skippedChecks, and a prior finding on that page carries instead of resolving. The rules that judge the URL, the response (headers, timing, size, MIME type), or the raw HTML still run: a11y/focus-visible, analytics/consent-mode, content/mime-type, core/doctype, crawl/html-size, integrity/kit-signature, perf/browser-required, perf/cache-headers, perf/compression, perf/http2, perf/ttfb, security/cookie-flags, security/https and the url/* rules. The cloud analysis rules (ai/llm-parsability, eeat/authority-signals) are skipped too, and the page is not sent to the cloud, so it costs no credits. The page is left out of the site-level data (no page_features row, no collectors), so the site rules that compare pages (word-count norm, asset divergence, NAP consistency, template discontinuity, duplicate titles) never count it, and it is not the template representative whose verdicts are copied to its cluster. Known gaps: the page's outgoing links are not read, so links/orphan-pages can flag a page only it linked to, and the site-metadata payload still includes the page. No input known today reaches this path (a class attribute is cut instead), so this is a safeguard (#1570)
• The macOS Intel (darwin-x64) binary is now built with Bun's baseline target, like the other x64 builds, so it no longer dies with an illegal instruction (exit 132) on Intel Macs whose CPU lacks AVX instructions #1562
• content/mojibake reports the Portuguese and Spanish lowercase accents. ã, á, ó, ú, Ã plus a soft hyphen, â, ê and õ in place of ã, á, ó, ú, í, â, ê and õ are detected, while correctly encoded text and uppercase words such as MANHÃ stay clean. (#1623)
• links/anchor-mismatch no longer warns on a German compound head noun or a generic listing label. Handschuhe linking to a page titled Fahrradhandschuhe shares a word, and a label made only of listing words (Alle Artikel, View all posts) pointing at an article or blog listing is counted as not judged. An unrelated anchor still warns. (#1626)