[HosseinKarami]

Summary

Version 1.3.6. Fixes from the sample check of a person. The check was on the seventh full audit of the large WordPress multisite, with 1.3.4, and on its first review cycle. The person answered 50 questions and checked 20 issues. On the 37 questions where the two AI passes agreed and the person gave a clear answer, the AI was right on 36. All 14 issues in scope had the right origin, source, and solution.

The details are in CHANGELOG.md.

The tests

• The image-of-text test reads the pixels of the image itself, at its size on the page, over the background color of its container. Before, it read an element screenshot. During an element screenshot the window changes its size, and on the multisite a script of the page moved the page. So the screenshot showed the heading above the image, which is text on a flat ground. On the multisite, 2,223 of the 3,711 questions that waited for the AI were images of text. 53 of their 58 answers were "no issue". An image from another site is read from its file, one time.
• The test also counts the ink: the pixels far from the main color. Small letters on a large ground have no second color of 2%.
• The test waits until a lazy image shows its file, not a placeholder from a data address.
• The list question reports the innermost container. Its evidence gives the line before the list and the lines that start with a number or a bullet. Before, it gave the first 5 lines of the container, and the AI missed a list of five points low in an article. The screenshot shows the lines.
• A label that only a screen reader reads is not a visible label: 1 or 2 px, clipped, or off the screen. The question names it.
• The player question names the alternatives, and says that the title of the frame does not count.
• The detection of image-of-text, list-without-list-code, and field-no-visible-label is now 1.3.6.

The results from run to run

• A page that the settings leave out now, by a skip pattern or a host that is not permitted, ends its earlier results. Its issues are "page removed" with the new reason "out of scope", its questions go, and the audit sends no request to it. Before, its results stayed "not checked" with no end. The report gives the counts.

The cycle

• review.md: the hidden label, the title of the frame, and the lines of a list.

Test plan

• npm test, with a new test file, images-lists-scope (21 checks).
• The new checks on the 1.3.5 code: 14 fail.
• The seeded faults: 74 of 74 found.
• The new image test on 30 pages of the multisite, read-only, at 390 px: the 5 images of text that the AI confirmed are found. Of the other 25 questions, 23 go. The 2 that stay are flat pictures with words.
• A full audit of the multisite with 1.3.6.