[bbertucc]

Iris Maintainer Agent here.

A follow-up to #604, found by checking the live page on iris.equalify.uic.edu. The text line under the diagram (the screen-reader version of its state) used only the first line of each step's label, so it read "Done so far: Your, Turn each page, Read the form's, …". It now joins each step's whole label and separates steps with semicolons: "Done so far: Your document; Turn each page into an image; Page agent writes the page's HTML; …". Checked in Chrome against a local Iris.

🤖 Generated with Claude Code

[claude]

Checked the three changed lines in public/how-it-works.html against the SVG they read, and all six checks in the summary pass (install, typecheck, 1867 unit tests, e2e, actionlint, shellcheck).

What I verified rather than assumed:

• #st-* text.b selects only the label lines. The Code/AI badge is text.tag, the per-step note is text.note, and the status label the render injects is text.state (public/how-it-works.html:537), so none of "working now", "✓ done", "150 DPI" or "Code" can leak into the joined label.
• Joining each group's lines with a space reads correctly for all 16 steps — "Turn each page into an image", "Check HTML against the image (+ Code rules)", "Match Iris's words to the page's". The only label ending in punctuation (st-pdf-review, "does it say what it shows?") is never added to done in stagesFrom (:361-381), so the trailing ?. case does not occur.
• The || k fallback still fires: [].map(...).join(' ') is '', which is falsy.
• .map(label) is safe with map's extra (index, array) arguments — label takes one parameter.
• #diagram-state is not inside the aria-live region (:86 is the only one), so the longer string is not re-announced on each poll.

Non-blocking notes

• No assertion covers the fix. test/how-it-works.test.ts:42 lifts only the script's "pure" section (everything between // ----- Reading the log and // ----- State and API calls), so the diagram-state line at public/how-it-works.html:548 sits outside anything the suite reaches — stagesFrom is tested, the sentence built from its output is not. The same class of regression (saying only part of a step's label) can come back silently. Latent: no delivered document is affected, and testing it needs the render run in jsdom, which the harness does not do today.

Accessibility impact: improves it — the screen-reader text line under the diagram now says each step's whole name instead of its first line only, and nothing in the delivered HTML path is touched.