Evidence — icon-only language trigger and a compact, positive menu (#406, PR #431)

Before/after for T-038: the navbar trigger loses its EN label and caret, and the menu stops rendering "not available" as a wall of disabled, greyed-out rows.

Before / after

	Before	After
Trigger size	88 × 19 px	38 × 38 px
Trigger visible text	EN	(none)
Caret (dropdown-toggle)	present, ::after inline-block	absent, ::after: none
Trigger accessible name	title="Language"	title="Language: English"
Menu width	243 px	144 px
Widest row	241 px	142 px
Disabled / aria-disabled rows	2	0
Bootstrap .active primary fill	1	0
Untranslated rows that are real links	0	2
auto chips on machine translations	0	1
Per-row "Not yet translated" strings	2	1 (the single footnote)

The panel variant is captured too (01-before-panel-variant.png, 02-after-panel-variant.png) — the include renders two markups and this change touches both.

Against the issue's acceptance criteria

• Menu shows no disabled rows — 2 → 0. Español and Deutsch are now <a> elements to the source page, so the existing zer0-lang preference handler (which matches a[data-lang]) fires on the fallback path too; it could never fire on the old span.disabled.
• Widest row ≤ 220px — 241px → 142px, measured with getBoundingClientRect(). PR #431 originally shipped this criterion as a static assertion on the CSS rule with an explicit caveat that nothing had measured it in a browser. This closes that gap: the old menu genuinely exceeded the cap, the new one is 78px inside it.
• Current language visually distinct without a primary fill — Bootstrap's .dropdown-item.active count is 1 → 0, replaced by .is-current (tint + bi-check2) while aria-current="true" is retained.
• Trigger is icon-only at all widths with an accessible name — the EN span and the caret are gone, the button is a 38px square, and aria-label / title are the only accessible name — so title now names the language being read rather than just "Language".
• One footnote instead of per-row text — 2 per-row strings → 1 footnote under a divider, referenced by each marked row via aria-describedby.

How this was measured — and the caveat

No Jekyll site was served. The agent sandbox has no bundler and could not install one, so bundle exec jekyll build and the usual evidence-kit.mjs run against localhost:4000 were both impossible.

Instead the include is rendered with liquidjs against a fixture and styled by the real assets/css/main.scss compiled with Dart Sass plus the repo's vendored Bootstrap CSS and JS, then measured in real Chromium via Playwright. "Before" is origin/main, "after" is this branch — same harness, same browser, same fixture, same viewport.

Two honest limits:

1. The fixture declares three target languages (fr translated, es/de not) where _config.yml ships only [fr]. That is deliberate — with a single, translated target none of the untranslated / chip / footnote states would render at all — but the absolute row counts above are the fixture's, not production's.
2. These are component crops, not full-page screenshots of the built site, so they do not show the toggle in the real navbar cluster at every viewport. test/visual/features/language-toggle.spec.js runs against the real built site in Test Suite (Ruby 3.3) and remains the authoritative check.