fix(ui): drop the stylesheet selectors that can never match an element

[argszero]

Summary

ui/css/style.css carried four rules whose names no longer appear on any element — not in the markup, not in the scripts, and not in the design baseline the stylesheet itself declares (ui/css/style.css:4 → docs/prototype/aitokenpool-console.html):

selector	status
.demo-hint (:754)	element removed with the demo-seed cleanup (#90 35291b3); the rule stayed, and its i18n key was already retired
.login-divider + ::before + ::after (:757/:759)	#159 f46882c moved the markup to the prototype's own name, .divider; the old rules kept a "historic alias" comment and nothing else
.wallet-note (:909)	#156 41bb6d1 rewrote the element as .wallet-hero-note; the old rule was left behind
.sidebar #toggle-mode-btn (:1242, inside the mobile media query)	the button it hides never reached the product; the #logout-btn beside it in the same selector list is live

A rule nobody carries still applies — writing class="wallet-note" today still gets its 12px — which is what makes these misleading: they read as live component rules for elements that no longer exist. Each of the four was introduced as a rename or a removal elsewhere, and the rule was left in place.

.delta (three rules) is deliberately kept: the prototype's statCard() really does emit <div class="delta ok">, so it is design-baseline vocabulary, not a leftover.

Related Issue

None — this came out of a repo-internal audit of selector/element correspondence, and there is no open issue for it.

Changes

• Delete the four orphan rules; the one inside the media query is removed from the selector list only (its sibling #logout-btn stays). No behaviour change: nothing renders these names.
• src/layout_gate.rs: new rule R4 every_selector_lands_on_an_element — every .class / #id in a selector prelude must occur on an identifier boundary in the app corpus (ui/index.html + ui/js/*.js) or the design-baseline corpus. Token roster derived from the selector text; script roster borrowed from js_gate so a new ui/js/*.js cannot slip past it; ⛔ no hand-written name list.
• ui/README.md: convention section stating the invariant, why a still-applying orphan rule is worse than a deleted one, and the gate's scope.
• ui/index.html: cache-bust token style.css?v=20260924-4 → -5.
• No config / data-structure change.

Gate scope (lexical, deliberately wide): it proves a name has a landing site, not that the landing site is the element the rule was written for (a same-name occurrence in a comment, a string, or dead code counts; mentioning a name ≠ wearing it). The baseline corpus is an intentional escape hatch — anything the prototype mentions passes.

Tests

• cargo test — 446 passed, 0 failed (baseline b59999b: 442) — the gate brings 4 new tests.
• cargo fmt --check — clean.
• cargo clippy --all-targets -- -D warnings — clean.
• Unit tests added/updated: every_selector_lands_on_an_element plus four companions — a self-proof that the scanners see something and do not read comments/strings, a nested-rule + two-corpus negative control (.delta clears only via the prototype; #toggle-mode-btn is unseen by both), a per-leg mutation test, and a leg-reading printer.
• A/B on the real tree: re-adding .wallet-note / .demo-hint to style.css makes every_selector_lands_on_an_element fail (restored byte-identical afterwards). Run against the pre-fix tree it fails naming exactly these four.

Checklist

• Branch name follows convention (fix/…)
• Conventional Commits format, no (#N) in the title
• Single responsibility, minimal change