The link modal stacks a search field directly above a dropdown. They were
built independently and had drifted into two different treatments:

  DesignSystemSearch     surface.enabled (translucent overlay)
                         decorative.level01 hairline, radius l
  DesignSystemDropdown   background.level01 (opaque)
                         decorative.level02 at 2px, radius xl

Stacked four pixels apart, the pair read as controls from two different
products. The opaque fill is the load-bearing half: on an elevated surface —
every modal this control appears in — background.level01 is a dark sunken
hole in dark theme, while the search field beside it sits a touch lighter
than its host. Two fields, opposite reactions to the same sheet.

The dropdown moves to the search field's shell. It has two real call sites
(the relationship selector and the AI settings header bar), against the
search field's sixteen, and the search field's treatment is the better
reasoned of the two.

Two differences are kept deliberately. Expanded, the dropdown border still
switches to interactive.enabled — the search field has no focus treatment to
match, and losing the open-state signal to win a cosmetic match would be a
bad trade. And the field stays taller, because it stacks a label above its
value where a small search field holds one line.

Nothing enforced the match, which is why it drifted. A new test renders both
in one tree and asserts fill, border and radius agree — against each other
and against the tokens, so a change moving both in the same wrong direction
still fails. Verified it bites by reverting the fill.

Part of lotti3-6fz (item 3).

Review on #3583 found the change fixed one of the two production pairings and
described the other backwards.

The AI settings header stacks its dropdown under an AiSettingsSearchBar whose
default maps to DesignSystemSearchSize.medium — radius m — while the change
pinned every dropdown at radius l. That pairing stayed mismatched, and the new
test could not see it because it only covered the small variant.

DesignSystemDropdownSize now mirrors the search sizes and selects the radius:
small -> l, medium -> m, defaulting to medium since the field is already the
height of a medium search field. It controls radius only; height stays
content-driven because the field stacks a label above its value. The
relationship selector asks for small, pairing it with the task picker's small
search field.

The pairing test runs both real pairings across both themes, and now pins the
radius and border width against the tokens as well as against the search
shell — comparing only the two components left them free to drift off the
documented contract together, which is the drift the guard exists to prevent.

Two documentation corrections:

  * the link modal builds the relationship dropdown *above* the search field,
    not below it. The README and the CHANGELOG entry both had it reversed.
  * surface.enabled is theme-relative — white at 6% in dark, black at 6% in
    light — so the field lifts off a dark host and insets slightly into a
    light one. "Sits lighter than whatever panel it's on" was only true in
    dark theme, and had been copied unqualified into the CHANGELOG and the
    Flatpak metainfo.