UI rework polish - #2666

Two behaviours the page offered twice, both visible on the same screen.

Row hover is opt-in through `clickable-table-row`, and the rule that suppresses
it everywhere else was written as `tr:hover > td`. The actors and workers tables
are antd `virtual` tables, whose body is divs — no `tr`, no `td` — so those two
went on lighting up under the pointer while the pools and templates beside them,
and every other static table in the app, had stopped. The rule now matches the
cell class antd actually paints the hover from, which exists in both bodies; the
pointer and pressed rules for clickable rows are rewritten the same way so the
mirror bug cannot appear. Selected rows keep their own hover.

The actor and worker columns carried a header of the page's own: a button around
the title, a bare arrow beside it, nothing outside those few words to click. It
avoided antd's `sorter` because a comparator reorders the page in hand, and one
page out of 410,110 reordered is not the cluster sorted. The concern was right
and the remedy left a reader two headers to learn. `sorter: true` is antd's
header with no comparator behind it, so the whole cell is the target, the
chevrons show the direction, and the table still reorders nothing — the order
goes out with the next read.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Signed-off-by: Torey Scheer <[email protected]>

…e rest

Eight review comments on kagent-dev#2666, the first of which is that the substrate sort
was never server-side. `GetSubstrateStatus` takes a namespace and nothing
else, and `localPage` orders the whole inventory in the browser before
slicing out a page — so three places asserted a sort that does not happen,
one of them text a reader sees: "Sorted by the server: actor, ascending".

`sorter: true` stays, because it is right for the opposite reason to the one
written down. The read has already ordered every row, so a comparator would
re-sort the hundred on screen. The strip beside each table now claims what is
true and worth claiming — the order holds across the whole inventory rather
than over the page in front of the reader — and `ServerOrder` is
`AppliedOrder`, since what it reports is what was applied and not who applied
it.

Carrying the order in the read key costs a full re-read of the inventory per
header click, to reorder rows the browser was just holding. That is now
written where the key is built instead of being left to be discovered.
Removing it means moving the sort and the slice out of the transport so a
reorder can be a re-render, which changes what the operation returns; it
belongs in its own change rather than in a header fix.

Each column's key is now typed as its sort field, so a key that is not one
fails to compile — the shape that let `pod` stand where `workerPod` belonged.
Actor `status` and worker `pool` resolve to the same ordering as `default`
because they *are* it; named rather than left to fall through, so a change to
the default has to decide what those headers mean.

Four smaller ones. The template form's name was marked required through a
condition that could not be false — it renders only while creating, and the
creating caller is never read-only — so it is unconditional and the
conditional rationale moves to the model configuration, which is the field it
describes. The hover rule's selected-row exception could never match: a row
is selectable only where it is also clickable, so this rule has already
passed it by. The expand-icon rule still keyed off `tr.clickable-table-row`,
the stale selector the rules above it stopped using. And the no-hover test
asserted a colour after a fixed wait, which passes just as well when the
pointer never arrived — it now asserts antd marked the cell hovered first, so
a missed hover fails instead of proving nothing.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Signed-off-by: Torey Scheer <[email protected]>

The Remote URL / Command toggle was reported as low contrast in dark mode, and
its text was not the problem: the labels measure 12.6:1 selected and 8.2:1
unselected. The control was. Its track measured 1.00:1 against the page — the
same colour — and its selected pill 1.12:1 against the track, behind an
elevation shadow antd paints at one percent white. Both halves were perfectly
legible and nothing said which one was chosen, or that the thing was a toggle.

The light theme is left alone, because its pill is white with a dark shadow
under it and that reads.

Fixed with an edge rather than a fill, the way the inputs are. A fill cannot
carry it at this end of the range: the page is near-black, so 3:1 on luminance
alone takes a mid-grey pill that reads as disabled. A one-pixel `borderStrong`
ring measures 3.5:1 against the track and 3.9:1 against the page, and is the
boundary every other control in the app is already outlined with, which leaves
the two fills with only the pill's shape to carry. Both toggles on the MCP form
are the only segmented controls in the app, and the fix is a theme token, so
neither page changes.

The spec gains a block of its own for non-text contrast, which is a separate
threshold that a check aimed at text walks straight past: 3:1 for the edge that
identifies the state, with the labels measured beside it so that edge cannot be
won by dimming them. Verified in both directions — without the theme change it
reports 1.01:1 and fails.

Its shadow colours are composited over what they sit on before being measured.
Read as opaque, a one percent white shadow measures as the most visible edge on
the page instead of the invisible one it is, which is the mistake this spec's own
note already records against page colours.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Signed-off-by: Torey Scheer <[email protected]>