fix(gui): pair the LAN switches, anchor the search row, and model pack order - #143

Two layout complaints from using the program, and the harness fix that lets the suite check one of them.

The two LAN switches share a row

They are one decision seen from two sides. Stacked, each had half a card of empty space beside it
and the second read as an afterthought under the first rather than as its mirror.

A checkbox takes a whole row by kind - a long label clips in half a card - so columns=2 on the
section did nothing by itself. Field.span is tri-state now: None decides by kind, True claims a
row, False pairs the field with its neighbour. Only three BOOL fields exist in the whole registry
and narrow_filter keeps its own row, so the Settings window is untouched.

The render guard measured buttons only, which is how a clipped checkbox label could have shipped -
and this change puts the two longest ones side by side. It measures checkbuttons now. Verified at
1366x768 in both languages with the widget count printed, so the "OK" is not vacuous: four checkboxes
per language, none clipped, the Polish pair asking 249 px and 195 px and getting both.

The search row is anchored at both ends

Third shape, and the first two failed the same way. Everything packed LEFT was reported as "a stray
label floating above the first section with the whole width empty beside it"; everything packed RIGHT
drew the same report mirrored, and an alignment fix in between did not settle it either.

So the side was never the problem: a row with ONE cluster in it has a band of nothing beside that
cluster wherever the cluster goes. Both ends are anchored now - label and box at the left margin, note
and count at the right - which is the shape the Connections toolbar already uses and nobody has
complained about. Measured: the label lands on the section cards' own left edge, 12 px further left
than it sat before, and the count ends a pixel from their right edge. On the Connections page that
label is flush with its table, so the two pages now agree about where a search box lives.

The idle state is the one the page is looked at in, so the right end carries Ctrl+F while the box is
empty - otherwise the row drops straight back into the rejected shape whenever nobody is searching.
Same label, same fixed width, so nothing moves when it turns into a position counter.

The fake tkinter models pack order

It answered pack_slaves in CREATION order while its docstring claimed pack order. Measured: pack A,
then pack B with before=a, and it still answered [A, B].

That is not cosmetic. A widget re-packed above an existing sibling looked right in the suite and landed
under the whole page on real Tk - exactly what the Control page does when a preference brings its search
bar back, which is why that order could until now only be checked by rendering. The parent keeps an
ordered list; pack appends, honours before=/after= and MOVES an already-packed widget;
pack_forget removes it. The test that had to settle for "the call carried before=" now asserts the
bar is back above the page body.

Written against __dict__ rather than getattr(w, "x", None): __getattr__ answers any unknown
attribute with a no-op callable, so the default never arrives and the "not set yet" branch is
unreachable. That bit the first draft of this fix.

Verification

293 guards covering everything touched, ruff, mypy, the GUI smoke and the real-Tk render check in
both languages. Four new entries in the mutation registry, all caught - plus one repaired: an older
entry quoted span=True on lan_mode and the bookkeeping check caught it going stale, which is what
that check is for.

Two existing tests were asserting "cleared = blank" as a proxy for "no position is shown" and now say
which idle text they mean.

Looked at on the live window as well, since two of the three are about how the page reads.