feat(gui): find a setting on the Control page by name

[donislawdev]

The Control page renders the whole field registry - a dozen sections, forty-odd fields, several of them folded away on a fresh install - and there was no way to reach a setting except by opening every section and reading. This adds a search box for it.

What it does

Type part of a field or section name, or a flag such as --loss. Every match is highlighted, the one you are on is filled in, and the counter in the corner points at it. Enter walks forward, Shift+Enter back, Escape clears, Ctrl+F puts the caret in the box.

• It marks and jumps, it does not filter. Hiding what does not match would fight the page's own layout: fields are packed left to right in rows, so a hidden field leaves a hole, and sections are spread over two columns by weights computed at build time, so sections disappearing rebalances the page under the reader.
• A folded section opens for the search and folds back afterwards. Through set_open, never toggle - toggling runs the accordion's callback, which persists the fold state, so searching would permanently unfold sections the user had chosen to keep closed.
• Names and CLI flags, not tooltip bodies. Searching the tips would make one word light up half the page, which is a symptom index rather than a way to find a field.
• Accents are optional. Polish labels carry diacritics and people type without them, so matching folds them away through NFKD. Without this "opoznienie" finds nothing while the accented spelling works, and no user can tell which half of the feature is broken.
• A field that renders in the Settings window is named, not missed. It is indexed but never offered as a jump, because this page cannot scroll to it.

Structure

gui/form_search.py holds the matching: registry plus the current translation in, entries in page order out. It has no Tk, so it is tested directly instead of through the fake-Tk subprocess. The page owns the highlighting, the scrolling and the fold rules.

The index is built after the language is set rather than at import time - a language switch rebuilds the UI precisely so nothing keeps the old words. The query survives that rebuild (and the two-column one) and dies with the process; ControlForm gained on_rebuilt so the marks can be put back on widgets that no longer exist.

Ctrl+F now goes through one dispatcher. Both pages with a search box bind it on the root, and a root binding without add= REPLACES the one before it, so the page built second would otherwise have taken the shortcut away from the other. From a page with no box it keeps the older behaviour and brings the connection table forward.

What the live pass found

The fake tkinter has one widget class for everything and never validates a style name, so the whole feature was also driven on real Tk. That is where the last defect came from: two hits can claim the SAME widget - a dropdown has no text of its own so it marks its section header, and that header is often a hit itself. It was painted as the current match and then repainted as an ordinary one by the later hit, so "1 / 5" appeared with nothing filled anywhere. The second claim on a widget is dropped now, which is what keeps the count honest: as many marks as it says, each in its own place.

Guarded

• 8 tests in tests/test_form_search.py (pure): the index covers every registry entry, blank and punctuation-only queries find nothing, case and substring matching, section names, CLI flags in three spellings, page order, the Settings-window split, and an accented label reachable without its accents.
• 12 in tests/test_gui_layout.py: marking and clearing, current versus other and its wrap-around, the highlight not reusing the page accent, the section-header fallback, both fold rules, the "it lives in the Settings window" answer, the empty result, the widget collision, surviving a form rebuild, one Ctrl+F reaching whichever box is in front, and the Settings window - which renders the same ControlForm - having no search box.
• 4 new entries in the mutation registry, all caught. One existing entry was replaced: it used to break the table's Ctrl+F by moving the binding off the root, and with two pages binding the same dispatcher that no longer breaks anything, so it now breaks the dispatcher's decision instead.

Full suite 1156 passed, smoke_gui.py OK, mutation registry 106 caught with the canary BROKEN, both real-driver smokes clean.