Status: PLAN / awaiting sign-off (written 2026-07-11). Nothing implemented yet.

Replaces the current full-screen name editor (screens/modal/FullKeyboard.qml → PagedPicker → SearchRoot/SearchContainer/SearchCriteria/ SearchResults + TilesetPicker) with a real keyboard outline: an ASDF/QWERTY deck whose letter and number keys each hold one game tile, with the physical key it's bound to shown as a superscript legend in the corner. Click the key, or just type it. Modifier combinations (Shift / Ctrl / Alt) switch pages; the pages are also clickable buttons above the deck.

1. The numbers (why the layout is what it is)

projects/db/assets/data/font.json holds 255 tiles (ind 1–255; 0 is the string terminator and is not a tile). Classified by the app's existing colour precedence (normal → control → picture → singleChar → variable → multiChar, from SearchResults.determineColor):

Category	Count	What it is
Normal	70	A–Z, a–z, space, ( ) : ; [ ] - ? ! . / ,, Pk Mn ♂ ♀ ×
Single-Char	49	digits 0–9, bold A–I/V/S/L/M, é, ', contractions ('d 'l 'm 'r 's 't 'v), quotes, …, ·, $, kana
Multi-Char	7	<pkmn> <poke> <pc> <tm> <trainer> <rocket> <......>
Variable	4	<player> <rival> <targ> <user>
Picture	114	tile01–tile48, tile4D, box-frames ╔ ═ ╗ ║ ╚ ╝, tileC0–tileDF, arrows ⮚ ⯈ ⯆
Control	11	<page> <line> <next> <para> <end> <cont> <cont_> <autocont> <done> <prompt> <dex>
Total	255

Assignable keys per page: letters (26) + number row (10) = 36. Pages needed: ⌈255 / 36⌉ = 8. Modifier combinations available: Shift / Ctrl / Alt = 2³ = 8. They line up exactly — one page per combination, no leftovers, nothing arbitrary.

The Space tile ( , ind 127) is the 255th and rides on the deck's real spacebar, which is where every human already expects it. So: 254 tiles across 8 × 36 alphanumeric keys (34 slots left empty), + Space on the spacebar.

2. Page order & the mapping doctrine

The category strip's order (minus "All") is the usage/priority order Twilight named: Normal → Single-Char → Multi-Char → Variable → Picture → Control. Pages are ordered by that, and cheap modifier combos get the high-usage pages.

But keys are not filled linearly by category. Three rules, in priority order:

Identity beats everything. If a tile is a key (A, a, 7, !), it lives on that key. A user should never have to learn where "A" is. This is why page 1 is uppercase A–Z + digits and page 2 is lowercase a–z — exactly like a real keyboard.
Mnemonic beats ergonomics. 's sits on S, 't on T, bold B on B, <player> on P, <pc> on C, <dex> on X. The superscript legend means an awkward-but-memorable key beats a comfortable-but-random one.
Ergonomics breaks ties. Where there's no identity and no mnemonic, the more-used tile gets the better key — home row (F J D K S L A G H) first, then the upper row, then the bottom row, then the number row (a reach). Frequency here is judged by category order first, then by how often the tile actually shows up in names/dialogue.

A tile is promoted to a better key on a later page rather than being buried on a bad key of an earlier page (Twilight's explicit instruction) — e.g. the box-frame glyphs and cursor arrows are the only genuinely useful Pictures, so they take page 5's home row rather than being dumped in tile order.

The pages

#	Modifier	Name	Contents
1	—	Letters	A–Z uppercase (Normal) + digits 0–9
2	Shift	Lowercase	a–z (Normal) + the 10 Normal symbols
3	Ctrl	Symbols	bold letters + é ' “ ” ‘ ’ … · $ ﹕ . + : ; [ ] / Pk Mn
4	Alt	Codes	contractions + <player>/<rival>/<targ>/<user> + Multi-Char words + kana
5	Shift+Ctrl	Tiles I	box frames + arrows + tile01–tile1B
6	Shift+Alt	Tiles II	tile1C–tile3F
7	Ctrl+Alt	Tiles III	tile40–tile48, tile4D, tileC0–tileD9
8	Shift+Ctrl+Alt	Controls	tileDA–tileDF + the 11 control codes

Pages 1–4 (the four cheapest combos) hold every Normal, Single-Char, Multi-Char and Variable tile — i.e. everything you'd ever put in a name. Pictures and Controls (the two you shouldn't) sit behind two- and three-key combos.

3. The key maps

Deck rows: A = number row 1 2 3 4 5 6 7 8 9 0, B = Q W E R T Y U I O P, C = A S D F G H J K L, D = Z X C V B N M.

Page 1 — Letters (no modifier)

Key	Tile	Key	Tile
1–0	1 2 3 4 5 6 7 8 9 0 (Single-Char)	Q–P	q w e r t y u i o p
A–L	a s d f g h j k l	Z–M	z x c v b n m
Space	(Normal)

Pure identity: the base layer of a real keyboard — press a letter key, get the lowercase letter. 36/36 used.

Corrected 2026-07-11 (Twilight). The first cut had uppercase here, reasoning that Gen 1 names are all-caps. That was wrong: it's the one place the deck would have contradicted every keyboard its user has ever touched. All-caps is what Caps Lock is for (§5).

Page 2 — Uppercase (Shift)

Letters: Q W E R T Y U I O P / A S D F G H J K L / Z X C V B N M — identity, exactly like a real keyboard. Number row = the 10 Normal symbols, matching the real Shift-row where the game has the glyph:

Key	Tile	Why
1	!	Shift+1 = ! on a real keyboard ✔
2	?	the other sentence-ender
3	.
4	,
5	-
6	♂ (<m>)
7	♀ (<f>)
8	× (<x>)	Shift+8 = * = multiply ✔
9	(	Shift+9 = ( ✔
0	)	Shift+0 = ) ✔

36/36. Page 2 is 100 % Normal — pages 1+2 together are the whole in-game-legal character set.

Page 3 — Symbols (Ctrl)

Key	Tile	Key	Tile	Key	Tile
A	bold A	I	bold I	Q	“ (<o">)
B	bold B	L	bold L	W	” (<c">)
C	bold C	M	bold M	R	‘ (<o'>)
D	bold D	S	bold S	T	’ (<c'>)
E	bold E	V	bold V	U	… (<...>)
F	bold F	J	' (apostrophe)	Y	· (<mdot>)
G	bold G	K	é (<e>)	Z	﹕ (<:>)
H	bold H	N	Mn (<mn>)	X	. (<.>)
		P	Pk (<pk>)	O	$ (Pokédollar)
1	:	2	;	3	[
4	]	5	/	6–0	(empty)

Ctrl+B = bold B is the same thing it means in every app on earth. 31/36 used.

Page 4 — Codes (Alt)

Key	Tile	Mnemonic
D L M R S T V	'd 'l 'm 'r 's 't 'v	the contraction sits on its own letter
P	<player>	Player
O	<rival>	Opponent (and it sits right next to Player)
Y	<targ> (Player Usage)	Your usage
U	<user> (Enemy Usage)	User code — next to Your
K	<pkmn>	PoKémon
E	<poke>	Poké
C	<pc>	Computer
J	<tm>	prime free key (T and M are taken by contractions)
H	<trainer>
G	<rocket>
F	<......>	prime key — …… is common in dialogue
1–5	あいうえお	kana strip
6–8	ァゥェ
A B I N Q W X Z, 9, 0	(empty)

26/36 used.

Page 5 — Tiles I (Shift+Ctrl)

The six box-frame glyphs are drawn as a box on the keys — Q=╔ W=═ E=╗, A=║, Z=╚ X=╝ — and the three cursor arrows take the right home keys: J=⯈ K=⯆ L=⮚. Those nine are the only Pictures anyone actually uses, so they get the best keys on the page.

The remaining 27 keys carry tile01–tile1B in reading order (number row → R T Y U I O P → S D F G H → C V B N M). 36/36.

Page 6 — Tiles II (Shift+Alt)

tile1C–tile3F, straight reading order: number row (1C–25), Q–P (26–2F), A–L (30–38), Z–M (39–3F). 36/36.

Page 7 — Tiles III (Ctrl+Alt)

Number row: tile40–tile48 + tile4D. Q–P: tileC0–tileC9. A–L: tileCA–tileD2. Z–M: tileD3–tileD9. 36/36.

Page 8 — Controls (Shift+Ctrl+Alt)

Number row 1–6: tileDA–tileDF. Control codes on mnemonic letters:

Key	Code	Key	Code
E	<end>	C	<cont>
L	<line>	V	<cont_> (alt continue, next to Cont)
N	<next>	B	<autocont>
P	<page>	D	<done>
A	<para>	R	<prompt>
X	<dex>

17/36 used. The most dangerous codes live behind the most deliberate key combination — you cannot hit <end> by accident.

Completeness

254 tiles on alphanumeric keys + Space on the spacebar = 255 = every tile, each exactly once, no duplicates. This is pinned by a unit test (§6).

4. Visual design

The current screen's problem is that it's a list of chips next to a filter form. The new one is a device: a keyboard deck you look at, not a form you read.

Deck. A rounded neutral chassis (light grey, subtle inner border) holding the key rows, centred in the page. Real-keyboard silhouette: number row + ⌫, QWERTYUIOP, ASDFGHJKL + ⏎, ⇧ ZXCVBNM ⇧, and a bottom row of Ctrl Alt ␣ Alt Ctrl.
Keycaps. Rounded rect, near-white face on the deck, 1 px border in its category colour, a faint wash of that colour in the face (≈8 %), the animated tile centred, and the physical key as a superscript legend in the top-right corner (muted grey). Hover → border goes full category colour, cap lifts (a soft shadow), and the detail pane fills. Press → the cap fills the category colour at ~20 % and depresses 1 px.
Structural keys (⇧ Ctrl Alt ␣ ⌫ ⏎) are neutral grey caps with no tile. The modifier keys light up in the accent colour whenever they're held or latched — so the deck itself shows you which page you're on.
Page strip — 8 buttons above the deck (the SegStrip pattern from the Pokemart screen), each showing the page name and its modifier badge (⇧, ^, ⌥). The active one is filled; holding a modifier previews/moves the selection live. Clicking a page latches it (so the whole screen is usable with the mouse alone, and on a touchscreen).
Left rail — the colour reference. Replaces the old category filter list. Six rows: colour swatch · category name · ⓘ. Hovering the ⓘ shows the category's explanation (the same copy the old HelpDots carried). It's a legend, not a control — nothing to click, nothing to get wrong.
Right pane — the detail split (kept, with its divider): hovering or clicking a key fills it with the colour bar, the tile's name/alias, its raw code, and its description.
Footer unchanged: the Name / Example toggle, the re-roll button, and the live NameDisplay.
Header keeps the "Simulated" group (Outdoor toggle + tileset combo) and the name field. The Grid/Tileset toggle is deleted and the raw tilemap view goes with it.

Category colours — unchanged (decided 2026-07-11)

The six Settings::fontColor* values stay exactly as they are (Normal grey #616161, Single-Char lime #9E9D24, Multi-Char amber #FF6F00, Variable green #388E3C, Picture blue #303F9F, Control purple #7B1FA2). A safe→dangerous re-ramp was proposed and rejected — the categories already read fine and the redesign is about the screen's layout/chrome, not the category identity. The deck applies them as borders + a faint face wash, so they stay legible without shouting.

5. Interaction

Click a key → its code is appended to the name.
Type on the real keyboard → same thing, and the on-screen cap flashes its pressed state.
Hold Shift / Ctrl / Alt → the deck flips to that page live, and drops straight back when you let go. Physical modifiers are momentary — nothing latches, exactly like the shift layer on the keyboard under your hands.
Click a modifier cap or a page button → the page latches. A mouse can't hold a chord and click a key at the same time; and a latched page is the only way in when the OS eats the chord.
⌫ Backspace is token-aware — it deletes a whole <code> (e.g. all of <player>), never one character out of the middle of one. (In edit mode, below, Backspace is an ordinary character backspace, as it should be.)
⏎ Enter commits and closes; Esc closes.
Space inserts the Space tile.

Caps Lock (added 2026-07-11 — it's a real one)

Caps Lock is not "latch the Shift page". It behaves the way it does on every keyboard:

Letters only. The number row keeps typing digits, so PIKA2 needs no unlocking. (This is the whole reason it can't just be a latched Shift.)
Ignored under Ctrl/Alt. Ctrl+B is bold B with the caps light on or off.
Shift inverts it. Caps + Shift = lowercase.

The consequence is that the deck can show two pages at once — uppercase letters above a digit row — which is exactly what a keyboard does. The rules live in C++ (FontKeyboard::pageForKey), pinned by tst_font_keyboard, so the QML just asks. Qt exposes no portable way to read the caps light, so if it was already on before the screen opened, the deck corrects itself from the OS's own answer (event.text) on the first letter you press.

The two modes — keyboard vs. edit (decided 2026-07-11)

Page 3 is the Ctrl page, so Ctrl+C is bold C and Ctrl+V is bold V — they can't also be copy/paste. Rather than hide that behind a focus rule, the screen has two explicit modes and says which one you're in:

Keyboard mode (the default): the deck is live and holds the keys. The name field is a read-only display of what the deck is building — no caret it won't honour. Backspace removes a whole tile. The field's pen button switches to…
Edit mode: the field becomes an ordinary text field — caret, selection, Ctrl+C/V/Z, and a character-by-character Backspace. The whole keyboard fades out and goes dead: it has no say in what you're typing and shouldn't pretend to. The pen is replaced by a check (apply) and a cross (discard), so an edit is something you commit or throw away — never something that half-happened. Nothing typed here reaches the name until the check is pressed; the cross puts the field back. Enter applies, Esc discards. Leaving edit mode hands the keys straight back to the deck.

Known OS-level risks (honest ones)

Shift+Alt and Ctrl+Shift are Windows' switch keyboard layout shortcuts when a user has more than one input language installed; Ctrl+Alt is AltGr on non-US layouts. In those setups the held-modifier shortcut for pages 5–7 may be eaten by the OS. Mitigation: every page is always reachable by clicking its page button or the on-screen modifier caps — the held-key combo is a shortcut, never the only way in. This is a design constraint, not a bug to chase.

6. Implementation

C++ (new)

projects/app/src/mvc/fontkeyboardmodel.{h,cpp} — FontKeyboardModel : QAbstractListModel, exposed as brg.fontKeyboardModel.

A static, hand-written layout table: page[8][36] of font ind (0 = empty slot), plus the spacebar's tile. This table is the mapping in §3 and is the single source of truth — the QML never guesses.
Q_PROPERTY int page (0–7); setting it resets the model to that page's 36 keys.
Roles per key: fontInd, keyLabel ("A", "7"), alias, code, tip, category, categoryColor, renderMode, isEmpty.
Q_INVOKABLE int pageFor(bool shift, bool ctrl, bool alt), Q_INVOKABLE int indForKey(int page, const QString& key) (physical typing → tile), Q_INVOKABLE QString chopLastToken(const QString& str) (token-aware backspace).

Why C++ and not a QML table: it's testable (below), it keeps the delegate dumb, and the layout is data, not markup.

Rendering the animated tiles (the performance trap)

Each image://font/... request runs a full tileset build (TilesetEngine::buildTileset). Firing that once per key per animation frame (36 keys × 8 frames) would melt the UI — this is the same class of problem that already froze the old hover tooltip.

So: request the whole 16×16 tile sheet once per frame (image://tileset/..., exactly what TilesetDisplay already does) and have every key clip its 8×8 cell out of that one shared image — a classic sprite-sheet: Item { clip: true; Image { source: sheetUrl; x: -col*w; y: -row*w } }. One provider call per frame for the entire deck, and QML's pixmap cache hands the same pixmap to all 36 keys. Tile ind → row = ind / 16, col = ind % 16.

Exceptions (they aren't single tiles):

Multi-Char / Variable keys render via the existing TilePreview (expanded, static frame) — so <poke> shows "Poké" and <player> shows the actual player name in game font. Static, because expanding a Variable every frame is exactly what froze the old tooltip.
Control codes have no glyph at all (length 0) → the cap shows its alias as text ("New Line", "End of text") in the Control colour.

QML

New, in fragments/controls/name-full/:

KeyboardDeck.qml — the chassis, the rows, the structural keys, modifier state, key capture.
KeyCap.qml — one cap (tile / preview / label + superscript legend + hover/press states).
TileGlyph.qml — the sheet-clipping animated tile renderer.
PageStrip.qml — the 8 page buttons.
ColorLegend.qml — the left colour-reference rail.

Rewritten: FullKeyboard.qml, NameFullHeader.qml (loses the Grid/Tileset toggle), DetailView.qml (restyled).

Deleted (dead once the deck lands): PagedPicker.qml, SearchRoot.qml, SearchContainer.qml, SearchCriteria.qml, SearchParam.qml, SearchResults.qml, TilesetPicker.qml — and their app.qrc entries. TilePreview.qml stays (the deck uses it). The C++ FontSearchModel / brg.fontSearch go with them in the same change (decided 2026-07-11) — nothing lingers.

Settings is untouched: the six fontColor* values stay as they are (§4).

Tests

projects/tests/tst_font_keyboard (new), pinning the thing that would be catastrophic to get wrong silently:

Completeness — every font ind 1–255 appears exactly once across the 8 pages + the spacebar. No tile is unreachable; no tile is duplicated.
No collisions — no page maps two tiles to the same key; no page exceeds 36 keys.
Validity — every mapped ind resolves in FontsDB.
Identity contract — page 1 key A is A, page 2 key A is a, page 1 key 7 is 7, … (a regression guard on the promise that the alphabet is where you expect it).
chopLastToken — removes a whole <code>, a plain char, and is safe on an empty string.

Plus tst_qml_screens must stay green (it loads every screen through the real engine).

Order of work

FontKeyboardModel + tst_font_keyboard (get the map provably right before any pixels).
New Settings palette.
TileGlyph (sheet clipping) + KeyCap — verify one cap renders and animates cheaply.
KeyboardDeck (rows, structural keys, modifier/page logic, key capture, token backspace).
PageStrip, ColorLegend, DetailView restyle.
Rewire FullKeyboard + NameFullHeader; delete the dead QML + qrc entries.
Screenshot pass on every page (mandatory manual review — overlaps, clipping, spacing), then in-app review with a save loaded.

7. Decisions (Twilight, 2026-07-11)

Lowercase is the base layer; Shift is uppercase. The first cut had it inverted (Gen 1 names are all-caps) — rejected, and rightly: it contradicted every keyboard the user knows. All-caps is what Caps Lock is for. (§3, §5)
Caps Lock is a real Caps Lock — letters only, ignored under Ctrl/Alt, inverted by Shift. (§5)
Physical modifiers are momentary; only clicking latches. (§5)
Two explicit modes — keyboard mode (deck live, field read-only, tile-wise Backspace) and edit mode (field is a normal text field, keyboard fades out + goes dead, check applies / cross discards), with the mode named in words on screen. (§5)
Space rides the spacebar — the only tile not on an alphanumeric key. (§1)
Category colours stay as they are — no re-ramp. (§4)
Dead search code went in the same change — FontSearchModel / brg.fontSearch and the six dead QML files were removed with the redesign. (§6)