T61: the terminal logs each caret the input method asks for and each move of the caret; on the owner's Windows PC, WeType put the candidate box at the caret in 23 compositions, so the owner closed the jump a user found unreproduced

AGENTS.md

@@ -118,3 +118,4 @@ ELYTerminal. A terminal on GPUI and Ely components. `PRD.md` says what it does;
 - 2026-10-19: a plain click on the line the shell reads moves its cursor there, which a user asked for after 0.1.6, as Ghostty's `cursor-click-to-move`: a press with no modifier, one click, let go in its own cell with no selection, while OSC 133 says the shell reads a line, B since the last A or C, and the program tracks no mouse at the release. A shell whose prompt start says `click_events=1` or `2`, as fish from 4.1 says on its own, hears a click past the prompt's last cell before the cursor, or, where no prompt cell of its own line comes before the input, as under fish's empty prompt, from the first cell of the newest line typed, known by its number, not its place, since a prompt drawn at the screen's top leaves older lines below it, down to that line's last row or the cursor's, and never on a prompt cell, itself, a left press in SGR, its row the screen's or counted from the prompt's first, as kitty and Ghostty send it, and moves its own cursor, combining marks and its own redraws included: fish moves by code point, and its redraws drop the wrap flags a count would follow. Any other shell gets Left or Right arrows, one per cell, each pressed and let go as the keyboard would, through the key encoder, so DECCKM, kitty's flags and win32-input-mode shape them; the count runs only on the cursor's line, soft wraps joined, past the prompt's cells, which the prompt's link marks, so bash 3.2, which sends no C, moves nothing once its command left the prompt's line. Cells after the last character typed count only up to the cursor, so a click past the end goes to the end; a wide character takes one arrow. Never on the alternate screen. One line only, as Ghostty's `cl=line`. A right prompt or an autosuggestion past the typed text counts as text, as in Ghostty, so a click on it may take the suggestion. The first click of a double click moves the cursor before the word selects, as in Ghostty.
 - 2026-10-19: files dropped on the terminal type their paths, which a user asked for after 0.1.6, as macOS Terminal does: each path followed by a space, the paths of one drop in its order, sent as a paste, bracketed when the program asked, and kept literal for the shell the session started, which its name says, as the integration reads it: for sh, bash, zsh, fish and any other, each character Ghostty escapes gets a backslash, and `^` and `~` too, which zsh's `extendedglob` reads; for pwsh and powershell, the default on Windows, single quotes, each single quote inside doubled, PowerShell's curly ones too, since PowerShell reads a backslash as a plain character and expands `$` inside double quotes; for cmd, double quotes, which keep `&`, `|`, `<`, `>` and `^` inert, the backslashes ending a path doubled, since a program's parser reads `\"` as a quote, as `C:\` ends, and a path holding `%`, or `!`, which cmd expands inside them under `/v:on`, stops the drop. Windows Terminal's quotes around a path with a space alone leave `$(...)` and `;` live in PowerShell. A program run inside the session, `cmd` under PowerShell among them, gets the session's own quoting. A path not in UTF-8, or holding a control character, which would act as it is typed, stops the drop, logged. gpui hands a drop to the hitbox under the pointer only while the last input was not the keyboard, and a file drag counts as no input of its own: after typing, a drop with no pointer move over the window since does nothing, and nothing in ELYTerminal sees it, which the owner chose to leave so. `shot`'s `drop@x,y` plays a drop through `Window::dispatch_event`, so AppKit's drag session itself goes untested.
 - 2026-10-19: fallback fonts, which a user asked for after 0.1.6, since the Chinese and Latin fonts people like seldom come as one: `font_fallbacks` in the settings lists installed families, tried in order for a character Font family lacks, through gpui's `Font::fallbacks`, which CoreText's cascade list, DirectWrite's fallback builder and cosmic-text's chain each honor; the system's own fallback follows the last. On Linux gpui's cosmic-text path keeps the fallback the last character took while it covers the next, so shaping clusters stay whole: with two fallbacks, a character only the second covers carries the next one, which the first covers too, into the second; macOS and Windows choose per character. The owner chose to leave that so. Each family must load as the terminal draws it, as Font family must, so a family gpui would pass over without a word stops the start or the save, the error naming `font_fallbacks`; an empty name is refused. The field is new beside `font_family`, so a settings file from before loads as it was. Settings › Appearance lists them under Family: each an Ely `FontPicker` whose pick replaces it and a button that removes it, then Add fallback, which opens an empty picker whose pick adds one, its button closing it; a refused pick shows the saved family again, as Family's does. Each row keeps an input of its own, and a removal drops the removed row's input and takes the focus away, since Ely's combobox keeps a focused input's text, and an input kept by place would show, and on Enter save, the family removed. Each picker's id is its input's entity, not its place: Ely's combobox keeps its state by id and listens to the input it first met, so a picker keyed by place, after a removal moved the inputs, wrote another row's family into its new input at the next blur. Ligatures and the bundled Nerd Font for private-use icons stand as before; the cursor preview keeps the family alone.
+- 2026-10-19: a user found the IME's candidate box jumping on Windows after 0.1.6. gpui places it, through `ImmSetCompositionWindow` and `ImmSetCandidateWindow`, only when a composition starts, at the caret the terminal reports, the grid cursor's cell from the last frame, hidden or not. On the owner's PC, WeType under powershell.exe and under a status line redrawn every 100 ms put the box at the caret in 23 compositions out of 23, so the owner closed it unreproduced. The terminal logs each caret the input method asks for, at debug, and each move of the caret, at trace, so the next report carries where the box was told to go.