[bhouse-nexthop]

Summary

Every repaint erased from the prompt down and printed the whole completion menu again, even when nothing in it had changed. On a 9600 baud serial console that is the dominant cost: a 64-entry ColumnarMenu is about 3.6 KB with colors, so each arrow key under an open menu takes several seconds to draw.

The painter now remembers the menu rows it drew. While the menu opens on the same row of an unscrolled screen of the same size, a repaint erases only the rows above the menu and resends only the menu rows that differ. Each remembered row carries the SGR state in effect where it starts, so a style that spans rows (a DescriptionMenu description) comes out the same on a redrawn row.

The menu is drawn whole whenever row-for-row reuse can't be proven to match a full paint:

• the screen scrolled, was resized, or the anchor is stale
• any row wraps, or holds a bare line feed (a ListMenu multi-line description)
• text after the cursor reaches the menu
• the menu is taller than the screen
• TERM=dumb
• resending the changed rows would cost more bytes than a full redraw

Anything that writes past the painter forgets the remembered rows: external messages, resizes, clears, $EDITOR, and the exit path.

There is no public API change.

Before / After

Bytes sent per event: a 64-entry ColumnarMenu, 80x24, DefaultCompleter, partial completions off, measured through the engine.

Event	ANSI on, before	ANSI on, after	ANSI off, before	ANSI off, after
Open menu	3672	3672	1375	1375
MenuRight	3672	349	1375	144
MenuDown	3672	586	1375	233
Repaint, menu unchanged	3672	87	1375	55
Type a character	3673	3673	1376	1376

Typing still redraws the menu whole. The engine marks the anchor stale whenever a host completer runs (#1130), and a stale anchor drops the remembered rows. #1297 lets an in-process completer opt out of that. Tab with partial completions on goes through the same path.

Additional notes

The tests compare screens rendered by a vt100 emulator, cell by cell with attributes and cursor, against a full paint of the same frame. They cover ANSI on and off; right and multi-line prompts; status lines; menus that grow, shrink, wrap or end on the last row; wide characters; tabs; bare line feeds; SGR spanning rows, including resets inside parameter lists and zero-valued color arguments; and the DescriptionMenu and ListMenu paths through the engine.

Two engine tests (columnar_menu_below_the_status_line_ignores_the_hidden_hint, status_line_stays_above_a_columnar_menu_taller_than_the_screen) fail on main when the checkout path is long, because DefaultPrompt prints the cwd and wraps. They pass from a short path, with or without this change.