RCA REPRO CORRECTION — preserve root/fix, replace the first 3x10 trace with this differential trace so the expected authoritative screen has no legitimate wrap scrap:

`ESC[1;1H界界ABCDEFTo ESC[2;1H ESC[2K ESC[3;1HSearch`

Width-aware authoritative VT: two 界 consume cols1-4; `ABCDEF` consumes cols5-10; `To` legitimately wraps to row2, then CUP row2 + EL clears it; final row2 is blank, row3 starts Search.

Current ScreenGrid: two 界 consume modeled cells1-2; `ABCDEFTo` fills cells3-10 on row1, so it believes row2 was always blank; CUP+EL is a no-op. Synthesized repaint emits row1's 10 scalar cells as 12 physical display cells, re-wraps `To` into physical row2, skips modeled-blank row2, then paints Search on row3. False stale `To` appears exactly above Search.

Regression: feed the raw trace to an independent width-aware stateful emulator for authoritative cells; separately feed ScreenGrid's synthesized repaint to a fresh instance of the same emulator; compare cells. Current code must fail. My earlier no-EL 3x10 example demonstrated overwide repaint but also had a legitimate original wrap, so it was not a sufficient state-equivalence oracle. The 1x20 CUP overwrite example remains valid.