[vyncint]

Two screen-format issues.

#584 — searches were quadratic. searchable_row asked is_ascii() — a pass over every cell — for every row a search visited. Now each cell decides for itself whether it needs NFC (an ASCII cell is NFC already), the needle is folded on its own, and contains keeps its one whole-screen check per query. Every public entry point (find, find_all, locate, mask_matching, and with regex matches, find_match, find_all_matches, mask_matches) goes through these.

1000x1000	release, before → after	debug, before → after
find	3233 ms → 5.3 ms	26,770 ms → 74 ms
contains, last cell non-ASCII	2993 ms → 8.2 ms	27,656 ms → 101 ms

The regression test times all entry points on 200x1000 against a reference on the same screen (one pass for searches, building the masked screen for masks), best-of-N, bound 10x: old code 27–464x, new 1.1–1.6x. Stress is dispatched on this branch because the test is timing-based.

#582 — the JSON reader accepted cells the text reader refuses. A cell holding ESC]0;PWNED BEL ESC[2J loaded, and render --text / diff then drove the reader's terminal. Now refused, naming the cell, in the grid and the styled history: a control character (C0/DEL/C1, by the same cell_width rule Screen::parse now shares), printable ASCII after a cell's first character, an empty or ASCII-first cell marked wide without U+FE0F, and a continuation that holds text or follows no wide cell; and 0x0 screens. Unknown MouseModes bits are masked, so len() and iter() agree.

The compatibility promise decided every rule: each refusal is something no release could write (verified against vt100 0.16.2 and 0.18.1, vte 0.15, and unicode-width 0.2.0–0.2.2 over every code point). Kept readable on purpose: keycaps like 1️⃣ (0.18 writes them wide), ⚠️ in both its 0.16 and 0.18 forms, U+FFFD, cells whose width only a newer unicode-width table disputes, and a wide character 0.16 stranded in the last column after a resize. The compat corpus passes; removing each rule in turn turns exactly one test red.

Not in scope (noted): the window title and the text history are not validated; an old JSON file with a stranded wide character loads but its render --text output is refused by Screen::parse (pre-existing).

Closes #582
Closes #584