Fix: Kitty keyboard mode stack should live in the daemon, not a frontend Map

[pullapprove5]

What was wrong

desktop/ui/components/Terminal/kitty-keys.ts kept the negotiated kitty keyboard mode stack in a module-level Map, re-derived per window by parsing the PTY bytes xterm happened to see. Two consequences, both confirmed against the checkout:

1. Cold reattach lost the mode. TerminalPane rebuilds a fresh xterm from terminalReplay, which is the 256 KiB Ring (core/src/terminal/ring.rs). A program that pushes kitty mode once at startup and then produces more than that much output has had its push dropped from the ring, so a new window (or a web reload) started with an empty stack and silently encoded every chord the legacy way while the program still believed the protocol was on.
2. Two windows disagreed. A module-level Map is per-JS-realm, so each window held its own answer and at most one was right.

What changed

The negotiated state moves to the daemon, next to the vte::Parser that status.rs already runs over every PTY byte for the session's whole life. The encoder stays in the frontend, where it needs a DOM KeyboardEvent.

• core/src/terminal/kitty.rs (new) — KittyKeyboard: the per-screen-buffer mode stacks, with push/pop/set-in-place, screen switching, reset, and the prompt-time leak clear. Same semantics the TypeScript had (five-bit masking on push, "unwind everything" on an over-deep pop, malformed modes ignored, depth-8 wrap), ported with its reasoning and its tests.
• status.rs — the scanner now implements csi_dispatch/esc_dispatch: the four …u negotiation sequences, DECSET/DECRST 47/1047/1049 for which stack is live, RIS and DECSTR. A CSI ? u query is answered from the session's own stack; the bytes come back through StatusScanner::take_reply.
• session.rs — the PTY writer moves from Session into Shared so the reader thread can write the query answer on the thread that parsed it. A program now gets its answer whether or not a window is attached, instead of round-tripping to the frontend. No new lock nesting: the scanner lock is released before the writer lock is taken, in both directions.
• wire.rs — SessionStatus.kitty_flags (#[serde(default)]), so the mode arrives on the wire the way phase/cwd/title already do.
• Frontend — kitty-keys.ts keeps the encoder and replaces the stack with a small cache of what the daemon last reported (setKittyFlags/kittyFlags/forgetKittyState); terminalSlice.applyTerminalStatus — the one funnel both the live status stream and the cold-reattach replay pass through — records it. registry.ts no longer registers negotiation handlers. TerminalStatus.kittyFlags added to the types and the shared fixture.

Also handles DECSTR (CSI ! p), which the finding noted was unhandled; like RIS it clears a stuck keyboard.

Deliberately left out

The finding also suggested driving the leak clear off on_poll(at_prompt, …) rather than only OSC 133;A. I did not do that, and I think it would be a regression: the poller fires on an interval, so a tick landing after a shell that negotiates kitty mode for its own line editor (fish, zsh with kitty support) has emitted its prompt mark and pushed its flags would wipe that push — and keep wiping it. The prompt mark is a one-shot signal that arrives before the shell's push; a repeating poll observation isn't. Clearing at the OSC 133 prompt mark is retained, and the per-screen stacks (the main leak defence) are unchanged.

What I could and couldn't verify

• rustfmt --edition 2021 --check is clean on every touched Rust file (it also parses them, so there are no syntax errors).
• No test run was possible in this sandbox, in either language: there is no Cargo registry cache and no network (cargo build --offline fails fetching dependencies), Zig is not installed so the terminal feature could not link libghostty even with deps, and desktop/node_modules is absent with npm ci --offline failing on ENOTCACHED — so tsc --noEmit, vitest and cargo test all could not run. The Rust and TypeScript changes are unit-test-covered but those tests are unrun; please run scripts/test before merging.
• New Rust tests: kitty.rs ports the stack/screen-scoping/leak scenarios from the deleted TypeScript suite; status.rs adds byte-level tests for push/pop/set, the query reply, alt-screen scoping, the prompt-mark leak clear, that command marks do not clear a running program's mode, and RIS/DECSTR. The TypeScript suite keeps its encoder tests plus the real-xterm DECCKM seam test, and gains tests for the reported-mode cache.

---

Opened by a PullApprove implementation run (implement-finding v1) for the finding PA-25 — Kitty keyboard mode stack should live in the daemon, not a frontend Map.
Merging this is what closes the finding as fixed.