fix: mirror keyboard protocol state onto the real terminal and decode encoded C-a everywhere

[kylecarbs]

Problem

In boo ui, with a kitty-protocol application focused (Claude Code and friends):

• Shift+Enter does not insert a newline. The terminal sends a plain \r, indistinguishable from Enter, so the app submits instead.
• Esc only registers on the second press. The app receives a bare 0x1b and, running in kitty mode, waits for the rest of a sequence that never comes; the next Esc flushes the first.

Both work in boo at. Two adjacent holes surfaced during review and are fixed here too:

• boo at's C-a prefix silently breaks under modifyOtherKeys. The attach repaint has always replayed an app's CSI > 4;2m (vim-class apps) onto the client's real terminal, but keys.zig never decoded the resulting CSI 27;5;97~ encoding of Ctrl+A that xterm-faithful terminals then send.
• C-a does not work on non-Latin keyboard layouts under kitty. A Cyrillic layout reports Ctrl+ф (CSI 1092:1060:97;5u); the parsers ignored the base-layout-key subfield.

Root cause (boo ui)

boo ui composites sessions through a local ghostty-vt terminal. When the focused app enables a keyboard protocol, the enable sequence is consumed by that local terminal and never reaches the user's real terminal, which keeps sending legacy encodings. boo attach does not have this problem because the daemon repaint replays keyboard state onto the client's real terminal (window.zig repaint with formatter.extra.keyboard).

Fix

• Mirror keyboard state: Ui.syncKeyboard writes the focused view's kitty flags (CSI = flags ; 1 u) and modifyOtherKeys state (CSI > 4 ; 2/0 m) to the real terminal whenever they change, matching what an attach repaint replays. Both are suspended while a UI prompt or key-driven mode reads byte-oriented input, restored after, and re-synced immediately after input handling.
• Parse encoded input: a shared keys.Protocols struct ({kitty, modify}) gates decoding in both parsers (keys.Parser for the daemon, InputParser for the UI):
  • kitty CSI-u: CSI 97;5u arms the prefix (releases, repeats, modifier keys handled per parser; in the UI a second discrete press dispatches the C-a C-a focus-last binding like a second raw 0x01), the command key decodes, everything else (Shift+Enter CSI 13;2u) passes through verbatim.
  • modifyOtherKeys: new keys.parseModify decodes CSI 27;mods;cp~ for the prefix and command keys; the protocol has no event types, so a repeated prefix stays armed in boo at and dispatches C-a C-a in boo ui, matching each parser's raw-byte semantics. Paste markers and F-keys share the ~ final but never the 27 marker.
  • non-Latin layouts: parseKitty now reads the cp:shifted:base subfields and keys.effectiveCp matches the prefix and command keys through the base-layout key when the primary codepoint is non-ASCII (Ctrl+ф = C-a on Russian layouts). ASCII primaries always win, so AZERTY-style Latin layouts keep their legacy typed-character semantics.
• Canonical Esc event (UI): the Esc key (kitty CSI 27u press, or a lone 0x1b after the 50ms flush) is a .esc event carrying the original bytes; it drives the existing cancel behaviors and otherwise forwards exactly as encoded. modifyOtherKeys never encodes an unmodified Esc, so no extra handling is needed there.
• Quit drain (UI): quitting restores the terminal through client.restoreTty (shared with plain attach), draining held-key repeats so a still-held C-a C-d cannot EOF the shell underneath.
• The UI hold grammar is active regardless of protocol state (decoding is not), so sequences in flight across a mirror change replay whole, and legacy F-keys no longer leak 15~ into prompts as text.

Legacy terminals are unaffected: without mirrored state the parsers decode nothing new, and terminals without protocol support ignore the mirror sequences.

Testing

• 33 parser unit tests across keys.zig and ui.zig: kitty prefix arming/release/repeat, double C-a press, encoded command keys, Esc press/release/modified, split-Esc disarm, modify prefix/command/repeat/passthrough/off/split, dual-protocol decode, Cyrillic prefix+command, AZERTY non-match, paste protection, mid-hold mirror flips, single-chunk F-key replay.
• 7 PTY integration tests: kitty flags and modifyOtherKeys mirror on/off in boo ui, Shift+Enter and Esc verbatim passthrough, kitty- and modify-encoded C-a intercepted without leaking (both boo ui and boo at), prompts suspend and restore both mirrors, boo at reattach repaint replays >4;2m.
• zig build test-all (113 unit + 66 integration) passes, integration re-run 3x for flake checking; zig fmt --check clean.

Decision log

• Mirroring over translation: translating legacy input to protocol encodings UI-side cannot work; the terminal never encodes Shift distinctly for Enter in legacy mode, so the information is lost before boo ui sees a byte. Mirroring reproduces exactly the contract boo at provides.
• Suspending mirrors during prompts instead of teaching every prompt the encodings: under kitty report-all-keys even plain printable keys arrive encoded, which would break rename/goto/confirm-kill byte handling.
• modifyOtherKeys mirroring was removed mid-review, then restored once both parsers decode CSI 27;mods;cp~: mirroring it without the decoder would have broken the C-a prefix whenever vim-class apps were focused. ghostty itself keeps Ctrl+letter legacy even in mode 2, but xterm encodes everything modified, so the decoder handles both worlds.
• Base-layout matching is gated on a non-ASCII primary: kitty reports cp:shifted:base, and matching base unconditionally would turn AZERTY's Ctrl+Q (cp q, base a) into the prefix — surprising, and different from its legacy 0x11 byte. Non-Latin layouts (the actual target) always have non-ASCII primaries.
• .esc carries the original bytes so an unconsumed Esc forwards in whatever encoding the terminal used.
• Consumed-press releases pass through by design: kitty apps must tolerate unmatched release events, so no swallow state is kept.
• No version bump: releases are separate chore: release PRs in this repo.

This PR was generated by Coder Agents on behalf of @kylecarbs.