boxen REPL: mousewheel = history, option-arrows insert b/f literals

[jsavin]

Problem

Two input-handling regressions in the boxen REPL:

3a) Mousewheel navigates history instead of scrolling output pane

Mousewheel up/down events are mapped to history-prev/history-next (same as up/down arrows) instead of scrolling the output pane.

3b) Option-Left / Option-Right insert literal b / f

macOS Terminal sends ESC b (option-left) and ESC f (option-right) for word-jump. The boxen input line is consuming the ESC and then inserting the following byte (b or f) as a literal character instead of decoding the sequence as a word-jump command.

Repro

Mousewheel:

1. frontier-cli
2. Issue any command that produces a few lines of output
3. Scroll mousewheel up
4. Observe: history navigates instead of pane scrolling

Option-arrows:

1. frontier-cli in macOS Terminal.app (or iTerm2 with default "Option as Meta = Esc+" setting)
2. Type some text on the input line
3. Press Option-Left
4. Observe: b inserted at cursor (should jump cursor one word left)
5. Press Option-Right
6. Observe: f inserted at cursor (should jump cursor one word right)

Expected

• Mousewheel: scrolls output pane (up = older content, down = newer). History should remain bound to up/down arrows only.
• Option-Left/Right: jumps cursor by one word in the input line, same as readline/bash convention.

Context

Discovered during C.6 soak after flipping default REPL from linenoise to boxen (PR #802). Linenoise handled both correctly (option-arrows via its built-in word-motion bindings; mousewheel naturally passed through to the terminal since linenoise didn't capture mouse events).

Grouping these two as one issue / one fix since both live in the boxen input-line key/event decoder.

Severity

P1 each — both are muscle-memory regressions for terminal users. Workaround: --plain falls back to linenoise.

[jsavin]

Reopened. PR #808 reverted in commit 799583b — the fix made things worse in real terminals (SGR mouse byte leakage into input, Option-arrows still broken because Terminal.app sends modifier 9 not modifier 3, copy/paste hijacked, click-to-cursor missing).

Architecture review (2026-06-29) recommends Option C: build a custom input decoder, keep termbox2 for rendering. The input path is where all the terminal-quirk complexity lives; termbox2's input is too thin for what we need long-term.

Specific gaps:

• Modifier 9 (Meta) CSI sequences not in termbox2's mod-cap table
• SGR mouse burst-read fragility (partial sequences fall through to printable-key path)
• No bracketed paste (?2004h)
• No kitty keyboard protocol
• tb_deinit always CLEAR_SCREEN before EXIT_CA (cosmetic — affects copy-on-exit)

Testing gap: tmux harness can't catch these because tmux normalizes mouse/key events. Need PTY replay with recorded byte sequences from Terminal.app / iTerm2 / Alacritty before any retry.

Not redispatching automatically — this is a design decision worth its own planning round.