Korean in a Tauri terminal: tmux underscores, then a WKWebView IME jamo leak I half-fixed

Two bugs wearing one costume

Typing Korean into a pane showed 한글 → __. English was perfect. That asymmetry is the whole tell: an underscore is an ASCII byte, so the bytes were round-tripping fine — something downstream was substituting them.

It was tmux running in non-UTF-8 mode. A Finder-launched Tauri bundle inherits essentially no locale, and pty.rs only forwarded LANG/LC_* when they were already set — no UTF-8 default. With no UTF-8 in the locale, tmux replaces every multibyte character with _; a Hangul syllable is 3 bytes / 2 cells → __. The fix is one block in pty.rs: default LANG=en_US.UTF-8 when the inherited one is missing or non-UTF-8 (an extension of the repo's RULE #5b — default a locale, don't just forward it).

The second costume: jamo leak

With __ gone, Korean now leaked as decomposed jamo — 안녕 → ㅇㄴㅎ.... I guessed twice and was wrong twice (a userAgent/isSafari tweak; then a terminal-recreation that regressed the working split panes). Both cost trust. So I stopped guessing and instrumented: a passive listener logging every compositionstart/update/end, keydown, and inputType to the Rust runtime.log, which I can actually read.

The capture was unambiguous: panes created at app launch fired zero composition events and isComposing was always false → xterm couldn't buffer the IME and shipped every intermediate jamo to the PTY. Panes created after the webview was ready (via split) fired composition events correctly and worked. So: WKWebView doesn't bind IME to a textarea created before its input method is ready, and a blur/focus can't repair it — only a freshly-created textarea binds. xterm's CompositionHelper, for the record, has no platform check at all; it relies purely on those events.

Where it landed

__ is fixed; Hangul renders as glyphs (added an Apple SD Gothic Neo fallback). The first-pane composition is unresolved — the honest workaround is to split once, which gives a freshly-bound pane. "Recreate the launch pane's terminal once the webview is ready, keeping its PTY/tmux session" is the right idea, but my first cut fired on every remount and broke split; it goes back to the backlog to be done once-guarded and tested.

The reusable win: instrument → standard sink → read beat guessing every time, and it's a live miniature of the platform thesis we just locked into the blueprint — agents emitting standard-format records that a reader can query. Full failure analysis in docs/ISSUES.md (2026-05-30).