fix(ime): preserve Korean IME composition through split-commit and Backspace

[ohah]

Description

Follow-up to #9730 (merged). That PR addressed the Japanese IME Enter-form-submit case, but two multi-byte IME flows on macOS still misbehave:

1. The in-progress jamo/character is invisible until the next keystroke (split-commit)

This is the bug tracked in #3944 — "Text input in multi-byte Japanese doesn't appear until confirmation". The same root cause affects Korean IME: typing "가나다" slowly, the 나 does not appear after 가; it only shows up once 다 is typed and the buffer re-renders.

Cause: When typing ㅏ after 간, the IME sends setMarkedText('가') → insertText('가') → setMarkedText('') → setMarkedText('나') within a single interpretKeyEvents: pass. The trailing setMarkedText('나') places 나 as a buffer-level selection in the input field. The queued commit 가 then dispatches through user_insert, whose selection-replacement semantics overwrite 나 with 가 — losing the new composing jamo until the next composition update.

#9730's imeTouchedMarkedTextDuringInterpret guard only makes unmarkText conditional. It preserves the markedText object inside host_view.m, but the Rust input field's buffer selection still gets clobbered by the commit.

2. Backspace requires two presses to delete a marked syllable

With a marked ㄷ on 가나ㄷ, the user observes 가나다 → 가나ㄷ → 가나ㄷ (no change) → 가나 — the second Backspace is wasted.

Cause: macOS Korean IME implements Backspace on a marked syllable as:

setMarkedText('ㄷ') + insertText('ㄷ') + setMarkedText('') + doCommandBySelector(deleteBackward:)

doCommandBySelector: is currently a no-op (// no-op as we will be handling control characters in KeyDown events), so the IME's deleteBackward: signal is ignored. Only the queued commit ㄷ is applied, leaving a stray plain-text ㄷ that the next Backspace then deletes.

Changes (crates/warpui/src/platform/mac/objc/host_view.m)

• Split-commit handling — In keyDownImpl:'s commit branch, when a new marked text was started during the same interpretKeyEvents:, dispatch ClearMarkedText → commit → SetMarkedText instead of just the commit. This avoids the buffer-selection overwrite. New lastMarkedSelectedRange instance variable holds the most recent selectedRange for the re-dispatch.
• deleteBackward handling — doCommandBySelector: now tracks deleteBackward: / deleteForward: signaled by the IME during interpretKeyEvents: via a new imeRequestedDeleteDuringInterpret flag. When set, keyDownImpl: suppresses the queued commit so the marked text simply disappears, matching native macOS apps' behavior.
• Always end IME state on commit-only flow — Removed fix(macos/ime): don't submit form when Enter confirms Japanese IME conversion #9730's !imeTouchedMarkedTextDuringInterpret guard around the trailing unmarkText. The new split-commit branch above is what now protects newly-started marked text in split-commit cases, so the pre-fix(macos/ime): don't submit form when Enter confirms Japanese IME conversion #9730 invariant (commit ends IME state) can be restored — without it the stale marked-text state leaks into the next keystroke and breaks Backspace.

Linked Issue

Fixes #3944. Also addresses the Korean-specific symptoms reported in #8919 (follow-up to #9730).

• The linked issue (Text input in multi-byte Japanese doesn't appear until confirmation #3944) is labeled ready-to-implement.

Screenshots / Videos

Testing

Manual (macOS, Apple Korean IME, cargo run --features ime_marked_text)

Scenario	master (#9730)	This PR
"가나다" typed slowly	❌ 나 missing until 다 is typed	✅ each jamo visible immediately
"가나다" + Backspace × 3	❌ 가나다 → 가나ㄷ → 가나ㄷ → 가나 (one press wasted)	✅ 가나다 → 가나ㄷ → 가나 → 가
"안녕하세요" + Enter	✅	✅ (no regression)
Dead key ⌥E + g	✅ ´g	✅ (no regression)
Plain English + shortcuts	✅	✅ (no regression)

Automated (app/src/editor/view/marked_text_tests.rs)

Two regression tests added:

• test_split_commit_with_new_marked_text — verifies split-commit leaves the buffer as <commit><new marked>
• test_korean_ime_backspace_clears_marked_syllable — verifies Backspace cleanly removes the marked syllable with no stray plain-text leftover

cargo test --features ime_marked_text -p warp --lib marked_text
# 7 passed (5 existing + 2 new)

Note on prior PR (#9713)

A previous attempt (#9713) added a (!handled || wasComposing) guard to dispatch committed text on handled key events. @abhishekp106 closed it in favor of #9730 citing concerns about double-handled events. This PR drops the wasComposing guard entirely — the Enter case is already covered by #9730 + the deleteBackward: handling here, so the guard isn't needed.

Agent Mode

• Warp Agent Mode

[oz-for-oss]

@ohah

This PR is not linked to an issue that is marked with ready-to-implement.

Issue-state enforcement details:

• Associated same-repo issues checked: Korean IME: Last composing character is dropped when pressing Enter #8919

• Required readiness label: ready-to-implement

Readiness check:

• Korean IME: Last composing character is dropped when pressing Enter #8919: missing ready-to-implement; readiness labels present: none

To continue, link this PR to a same-repo issue such as Closes #123 in the PR description, and make sure that issue has ready-to-implement.

Powered by Oz

[oz-for-oss]

@ohah

This PR is not linked to an issue that is marked with ready-to-implement.

Issue-state enforcement details:

• Associated same-repo issues checked: #8919

• Required readiness label: ready-to-implement

Readiness check:

• #8919: missing ready-to-implement; readiness labels present: none

To continue, link this PR to a same-repo issue such as Closes #123 in the PR description, and make sure that issue has ready-to-implement.

Powered by Oz