Preserve committed text during rapid WGPU key input

[kortexa-ai-bot]

Problem

During the native mini-app experiment in https://github.com/kortexa-ai/zendo.sh/issues/25, rapid injected text into Warren textInput changed the actual input string. Loading the resulting path produced ENOENT with the corrupted path, so this was not just a text-rendering defect.

Reproduction

• macOS 26.6.2 arm64, M4 Pro.
• Electrobun 2.0.2-beta.13; Cottontail 0.6.0-canary.13, revision b1130364d5cc70882241753cf6c1ad7a9db730a4; runtime SHA-256 098ca38de283a8bcdd87687f8245e0fa5c5d9f5086480c98c3f75b0691fa9d8f.
• Native Warren createUIWindow with a controlled textInput using signal, onInput: setValue, and onSubmit.
• Focus the field and inject /tmp/zendo-25-ui-fixture.mjs as a rapid sequence with CUA native typeText, then Return.
• Observed value: /zzzzznn22225i---ixurrjjjis (rather than the submitted path).
• A subsequent short burst also changed characters. Discrete key calls improved the result but still lost a leading slash when sent immediately after Cmd+Backspace.

The reproduction uses synthetic native key events. Normal typing rates, IME, and other platforms have not been tested. Root cause is not confirmed. Inspect the native WGPU key callback, FFIType.cstring ownership across deferred/threadsafe callbacks, and input event ordering. If the defect belongs in Cottontail, create a native dependency there with a minimal reproduction.

Acceptance

• Preserve exact ASCII and Unicode text, including punctuation, across bounded rapid key bursts and modifier transitions.
• Retain a native regression fixture under the exact shipped Cottontail artifact, plus a normal typing check.
• Do not paper over committed text corruption with a US keyboard map.
• Integrate the fix into Zendo before treating its composer text-entry path as validated.

Current workaround

The mini-app workbench uses a native Open file dialog to choose source modules; built-in counter and GPU fixtures also launch by button. This does not fix Warren text input.

[kortexa-ai-bot]

The rapid-input corruption has a confirmed Cottontail cause and a tested local fix: kortexa-ai/cottontail#4 (comment). Its native blocking relationship is attached to this issue. Publication is pending explicit fork Git authorization.

In the isolated patched Zendo app, rapid ASCII paths and Cmd+Backspace followed immediately by new input now reach Run file unchanged.

This issue remains open: composed Unicode needs a committed-text path in the native editor. A temporary app event recorder receives only a and z for CUA typeText('aé漢🙂z'); direct Option+E followed by E yields an empty text event then plain e. The FFI fixture independently proves exact Unicode transport. The macOS WGPU view forwards [event characters] without NSTextInputClient or marked-text handling. Warren also still needs multiline editing/selection/clipboard for Zendo blackboardsh#26. These results are not evidence of IME support.

Next implementation should provide native committed/composed text semantics with real keyboard/IME fixtures. AppKit's text-input contract: https://developer.apple.com/documentation/appkit/nstextinputclient?language=objc.

[kortexa-ai-bot]

The runtime string-lifetime fix is now committed and pushed to our Cottontail fork as 5b4910fb89845b48d427c55e8a78ea1735c5efcb, with upstream PR blackboardsh/cottontail#4 open for review.

The patch is unchanged from the native validation recorded above. Cottontail #4 is in Project 4 / In review. This issue stays open for the native committed-text/IME gap and for integration with a reproducibly selected runtime artifact; Zendo's dependency pins have not changed.