d3ro-voice/scripts/verification
2026-11-01 19:39:06 +09:00
..
typing-fixture feat(commands): multi-select ordered command pipeline with step reordering and single-pass composition (AI-05, AI-07) 2026-10-31 02:54:54 +09:00
typing-runtime feat(commands): multi-select ordered command pipeline with step reordering and single-pass composition (AI-05, AI-07) 2026-10-31 02:54:54 +09:00
README.md feat: 회의 파일 가져오기와 입력 엔진 배포 보완 2026-11-01 19:39:06 +09:00
typing-race-proof.mjs feat: 회의 파일 가져오기와 입력 엔진 배포 보완 2026-11-01 19:39:06 +09:00

Typing-loss verification

These tools separate a proven service race from unverified claims about physical keys and real CPU/model load.

  1. Run node scripts/verification/typing-race-proof.mjs from the checkout root. It overlays three services from the pre-fix commit f49acf01057cc100b9caeace6eada26cf7bf4b54 in memory while keeping current dependencies and tests. Override with --baseline <commit> when investigating another baseline; the report records both commits and source hashes. A successful comparison requires the old services to corrupt the next Hangul word in three independent timing cases, the current services to preserve it, and unchanged-input correction to work in both. The other 13 assertions cover UIA snapshot contracts. This is not a complete installed-release A/B.
  2. Build typing-fixture/build.ps1 and typing-runtime/build.ps1. Quit the installed app and manually open the isolated runtime launcher as required by repository AGENTS.md. Do not run run-desktop.bat for this experiment; it kills unrelated Electron processes. Keep the test profile separate from the personal profile. The runtime receipt records the product bundle hash and product package version; the wrapper itself is not a released installer.
  3. Run the fixture with correction off, then correction on, saving each JSON under a distinct name. Use default Hangul IME mode first. Require exact calibration text and native IME composition messages. Compare all three requested paces under idle and bounded CPU load. A missing/aborted case, mismatched text, partial SendInput return or absent IME evidence is not a pass. Report achieved pacing, accepted key events and CPU samples during the input window, not just the requested load.
  4. Prove the correction path is active with Latin layout correction mode and a correction-enabled profile. A zero-loss Hangul run alone could simply mean the correction path never ran. Disable learned-text persistence throughout.
  5. For model loading, record local Ollama /api/ps before and after a request, request start/end, model load duration when available, and fixture input timestamps. Use only a model already installed locally, and do not unload a model being used by the user. A request made while a model is already resident is a warm-model test. A load ending before typing starts is not load-overlap proof. Record product-triggered warmup separately from direct Ollama stress.
  6. After a clean WinForms result, expand to Chromium/Electron and terminal input surfaces, then physical keyboard typing. Preserve failures and restore the normal installed app after closing the isolated runtime. Do not mark the full typing-loss issue verified on the basis of synthetic scan codes alone.

The fixture edits only its own text field, aborts when focus leaves, and bounds every CPU worker to 20 seconds. No clipboard or Unicode injection is used for the test corpus. The narrow check-to-SendInput focus race and synthetic-versus-hardware input remain limitations.

Windows contracts checked for this method:

  • SendInput returns the count inserted, serializes each array, and does not reset existing key state. Do not count a short return as success or retry a partial replacement.
  • GetLastInputInfo is session-local and its tick value need not increase. Compare change, not wall-clock ordering.
  • LowLevelKeyboardProc has a timeout and can be silently removed. Zero missing text alone does not establish continued hook health; use a positive correction control.