[windows/capsule] 修复胶囊和悬窗生命周期 P0

[Cooper-X-Oak]

This was generated by AI during triage.

Parent

#478

Triage

• Category: bug
• State: ready-for-agent
• Requested state label: ready-for-agent
• Requested repository labels (not applied by this run due label permissions): bug, windows, priority: high

What to build

修复 Windows beta 上胶囊/悬窗生命周期的 P0 断点：触发语音输入后必须显示录音胶囊；悬窗提示“再按一次快捷键结束”时，第二次快捷键必须实际结束录音并进入提交/处理流程；Esc 或点击退出不能造成应用未响应白框。

该 vertical slice 覆盖从 hotkey event 到 coordinator state、capsule window visibility、recording stop/cancel、idle cleanup 的完整路径。

Acceptance criteria

• Windows 10 22H2 下，语音输入触发后能稳定显示录音胶囊。
• 录音中再次按提示快捷键，可以稳定结束录音并进入提交/处理流程。
• Esc 取消和点击悬窗退出不会出现“应用未响应”白框。
• 日志能看到完整状态迁移：hotkey press -> capsule shown -> recording -> stop/cancel/submit -> idle。
• [windows] 悬窗弹出后再次按快捷键无响应 #476 复现路径重新验证时，不再出现再次按快捷键无响应。
• [Windows] [bug] 语音输入不显示录音胶囊 #470 复现环境重新验证时，不再出现语音输入无胶囊提示。

Blocked by

None - can start immediately.

Evidence

• [windows] 悬窗弹出后再次按快捷键无响应 #476: Windows 环境下，悬窗显示“录音中...再按一次右 Alt 结束并提问”后，再次按右 Alt 没有可见响应，悬窗停留在录音中状态。Reporter 提供截图、视频和 openless-2026-05-18T08-18-31.log。
• [Windows] [bug] 语音输入不显示录音胶囊 #470: Windows 10 专业版 22H2 / OpenLess 1.3.2，每次语音输入时桌面不显示任何 UI 提示或语音胶囊。
• 希望支持自动加入词典/OCR 导入，并修复 Windows 窗口卡死和白边问题 #445: 按 Esc 退出或单击小窗口时出现“应用未响应”的白框，需要等待一会儿才能退出。

Expected / Actual

Expected: 胶囊生命周期和界面提示一致；用户能看到录音状态，能通过提示快捷键结束录音，能通过 Esc 或点击取消，并且所有路径都会回到 idle。

Actual: Windows beta 用户报告胶囊不显示、二次快捷键无响应、Esc/点击退出卡成未响应白框。

Validation gate

• 在 openless-all/app 下运行：powershell -ExecutionPolicy Bypass -File .\scripts\windows-smoke-suite.ps1 -SkipRealAsr -SkipPrivacy。
• 如果修改 capsule 生命周期，补充或运行：powershell -ExecutionPolicy Bypass -File .\scripts\windows-capsule-lifecycle-smoke.ps1。
• 人工复测 [windows] 悬窗弹出后再次按快捷键无响应 #476 视频路径：打开悬窗、进入录音中、再次按提示快捷键、确认状态进入结束录音/处理。
• 人工复测 [Windows] [bug] 语音输入不显示录音胶囊 #470 环境等价路径：Windows 10 22H2 上启动语音输入，确认胶囊可见。

[Cooper-X-Oak]

This was generated by AI during diagnosis.

Evidence update - 2026-05-18

Local validation environment: Microsoft Windows 11 Home Chinese, version 10.0.26200, build 26200, 64-bit. This is not the reporter's Windows 10 22H2 environment, so Windows 10 confirmation remains required before closing #470/#445.

#470 capsule visibility

Expected: dry-run dictation start should emit CapsuleState::Recording; the OpenLess Capsule Tauri window should exist hidden before recording, become visible during recording, and hide after cancel.

Actual: replaced the flaky synthetic keybd_event smoke with deterministic single-instance CLI triggering plus process-scoped Win32 EnumWindows.

Validation passed:

cd openless-all/app
powershell -ExecutionPolicy Bypass -File .\scripts\windows-capsule-lifecycle-smoke.ps1 -ExePath (Resolve-Path '.\src-tauri\target\debug\openless.exe').Path -TimeoutSeconds 20

Observed result:

StartedDryRun : True
Cancelled     : True
Before        : exists=True visible=False
AfterStart    : exists=True visible=True
AfterCancel   : exists=True visible=False

Divergence layer: current-source product path on this machine reaches window visibility correctly. The previous smoke loop was unreliable because synthetic low-level hotkey injection and global title lookup could fail independently of product behavior.

#445 floating-window responsiveness

Expected: capsule click cancel, QA Esc close, and QA close-button click should reach coordinator, hide the relevant floating window, and not produce Windows hung-window evidence.

Validation passed:

cd openless-all/app
powershell -ExecutionPolicy Bypass -File .\scripts\windows-floating-window-lifecycle-smoke.ps1 -ExePath (Resolve-Path '.\src-tauri\target\debug\openless.exe').Path -TimeoutSeconds 20

Observed result:

CapsuleClickCancel : visible=True hung=False responsive=True -> visible=False
QaEscClose         : visible=True hung=False responsive=True -> visible=False
QaButtonClose      : visible=True hung=False responsive=True -> visible=False

Divergence layer: current-source capsule/QA close paths on this machine do not reproduce the reported "app not responding" white box. Remaining uncertainty is environment-specific: Windows 10 22H2 / reporter machine / exact beta build.

#476 second QA hotkey regression

Behavior fix: when QA panel is visible and QaPhase::Recording, route the QA hotkey to handle_qa_option_edge instead of closing/cancelling the panel.

Regression gate passed:

$env:OPENLESS_EMBED_COMMON_CONTROLS_V6_MANIFEST = '1'
cargo test --manifest-path src-tauri/Cargo.toml --lib coordinator::tests::qa_hotkey_during_recording_submits_without_closing_panel -- --exact --nocapture

Also verified default app build still passes without the test manifest opt-in:

cargo build --manifest-path src-tauri/Cargo.toml --bin openless

Current state

Ready for PR packaging from the local evidence side. Keep the source issue validation open for Windows 10 22H2 confirmation before claiming #470/#445 fully closed.

[Cooper-X-Oak]

This was generated by AI during triage.

Evidence update - 2026-05-18 real-machine rerun

Local validation environment: current Windows desktop workspace, branch fix/windows-handoff-core, debug binary rebuilt from local source with D-local temp/cache/target paths.

What passed

• cargo build --manifest-path openless-all/app/src-tauri/Cargo.toml --bin openless passed.
• Runtime launch smoke passed: process responded and log reported hotkey listener installed.
• OS hook smoke passed for the current local configured trigger path: low-level Windows hook observed vk=163 and coordinator received the hotkey press.
• Capsule lifecycle smoke passed for CLI start + explicit cancel: OpenLess Capsule existed hidden before recording, became visible during dry-run recording, then hid after cancel.
• PR Fix Windows hotkey conflicts, startup cancellation, and beta updater race #490 unit gates passed: action_shortcut 3/3, starting 6/6, beta_manifest 5/5.
• Right Alt mapping unit gate passed: windows_option_triggers_keep_left_and_right_alt_separate confirms RightAlt/RightOption map to VK_RMENU (vk=165).

What the reporter log confirms

openless-2026-05-18T08-18-31.log used vk=165, so the reporter path is Right Alt, not Right Ctrl. The log shows repeated stop/toggle edges during Starting, for example toggle stop edge during Starting — queued, then applying pending_stop edge → end_session immediately. That is the state-machine slice covered by PR #490.

New/remaining concern

The floating-window lifecycle smoke is now failing reproducibly on this machine at the QA Esc close step:

QA Esc close did not reach coordinator.

The fresh log stops after:

[coord] QA panel opened (awaiting Option to record)

There is no QA panel closed, history cleared after the scripted Esc. Earlier local evidence on this issue said the same smoke passed; this rerun supersedes that as a current risk signal. This looks like a QA window focus / frontend Esc delivery problem, not the same as the #490 Starting stop-state fix.

Gate decision

Do not close this issue from #490 alone. Treat #490 as covering the log-confirmed Starting repeated stop and action-hotkey retry slice, while keeping this issue open for QA Esc / floating-window close validation and a physical Right Alt HITL gate on the reporter-like path.

[Cooper-X-Oak]

Closing per author request. I am pausing involvement with this project for now.