ZQ-77: add composer dictation and Desktop native-input foundation

[zqbake]

Summary

• add opt-in Gateway 2.1.0 composer dictation with an independent Qwen ASR adapter, capability gating, bounded session state, exact-once commit receipts, and explicit Memory-correction handling
• wire Web and TUI composers for partial/final text, keyboard interleaving, deterministic recent-range edits, continuous dictation, pause/cancel, and existing composer submission
• integrate the macOS system-input feature into the same Desktop product: a packaged universal InputMethodKit bundle, Desktop-owned transient Bridge, signed local IPC, lifecycle settings, global shortcut, shared Gateway/microphone ownership, and no independent daemon or provider credential
• install, repair, register, status-check, and uninstall Qwen Input from the Desktop settings flow; the user must explicitly enable and select it in macOS, and Qwen never silently changes the selected input source

Architecture and safety

Desktop renderer (existing microphone + Gateway WS) -> Electron main (narrow IPC) -> QwenInputBridge (signed stdio/socket broker) -> Qwen Input (InputMethodKit) -> focused macOS text client

• composer and native-input features remain disabled by default
• the IME and Bridge do not capture audio, connect to providers, hold provider credentials, or persist transcript/draft/target text
• partial text uses the target's marked-text composition; final/edit operations are limited to ranges owned by the current session and guarded by session/generation/target/sequence checks
• secure, stale, hidden, focus-lost, replayed, malformed, and uncorrelated operations fail closed
• Bridge loss removes the owned partial, reconnects to a replacement Bridge, and never writes into a new target
• ordinary keyboard input passes through; starting while another input source is selected fails visibly and leaves that source unchanged
• Accessibility, Input Monitoring, Full Disk Access, and administrator access are not required for the base InputMethodKit path; optional Voice Send remains separate and disabled

Installed-app and GUI evidence

On macOS 26.5.1 arm64 with the Debug/ad-hoc build:

• real install/register/status/repair/uninstall completed without administrator access; macOS System Settings was used to explicitly add and enable Qwen Input
• TextEdit accepted real marked partial, final, pause/resume, deterministic replace, and cancel operations
• Safari textarea and contenteditable accepted real partial/final/edit operations
• Safari password input failed closed without insertion; focus loss rejected the later final and did not write into the new target
• terminating the Bridge during an active marked partial removed the partial and allowed a replacement Bridge to reconnect to the same live IME controller
• cleanup restored ABC, removed the installed input method and runtime socket, and left no Bridge/IME process

Automated verification

• npm run native-input:test
• node --test desktop/test/native-input-*.test.mjs (34/34)
• npm test
• npm run lint
• npm run build
• git diff --check
• arm64 and universal DMGs pass hdiutil verify; packaged Desktop, Bridge, and IME pass strict/deep code-sign verification
• the universal Desktop, Bridge, and IME each contain both x86_64 and arm64

Remaining release gates

This PR remains Draft. The installed InputMethodKit path has been exercised in TextEdit and Safari, but the physical-microphone/provider path in the packaged Desktop UI, raw Terminal/Secure Keyboard Entry matrix, Intel runtime, and Developer ID hardened-runtime/notarized/stapled release have not been completed. The local artifacts are ad-hoc signed and are not production release evidence. The remaining matrix is documented in docs/desktop/native-input-testing.md and its Chinese counterpart.

Closes ZQ-77

Related #165