fix(gateway): skip the auto-TTS voice reply for the desktop platform

[liuhao1024]

What does this PR do?

Stops the double auto-TTS playback on the desktop surface. When voice.auto_tts is enabled, two independent paths synthesize and play the same reply on the desktop: the gateway's _send_voice_reply (calling text_to_speech_tool and delivering via adapter.send_voice) and the desktop app's useAutoSpeakReplies hook (calling /api/audio/speak → playSpeechText). Neither knows about the other, so every desktop reply plays twice — logs show one "Generating speech" but two "TTS audio saved" entries per reply from two different PIDs (#90297).

_should_send_voice_reply now returns False for the desktop platform, deferring to the desktop's own hook (the authoritative speaker for that surface — it fires on every reply regardless of mode, and the gateway has no way to know whether the client spoke). This is the issue's first suggested fix; the alternative (teaching the hook to detect an already-spoken gateway message) adds cross-process coordination the simple skip avoids. Every other platform keeps its gateway voice reply unchanged.

Related Issue

Fixes #90297

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• gateway/run.py — _should_send_voice_reply skips event.source.platform.value == "desktop" right after the mode check, with the double-synthesis mechanism documented inline (auto_tts plays audio twice on desktop — gateway _send_voice_reply and useAutoSpeakReplies both fire #90297)
• tests/gateway/test_desktop_no_gateway_reply.py — three regressions: desktop is skipped even when global auto_tts would qualify; desktop is skipped even with an explicit all mode for the chat (the hook fires regardless); and Telegram with auto_tts on still gets its gateway voice reply (the skip is desktop-only)

How to Test

1. .venv/bin/python -m pytest tests/gateway/test_desktop_no_gateway_voice_reply.py -q — should pass (3 passed)
2. Observed result: on unmodified main (verified via git stash control), both desktop tests fail — _should_send_voice_reply returns True for the desktop platform and the gateway sends a voice reply the desktop hook will speak again; with this fix both shapes return False
3. Regression: tests/gateway/test_auto_voice_reply_format.py + tests/gateway/test_voice_mode_platform_isolation.py — 10 passed (voice-reply format selection and per-platform mode isolation unchanged)

Checklist

• Code follows the project's style guidelines
• Self-review completed
• Comments added for complex logic (why the desktop hook is authoritative and the gateway cannot know whether the client spoke)
• Tests added that prove the fix (both desktop shapes + other-platform guard)
• All new and existing tests pass locally (3 + 10 passed; main-control fail confirmed)
• Platform: macOS (pure predicate logic, platform-independent)

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Correct fix for the double-speak (#90297): the guard sits after the mode/auto-TTS gate and before any return True, so it covers every qualifying path, and the tests pin all three mode combinations. Suggestions:

1. gateway/run.py:21976 — the skip is completely silent, unlike every other exit in _should_send_voice_reply which logs "Auto voice reply skipped: ...". Add the same log line (with reason="desktop surface speaks locally") so a user who enabled /voice all on desktop and hears nothing isn't left debugging a mystery.
2. gateway/run.py:21976 — explicit per-chat /voice all is now silently ignored on desktop: there is no remaining way to get a gateway-synthesized reply on that surface even when the client hook is disabled. If that's intended, document it next to the voice-mode command help; otherwise consider keying the skip off the desktop hook being active instead of the platform alone.
3. gateway/run.py:21976 — magic string "desktop" is compared here and re-spelled in three test places plus wherever the session routing layer produces it. Promote a shared constant (or a real Platform member) so a future case drift ("Desktop", "desktop-app") can't silently disable the guard.
4. tests/gateway/test_desktop_no_gateway_voice_reply.py:18 — from types import SimpleNamespace # noqa: F401 (docstring reference only) is dead code; imports don't serve docstrings. Drop it.
5. tests/gateway/test_desktop_no_gateway_voice_reply.py:24 — _DesktopPlatform(str) with .value returning self is a clever stand-in; since the stated root cause is that the value comes from the session routing layer, importing/borrowing that layer's own factory would keep the stub honest if the runtime shape ever changes. Acceptable as-is with the docstring rationale. (nit)

No blocking issues found.

[liuhao1024]

Thanks — addressed in d1e9047:

• Silent skip (point 1): fixed. The desktop exit now logs Auto voice reply skipped: desktop surface speaks locally (mode=%s chat=%s) at debug, matching the other exits in _should_send_voice_reply.
• Dead import (point 4): dropped.
• /voice all having no desktop effect even with the client hook off (point 2) and the shared "desktop" constant (point 3): both change behavior/scope beyond this fix — whether the skip should key off the desktop hook being active rather than the platform, and promoting a shared constant across the session routing layer, are maintainer calls. Noted here so they aren't lost; happy to follow up on either if a direction is preferred.
• Test stub (point 5): keeping as-is with the docstring rationale, per your note.