test(gateway): stop TTS state leaking into later tests

[OmarB97]

What changed

• Seed HERMES_VOICE_TTS=0 with pytest monkeypatch before exercising the runtime TTS toggle.
• This makes pytest own and remove the variable at teardown instead of leaving TTS enabled for later tests.

Root cause

monkeypatch.delenv(..., raising=False) registers no undo when the key starts absent. The tested voice.toggle handler then sets HERMES_VOICE_TTS=1, which leaked through the rest of tests/test_tui_gateway_server.py. A later fake completion containing partial answer complete consequently generated an MP3 and launched /usr/bin/afplay on the developer machine.

Verification

• scripts/run_tests.sh tests/test_tui_gateway_server.py -q -k "voice_toggle_tts_branch_also_carries_record_key or session_activate_returns_inflight_stream_before_completion" — 2 passed.
• macOS unified audio log for that run — 0 afplay audio events.

Scope

One test-only environment setup line plus the explanatory comment.