[korexil]

• New FAQ section (zh/en/ja) with the symptoms people actually search for: stuttering, a long wait before speaking, silent VRAM spill, not fitting in 8 GB, uneven loudness, a pop before speaking, non-ASCII paths, WinError 206 / cu.obj.d. Every answer points at the measurement or pitfall it comes from and says where it was measured.
• The claim that the LLM and the TTS together need about 75% of the GPU had no measurement behind it; removed from pitfalls README: before/after clip and key numbers up front #9, how-it-works and the llamacpp_paced header (patch re-synced). The small-kernel explanation is now marked as unconfirmed.