Non-Mac `write` backend: draft via llama.cpp GGUF (and/or CUDA) without MLX

[telivity-otaip]

Why

write currently requires Apple Silicon + MLX. That caps adoption hard. filter already has a GGUF/llama.cpp path and train has optional CUDA — contributors who bring write to Linux/Windows NVIDIA (or CPU GGUF) unlock a much larger audience and star/fork traction.

Goal

Make personality-protect write work when MLX is unavailable by selecting an alternate local backend:

1. GGUF / llama.cpp (preferred first milestone — already partially supported via [gguf] extra)
2. Optional follow-up: CUDA / transformers path for NVIDIA boxes

Still local-only. No cloud fallback.

Acceptance criteria

• On non-Apple hosts with GGUF weights present, write --channel post produces a draft
• Backend selection: auto prefers MLX on Apple Silicon, else GGUF if installed, else clear error
• Same prompt / style-card / retrieval contract as the MLX path (adapter=none default)
• Article channel either works or errors with an explicit "not yet" message (document either way)
• README Hardware section updated (what works where)
• Tests mock the llama.cpp binding so CI stays download-free
• Memory/timeouts documented for 16 GB machines

Pointers

• src/personality_protect/write.py
• src/personality_protect/filter.py (existing GGUF usage)
• src/personality_protect/download.py
• pyproject.toml optional extras gguf, cuda, mlx

Out of scope

• Hosted APIs / OpenAI-compatible remote endpoints
• Shipping weights in git

Difficulty

Hard — but the highest-leverage platform expansion in the repo. Issue farmers with llama.cpp experience welcome.