fix(models): gate Browse installs to chat brains only

[quiet-node]

Description

Browse-all and paste-repo no longer treat every GGUF as a chat model. Multi-file HF repos often ship vision projectors and helpers next to text weights; users were downloading those small files, seeing Ready, then failing at engine start (for example unsupported model architecture: 'clip'). This change classifies artifacts by role and only offers supported chat weights as primary downloads.

Key changes

• Add a pure GGUF role classifier (Primary / Projector / Helper) driven by metadata when present, with filename as a soft list-time signal.
• Browse and paste listings show only chat-weight candidates; mid-name projectors (for example Bonsai-27B-mmproj-*.gguf) and helpers (dspark / MTP / draft) are not standalone downloads.
• Brain install still auto-attaches a projector companion by role onto one manifest row for --mmproj.
• Finalize and load refuse non-primary artifacts with clear user-facing copy; CLIP load errors map to projector messaging.
• Docs updated for Browse vision detection; staff picks and text-only single-file models keep existing behavior.

Testing

• Unit tests cover classification, listing, resolve attach/reject, finalize gate, and load preflight paths.
• bun run test:all:coverage (100% line gate) and bun run validate-build passed in the worktree.