Harden GGUF role gating and fold projector size into the memory gate

[quiet-node]

Description

Thuki re-reads a GGUF's header at install and load so non-chat models can never become a Ready primary chat model, and folds a vision model's projector size into the pre-load memory gate so vision loads are sized accurately. This closes the gap where a renamed companion, an embedding model, or a mis-sized projector could slip past the browse-time filename heuristics.

Key changes

• Architecture gating at finalize/load. The local blob header is re-read before a model can become Ready or load as -m. Known non-chat families (embedding/encoder arches such as bert, t5, and clip, plus audio tokenizers) and projector/adapter roles are rejected with a clear "not a chat model" message. Missing architecture stays soft (filename heuristics) so an incomplete or list-time header never bricks an install.
• Projector size in the memory gate. A vision model's projector (mmproj) size, taken from the on-disk blob length when present or the curated registry size otherwise, folds into the pre-load memory gate and the Library/Discover fit hint through one shared helper, so the gate and the fit estimate never size a projector differently.
• Projector companion ranking. When a repo ships several projectors, Thuki ranks them like llama.cpp: shared directory depth with the chat weights first, then the closest quant bit-width.
• Browse-all listing. Only primary text weights with complete LFS metadata are listed; projectors, embedders, draft/MTP/dspark helpers, and LoRA adapters never appear as standalone chat downloads.
• Clearer engine-start errors. A wrong projector paired with text weights surfaces a fixable "Vision projector mismatch" message. The classifier requires the failure to actually name a projector, so a text-only load's error keeps its raw reason instead of being mislabeled a projector problem.

Testing

cargo test + cargo llvm-cov (100% line coverage), cargo clippy -D warnings, and the full frontend + backend build gate.

Notes

• DENIED_PRIMARY_ARCHES is a baked-in, pin-scoped denylist: an engine bump may need it extended when new non-chat families appear. It lives in config/defaults.rs with the other baked-in constants and is documented in docs/configurations.md.
• The Library fit badge for a Browse-installed vision model with no registry entry counts weights only, while the load gate folds the projector blob. This is an under-estimate on the badge (never a false block), kept deliberately because the badge builder has no model-store handle.