Recommend first-party Qwen/Gemma GGUF models

[danyurkin]

What

Rework the in-app Recommended models list to a curated set of first-party Qwen & Gemma GGUF models from our own AtomicChat/* Hugging Face org, spanning three hardware tiers (light / mid / heavy).

Why

The previous list mixed third-party repos (mlx-community/*, unsloth/*) with a couple of first-party entries, and one entry (AtomicChat/gemma4-e4b-it-GGUF) only resolved via an HF rename-redirect — its id did not match the canonical repo name and would break if the redirect were ever dropped. This change:

• points every recommendation at a first-party repo (downloads/stars land on our org, not mlx-community),
• uses exact, case-correct repo ids (verified against the HF API, no redirects),
• gives a clean tier ladder so users on weak and strong machines both get a sensible default.

New list (6 entries, all GGUF, all platforms)

Tier	model_name	Badge
Light (≤8 GB)	AtomicChat/Qwen3.5-4B-GGUF	Everyday use
Light (≤8 GB)	AtomicChat/gemma-4-E4B-it-GGUF	Everyday use
Mid (16 GB)	AtomicChat/Qwen3.5-9B-GGUF	Everyday use
Mid (16 GB)	AtomicChat/gemma-4-12b-it-GGUF	Everyday use
Heavy (32 GB+)	AtomicChat/qwen36-27b-GGUF	Everyday use
Heavy (32 GB+)	AtomicChat/gemma-4-31B-it-GGUF	Everyday use

What was removed

• The broken-id AtomicChat/gemma4-e4b-it-GGUF (replaced by the canonical AtomicChat/gemma-4-E4B-it-GGUF).
• Third-party mlx-community/* MLX entries and the unsloth/Llama-3.2-3B-Instruct-GGUF finetuning entry.
• The Laguna-XS-2.1 / qwen3-coder-30b-a3b coding entries.

Intentional trade-offs (follow-ups)

• No MLX / macOS entries in this PR. We currently have zero first-party MLX quants of these Qwen/Gemma models. Rather than keep recommending mlx-community/*, MLX is deferred: the 6 MLX-4bit quants are tracked in ATO-288 and will be added back in a follow-up PR (description_key: "hub:recForMlx", platforms: ["macos"]) once published under AtomicChat/*.
• No coding badge in this PR. Dropped for now; can be reintroduced with a first-party coding model.

Verification

• ✅ All 6 model_name ids return HTTP 200 on the HF API with no canonical-name redirect (exact case-sensitive match), and each repo contains .gguf files.
• ✅ recommended.json validates against models/schema.json (ajv draft-07 + per-entry manual check): required keys present, every description_key matches ^hub: and reuses the existing hub:recEverydayUse locale key, no stray properties, updated_at is valid ISO-8601.

🤖 Generated with Claude Code