## Summary

The "Recommended for coding agents" cards (`apps/code`) previously
showed the icon of whichever provider served the model cheapest. This is
misleading — e.g. **GLM-5.2** rendered the cheapest serving provider's
logo rather than its actual maker.

Now each card shows the model's **brand/family logo** based on
`model.family`, where `family` denotes the model's **creator**.

## Changes

- **Icon by family**: Added `getModelFamilyIcon(family)` +
`ModelFamilyIcons` map in
`packages/shared/src/components/provider-icons.tsx`, reusing existing
SVG React logos and falling back to the LLM Gateway logo for families
without a dedicated brand icon. Added three new brand icons that didn't
exist yet: `GeminiIcon`, `MetaIcon`, `NvidiaIcon`.
- **`CodingModelsShowcase`** now derives its card icon from
`model.family` instead of the cheapest provider mapping (pricing/context
still come from the cheapest provider).
- **Renamed the `glm` family to `zai`**: family should denote the
creator, and GLM models are made by Z.ai. Updated all GLM model
definitions (`packages/models/src/models/zai.ts`) and the gateway
fallback test fixtures. Dropped the now-redundant `glm` icon alias since
the `zai` family already maps to the Zai logo.

## Family → logo coverage

`alibaba, anthropic, atlascloud, bytedance, deepseek, elevenlabs,
google→Gemini, meta, minimax, mistral, moonshot, nvidia, openai,
perplexity, reve, sakana, xai, xiaomi, zai` all resolve to a brand icon;
unknown families fall back to the LLM Gateway logo.

## Testing

- `pnpm turbo run build --filter=@llmgateway/models
--filter=@llmgateway/shared --filter=code` ✅
- `pnpm vitest run apps/gateway/src/fallback.spec.ts` ✅ (50 passed)
- `pnpm lint` / `pnpm format` ✅

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>