Add missing CF Workers AI models: llama-3.3-70b-fp8-fast, qwen2.5-coder-32b, gpt-oss-20b, mistral-small-3.1, qwen3-30b, llama-3.2-1b/3b, kimi-k2.7-code

[stackbilt-admin]

Summary

Running bildy (https://github.com/Stackbilt-dev/bildy) against the live CF Workers AI model catalog via `GET /accounts/{id}/ai/models/search?task=Text+Generation` reveals 7 models missing from `model-catalog` that are material to routing quality.

Missing models (CF API confirms all active)

Model	Suggested use cases	Notes
`@cf/meta/llama-3.3-70b-instruct-fp8-fast`	`BALANCED`, `HIGH_PERFORMANCE`, `TOOL_CALLING`	Best value model on Workers AI right now. Fast FP8 quant. Primary gap for `planning` route class.
`@cf/qwen/qwen2.5-coder-32b-instruct`	`COST_EFFECTIVE` (code)	Purpose-built for code generation. Critical for `code_draft` routing.
`@cf/openai/gpt-oss-20b`	`COST_EFFECTIVE`, `TOOL_CALLING`	Smaller sibling of `gpt-oss-120b` (already active). Good cheap tool-calling option.
`@cf/mistralai/mistral-small-3.1-24b-instruct`	`BALANCED`, `TOOL_CALLING`	Strong quality/cost ratio, vision capable.
`@cf/qwen/qwen3-30b-a3b-fp8`	`BALANCED`, `HIGH_PERFORMANCE`	Qwen3 generation, state-of-the-art for cost tier.
`@cf/meta/llama-3.2-1b-instruct`	`COST_EFFECTIVE`	Ultra-cheap for dead-simple summary/classify turns.
`@cf/meta/llama-3.2-3b-instruct`	`COST_EFFECTIVE`	Pair with 1b above as step-up model.
`@cf/moonshotai/kimi-k2.7-code`	`TOOL_CALLING`, `BALANCED`	Code-focused variant of kimi-k2.6 (already active).

Impact on bildy routing

Without `llama-3.3-70b-fp8-fast`, bildy's `planning` and `code_draft` route classes fall back to `gpt-oss-120b` which is the right capability tier but not the best cost model. Without `qwen2.5-coder-32b`, code generation routing has no purpose-built option.

Intentionally excluded

• LoRA variants (`-lora` suffix) — fine-tune infrastructure, not general routing
• `@cf/meta/llama-guard-3-8b` — safety classifier, not chat
• `@cf/aisingapore/gemma-sea-lion-v4-27b-it` — SEA-localized, niche