fix(skills): correct image generation error format and provider options

[amelia-letta]

Summary

• Builtin-skill-watch: image-generation@9ead5f0480b2-5493aa96fe388894
• Selected skill: image-generation

Stale claims and current owning source

1. The skill claimed 400/500 errors both return { "message": "..." }. In reality, 400 validation errors return a ZodError shape {"issues":[...],"name":"ZodError"} while 500 errors return {"message":"..."}. An agent following the skill literally would read response.message on a 400 validation error and get undefined.
2. size was annotated "(OpenAI)" but the server also uses it for Flux (parsed into width/height sent to bfl.ai), and Gemini supports an aspect_ratio option that the skill omitted.

Owning source: letta-ai/letta-cloud libs/sdk-cloud-api/src/lib/imagesContract.ts (generateImageBodySchema: prompt 1–32000, n int 1–4, input_images max 14, aspect_ratio) and libs/tmp-cloud-api-router/src/lib/images/imagesRouter.ts (parseSize → Flux width/height, aspect_ratio → Gemini responseFormat, 402 credits_required, 500 message body).

Focused validation

• Live probes of POST /v1/images/generations: invalid provider → 400 ZodError listing ["openai","gemini","flux"]; n=5 → 400 ZodError "must be less than or equal to 4"; missing prompt → 400 ZodError "Required"; valid gemini request → 200 with b64_json image and billing.credits_charged
• Source inspection of imagesContract.ts and imagesRouter.ts in letta-cloud
• bun run check — all 12 checks passed

Test plan

• Review that the error format description matches actual API behavior
• Verify size/aspect_ratio provider annotations match the contract