feat(cloudflare): add vision support (Gemma 4, Llama 4 Scout, Llama 3.2 Vision)

[stackbilt-admin]

Summary

• CloudflareProvider previously had zero vision/image code — request.images sent via the CF path was silently dropped. This wires image input through to CF's OpenAI-compatible image_url content-part shape (base64 data URIs).
• Adds three vision-capable models: @cf/google/gemma-4-26b-a4b-it (256K ctx + tools), @cf/meta/llama-4-scout-17b-16e-instruct (multimodal + tools), @cf/meta/llama-3.2-11b-vision-instruct.
• Factory's analyzeImage() now dispatches to CF when configured; getDefaultVisionModel() falls back to Gemma 4 for CF-only deployments. Non-vision models and HTTP image URLs throw ConfigurationError with the vision-capable alternatives named.

Bumps to v1.3.0 (additive, strict semver).

Motivation

Reported by the foodfiles team — analyzeImage() routed to CF would silently drop the image. Options evaluated: call CF AI directly (strands the abstraction), fall back to Anthropic vision (works but costs per image), or properly add CF vision to the library. This PR is the proper fix.

Changes

• src/providers/cloudflare.ts (+104/-3): supportsVision = true, three new model entries in getModelCapabilities(), attachImagesToLastUserMessage + buildImageDataUrl helpers, validation for non-vision model + HTTP URL.
• src/factory.ts (+1): CF fallback in getDefaultVisionModel.
• src/__tests__/cloudflare.test.ts (+93): 6 new tests — payload shape, multi-image, data-URL passthrough, HTTP-URL rejection, non-vision rejection, capability flags.
• CHANGELOG.md, package.json: v1.3.0 entry.

Test plan

• npm run typecheck — clean
• npm test — 219/219 passing (14 cloudflare tests, 6 new)
• Foodfiles integration: verify recipe extraction works end-to-end against @cf/google/gemma-4-26b-a4b-it once published

Consumer usage

factory.analyzeImage({
  image: { data: base64, mimeType: 'image/jpeg' },
  prompt: RECIPE_EXTRACTION_PROMPT,
  model: '@cf/google/gemma-4-26b-a4b-it',
});

🤖 Generated with Claude Code