feat(firebase-ai-logic-basics): add image generation and on-device hybrid inference guides

[AustinBenoit]

Summary (Stack 4/5)

Extracts the remaining Web-only usage sections from web_setup.md into dedicated capability guides (image_generation.md and on_device_hybrid.md), completing the SKILL.md routing table and trimming web_setup.md to its final setup-only state.

Changes

• Adds image_generation.md (Nano Banana responseModalities: [TEXT, IMAGE] across all 4 platforms; Imagen deprecation notice).
• Adds on_device_hybrid.md (Web Chrome mode: InferenceMode.PREFER_ON_DEVICE + inCloudParams / onDeviceParams + mobile platform stubs).
• Finalizes web_setup.md and SKILL.md.

[gemini-code-assist]

Code Review

This pull request reorganizes the Firebase AI Logic documentation by extracting web-only patterns into dedicated capability guides for "Image Generation (Nano Banana)" and "On-Device Hybrid Inference (Web)", updating the capability table in SKILL.md and cleaning up web_setup.md. The review feedback identifies critical improvements for the code examples in the new guides, specifically recommending the use of safe calls in Kotlin to prevent runtime crashes when handling empty or blocked model responses, and adding an error handler to the JavaScript FileReader promise to avoid hanging indefinitely on file read failures.

[gemini-code-assist]

In Kotlin, candidates can be empty, and content within a Candidate can be null (for example, if the response was blocked by safety filters). Calling .first() on an empty list throws a NoSuchElementException, and accessing .content directly can lead to a NullPointerException or compilation error.

Using safe calls (firstOrNull() and ?.) is highly recommended to prevent crashes.

Suggested change

	val generatedImageAsBitmap: Bitmap? = model.generateContent(prompt)
	// Handle the generated image
	.candidates.first().content.parts.filterIsInstance<ImagePart>().firstOrNull()?.image
	val generatedImageAsBitmap: Bitmap? = model.generateContent(prompt)
	// Handle the generated image
	.candidates.firstOrNull()?.content?.parts?.filterIsInstance<ImagePart>()?.firstOrNull()?.image

[gemini-code-assist]

Similarly to the generation example above, candidates can be empty and content can be null if the model's response is blocked or fails. Use safe navigation to avoid potential runtime exceptions.

Suggested change

	val generatedImageAsBitmap = model.generateContent(prompt)
	// Handle the generated image
	.candidates.first().content.parts.filterIsInstance<ImagePart>().firstOrNull()?.image
	val generatedImageAsBitmap = model.generateContent(prompt)
	// Handle the generated image
	.candidates.firstOrNull()?.content?.parts?.filterIsInstance<ImagePart>()?.firstOrNull()?.image

[gemini-code-assist]

In the fileToGenerativePart helper function, the FileReader promise does not handle the onerror event. If reading the file fails (e.g., due to permission issues or file corruption), the promise will remain pending indefinitely, which can hang the application flow. Adding an onerror handler to reject the promise ensures robust error handling.

Suggested change

	async function fileToGenerativePart(file) {
	const base64EncodedDataPromise = new Promise((resolve) => {
	const reader = new FileReader();
	reader.onloadend = () => resolve(reader.result.split(',')[1]);
	reader.readAsDataURL(file);
	});
	return {
	type: "inlineData",
	inlineData: { data: await base64EncodedDataPromise, mimeType: file.type },
	};
	}
	async function fileToGenerativePart(file) {
	const base64EncodedDataPromise = new Promise((resolve, reject) => {
	const reader = new FileReader();
	reader.onloadend = () => resolve(reader.result.split(',')[1]);
	reader.onerror = reject;
	reader.readAsDataURL(file);
	});
	return {
	type: "inlineData",
	inlineData: { data: await base64EncodedDataPromise, mimeType: file.type },
	};
	}