feat: comic pipeline v0.10.0 — iterative chaining, title slugs, comics subfolder, 16:9v0.10.0

- Iterative image chaining: each page receives cover + prev page as pixel-level
  reference (multimodal) so the model copies appearance directly, not from text alone
- --story-theme flag overrides random genre pick per run
- Gemini now generates a short comic title (---COMIC TITLE--- separator) used as
  the output slug; all files named <slug>_cover.png, <slug>_page_1.png, etc.
- All comic output stored under comics/<slug>/ subdirectory
- Aspect ratio changed to 16:9 (ThinkPad X1 Gen 9 screen fit; 16:10 not supported)
- Bulgarian language rule moved to top of every prompt for stronger compliance
- Dynamic story-specific cover masthead + publisher logo badge matching story genre
- --story-theme "a thrilling space..." forces genre without touching random pool

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2 files changed, 72 insertions, 3 deletions

diff --git a/internal/image/nanobanana.go b/internal/image/nanobanana.go
index 8bcd945..befe74c 100644
--- a/internal/image/nanobanana.go
+++ b/internal/image/nanobanana.go
@@ -114,7 +114,16 @@ func (c *NanoBananaClient) Search(ctx context.Context, opts *SearchOptions) ([]S
 	fmt.Printf("Nano Banana Image Generation Prompt (%d chars): %s\n", len(prompt), prompt)
 	fmt.Printf("Nano Banana Image Generation: Using model '%s' with aspect ratio '%s'\n", c.modelName(), aspectRatio)
 
-	imageBytes, mimeType, err := nanoBananaGenerateImage(ctx, c, prompt, aspectRatio)
+	var imageBytes []byte
+	var mimeType string
+	// Use multimodal chaining when reference images are available — this keeps
+	// character appearance consistent across pages far more reliably than
+	// injecting a text-only character bible.
+	if len(opts.ReferenceImages) > 0 {
+		imageBytes, mimeType, err = c.generateImageWithRefs(ctx, prompt, aspectRatio, opts.ReferenceImages)
+	} else {
+		imageBytes, mimeType, err = nanoBananaGenerateImage(ctx, c, prompt, aspectRatio)
+	}
 	if err != nil {
 		if searchErr, ok := err.(*SearchError); ok {
 			return nil, searchErr
@@ -420,6 +429,61 @@ func (c *NanoBananaClient) generateImage(ctx context.Context, prompt, aspectRati
 	return imageBytes, mimeType, nil
 }
 
+// generateImageWithRefs sends reference images alongside the text prompt so the
+// model can match character appearance from the existing pages. This implements
+// the iterative chaining technique: each new page is conditioned on the visual
+// look established by previous pages rather than relying on text descriptions alone.
+func (c *NanoBananaClient) generateImageWithRefs(ctx context.Context, prompt, aspectRatio string, refs [][]byte) ([]byte, string, error) {
+	if aspectRatio == "" {
+		aspectRatio = nanoBananaAspectRatio
+	}
+	cfg := &genai.GenerateContentConfig{
+		ResponseModalities: []string{string(genai.ModalityImage)},
+		ImageConfig:        &genai.ImageConfig{AspectRatio: aspectRatio},
+	}
+
+	// Build multimodal content: reference image bytes first, then the instruction
+	// + prompt text. Leading with images means they are processed before the text.
+	parts := make([]*genai.Part, 0, len(refs)+1)
+	for _, ref := range refs {
+		if len(ref) > 0 {
+			parts = append(parts, &genai.Part{
+				InlineData: &genai.Blob{MIMEType: "image/png", Data: ref},
+			})
+		}
+	}
+	refNote := fmt.Sprintf(
+		"The %d reference image(s) above show the EXACT character appearance that must be preserved. "+
+			"Every character — same face, same age, same hair, same clothing, same animal breed and markings — "+
+			"must look IDENTICAL in the new image. Now generate:\n\n",
+		len(refs),
+	)
+	parts = append(parts, &genai.Part{Text: refNote + prompt})
+
+	resp, err := c.client.Models.GenerateContent(ctx, c.modelName(),
+		[]*genai.Content{{Role: string(genai.RoleUser), Parts: parts}},
+		cfg,
+	)
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: nanoBananaSource,
+			Code:     "API_ERROR",
+			Message:  fmt.Sprintf("failed to generate image with refs: %v", err),
+		}
+	}
+
+	imageBytes, mimeType, err := extractGeneratedImage(resp)
+	if err != nil {
+		return nil, "", &SearchError{
+			Provider: nanoBananaSource,
+			Code:     "NO_RESULTS",
+			Message:  err.Error(),
+		}
+	}
+
+	return imageBytes, mimeType, nil
+}
+
 func extractGeneratedImage(response *genai.GenerateContentResponse) ([]byte, string, error) {
 	if response == nil {
 		return nil, "", fmt.Errorf("no response from Gemini")
diff --git a/internal/image/search.go b/internal/image/search.go
index 8de97e0..80d5e33 100644
--- a/internal/image/search.go
+++ b/internal/image/search.go
@@ -27,8 +27,13 @@ type SearchOptions struct {
 	Page         int    // Page number (1-based)
 	ImageType    string // Type: "photo", "illustration", "vector", "all"
 	Orientation  string // Orientation: "horizontal", "vertical", "all"
-	CustomPrompt string // Custom prompt for AI image generation
-	AspectRatio  string // Override aspect ratio (e.g. "9:16"); empty = provider default
+	CustomPrompt    string  // Custom prompt for AI image generation
+	AspectRatio     string  // Override aspect ratio (e.g. "9:16"); empty = provider default
+	// ReferenceImages holds raw PNG bytes of previously generated images.
+	// When non-empty, the NanoBanana client sends them as multimodal content
+	// alongside the text prompt so the model can match character appearance
+	// across pages (iterative chaining technique).
+	ReferenceImages [][]byte
 }
 
 // DefaultSearchOptions returns sensible defaults for Bulgarian word searches