feat: Wave 2 eval harness + image pipeline instrumentation (v0.13.0)

Closes #119

- Extract PRAutoBlogger_Image_Scene_Parser from image prompt builder (331→300 lines)
- Add frozen eval dataset (20 peptide topics at eval/dataset.json)
- Add LLM-as-judge scorer (gemini-flash-1.5 via OpenRouter, 3 axes)
- Add eval runner orchestrator with dry-run and Opik push support
- Add WP-CLI command: wp prautoblogger opik:eval [--limit=<n>] [--dry-run]
- Add image pipeline Opik instrumentation (image_prompt_rewrite span)
- Add content generator eval mode (suppresses publish + image gen)
- Version: v0.13.0

includes/core/class-image-scene-parser.php

@@ -0,0 +1,91 @@
+<?php
+declare(strict_types=1);
+
+/**
+ * Parses LLM output into scene/caption pairs and synthesises fallback prompts
+ * from article titles when the LLM call fails.
+ *
+ * Who calls it: PRAutoBlogger_Image_Prompt_Builder (static calls).
+ * Dependencies: None.
+ */
+class PRAutoBlogger_Image_Scene_Parser {
+
+	/**
+	 * Parse the LLM response into separate scene and caption parts.
+	 *
+	 * Expected format from the LLM:
+	 *   Scene description text here.
+	 *
+	 *   "Caption punchline here."
+	 *
+	 * Falls back gracefully: if no blank-line separator is found, the entire
+	 * response becomes the scene with an empty caption.
+	 *
+	 * @param string $raw Raw LLM response text.
+	 * @return array{scene: string, caption: string}
+	 */
+	public static function parse_scene_and_caption( string $raw ): array {
+		// Split on blank line (the format the system prompt requests).
+		$parts = preg_split( '/\n\s*\n/', $raw, 2 );
+
+		if ( count( $parts ) >= 2 ) {
+			$scene   = trim( $parts[0] );
+			$caption = trim( $parts[1] );
+			// Strip surrounding quotes from caption — the LLM wraps it in quotes.
+			$caption = trim( $caption, '"\'' );
+			return array(
+				'scene'   => $scene,
+				'caption' => $caption,
+			);
+		}
+
+		// No blank-line separator found — treat entire response as scene.
+		return array(
+			'scene'   => trim( $raw ),
+			'caption' => '',
+		);
+	}
+
+	/**
+	 * Extract the first paragraph from HTML content.
+	 *
+	 * Removes all tags and captures the text up to the first double-newline
+	 * or 200 characters, whichever comes first.
+	 *
+	 * @param string $html HTML content.
+	 * @return string Plain text of the first paragraph.
+	 */
+	public static function extract_first_paragraph( string $html ): string {
+		$text  = wp_strip_all_tags( $html );
+		$paras = preg_split( '/\n\n+/', $text );
+		$para  = isset( $paras[0] ) ? trim( $paras[0] ) : '';
+
+		if ( strlen( $para ) > 200 ) {
+			$para = substr( $para, 0, 200 ) . '...';
+		}
+
+		return $para;
+	}
+
+	/**
+	 * Rule-based fallback when LLM rewriting is unavailable.
+	 *
+	 * Kept deliberately simple — this only runs if OpenRouter is down.
+	 *
+	 * @param string $title           Main heading / topic.
+	 * @param string $supporting_text Additional context.
+	 * @return array{scene: string, caption: string} Scene for image gen, caption for HTML.
+	 */
+	public static function synthesize_visual_concepts_fallback( string $title, string $supporting_text ): array {
+		// Fallback produces a simple comic concept when the LLM is unavailable.
+		$scene = sprintf(
+			'A cartoon scientist in a lab coat looking bewildered while examining something related to: %s.',
+			$title
+		);
+
+		return array(
+			'scene'   => trim( $scene ),
+			'caption' => 'Science is full of surprises.',
+		);
+	}
+}