refactor(llm): split 694-line LLM class into 4 focused files

The monolithic Peptide_News_LLM class mixed HTTP transport, prompt
construction, batch orchestration, and AJAX handling. Split into:

- class-peptide-news-llm-client.php (199 lines) — OpenRouter HTTP
  transport with retry logic and response parsing
- class-peptide-news-llm-prompt-builder.php (142 lines) — keyword
  and summary prompt construction, article text fetching
- class-peptide-news-llm-ajax.php (87 lines) — bulk-generate AJAX
  handler
- class-peptide-news-llm.php (300 lines) — orchestrator with
  backward-compatible proxies for is_valid_model(), call_openrouter(),
  and call_openrouter_with_usage()

Extracted run_llm_task() helper to deduplicate the keyword/summary
processing blocks. All public API preserved — no test changes needed.

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

includes/class-peptide-news-llm-ajax.php

@@ -0,0 +1,87 @@
+<?php
+declare( strict_types=1 );
+/**
+ * AJAX handler for bulk AI summary generation.
+ *
+ * Processes one article per request so the admin JS can loop safely
+ * within PHP's max_execution_time. Reports progress (processed count
+ * and remaining count) so the UI can show a progress bar.
+ *
+ * Registered via wp_ajax_peptide_news_generate_summaries hook.
+ * Depends on Peptide_News_LLM for orchestration and
+ * Peptide_News_Cost_Tracker for budget awareness.
+ *
+ * @since      2.5.0
+ * @see        class-peptide-news-llm.php  Orchestrator that does the actual processing.
+ */
+class Peptide_News_LLM_Ajax {
+
+	/**
+	 * Bulk-generate AI summaries for articles that are missing them.
+	 *
+	 * Processes one article per request (to stay within max_execution_time).
+	 * The admin JS UI loops automatically until all articles are done.
+	 *
+	 * Side effects: nonce + capability check, option reads/writes, DB queries,
+	 * delegates to Peptide_News_LLM::process_unanalyzed(), sends JSON response.
+	 */
+	public static function generate_summaries(): void {
+		check_ajax_referer( 'peptide_news_admin', 'nonce' );
+		if ( ! current_user_can( 'manage_options' ) ) {
+			wp_die( esc_html__( 'Unauthorized', 'peptide-news' ), 403 );
+		}
+
+		if ( ! Peptide_News_LLM::is_enabled() ) {
+			wp_send_json_error( 'AI Analysis is not enabled. Please enable it and set an OpenRouter API key in Peptide News → Settings.' );
+		}
+
+		Peptide_News_Logger::info( 'Bulk AI summary generation triggered by admin.', 'llm' );
+
+		// Auto-correct known broken model names to a working free model.
+		$broken_models = array( 'qwen/qwen3.6-plus:free', 'qwen/qwen3.6-plus-preview:free' );
+		$working_model = 'google/gemma-3-27b-it:free';
+		foreach ( array( 'peptide_news_llm_keywords_model', 'peptide_news_llm_summary_model' ) as $opt ) {
+			if ( in_array( get_option( $opt, '' ), $broken_models, true ) ) {
+				update_option( $opt, $working_model );
+			}
+		}
+
+		global $wpdb;
+		$table = $wpdb->prefix . 'peptide_news_articles';
+
+		$remaining = (int) $wpdb->get_var(
+			"SELECT COUNT(*) FROM {$table}
+			 WHERE is_active = 1
+			   AND ( ai_summary = '' OR ai_summary IS NULL )"
+		);
+
+		if ( $remaining < 1 ) {
+			wp_send_json_success( array(
+				'processed' => 0,
+				'remaining' => 0,
+				'message'   => 'All articles already have AI summaries.',
+			) );
+		}
+
+		// Process ONE article per request to stay within PHP's
+		// max_execution_time (often 30s on shared hosting).
+		$processed = Peptide_News_LLM::process_unanalyzed( 1, true );
+
+		// Recount remaining after this batch.
+		$still_remaining = (int) $wpdb->get_var(
+			"SELECT COUNT(*) FROM {$table}
+			 WHERE is_active = 1
+			   AND ( ai_summary = '' OR ai_summary IS NULL )"
+		);
+
+		$last_run = get_option( 'peptide_news_last_llm_process', array() );
+		$errors   = isset( $last_run['errors'] ) ? $last_run['errors'] : array();
+
+		wp_send_json_success( array(
+			'processed' => $processed,
+			'remaining' => $still_remaining,
+			'message'   => sprintf( '%d article(s) summarized, %d remaining.', $processed, $still_remaining ),
+			'errors'    => $errors,
+		) );
+	}
+}