Phishing email detection remains a critical challenge in cybersecurity, with existing datasets suffering from outdated vocabulary, monolingual coverage, and non-representative negative samples. This paper presents an end-to-end pipeline for generating and detecting multilingual synthetic phishing emails that combines Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), and a self-correction quality gate. The pipeline produces a dataset of 16,818 emails across five European languages (Romanian, English, German, French, Italian). We evaluate five classification architectures, namely TF-IDF+LR, mDistilBERT, XLM-RoBERTa, mDeBERTa-v3, and ModernBERT, coupled with a systematic analysis of scaling laws. All models reach F1~1.0 from only 500 training examples. Cross-domain evaluation reveals a severe generalization gap (FNR=97.8%+/-0.5%, 3 seeds) when training exclusively on synthetic data; mixing in as little as 17\% real examples resolves it (FNR$->3.8%+/-1.3%), and a real-only control shows this gain comes mainly from having any real data at all, with mixed training's distinct value being that it retains full multilingual synthetic coverage that a real-only classifier cannot. Phishing generated by Qwen2.5-7B-Instruct evades a classifier trained on standard phishing in 40--44% of cases regardless of Group Relative Policy Optimization (GRPO) fine-tuning (base: 40.2%+/-12.9%, GRPO: 43.5%+/-11.0%, 3 seeds), indicating the adversarial advantage stems from generator stylistic diversity rather than reward-optimized evasion; an adversarial iterative game further shows the defender adapts completely after a single retraining round with 200 examples. Explainability analyses using LIME and Integrated Gradients confirm that classifiers learn relevant semantic phishing features rather than stylistic artifacts.