Digital text images, due to their information-bearing characteristics, are especially vulnerable to manipulation by adversaries using editing tools, hence undermining their authenticity and trustworthiness. Current research predominantly emphasizes the detection of manipulated regions, while investigations into trustworthy recovery remain relatively limited. To overcome these limitations, we present a recovery framework for tampered regions of text images, which includes two essential modules: watermark ID embedding and tampered region recovery, aimed at enhancing the reliability of text images. The watermark embedding module employs optical character recognition and BCH coding to embed error-correctable watermarks, guaranteeing semantic traceability even in cases of significant tampering. The recovery module concurrently incorporates a diffusion-based inpainting technique with large language model reasoning for the first time, enabling high-fidelity recovery of manipulated text. Experimental results showed that the proposed framework achieves enhanced recovery accuracy and demonstrates considerable potential for practical applications in maintaining the integrity and reliability of real-world text images.