bug: candle-glm-ocr duplicates characters inside long hyphenated identifiers

Description

candle-glm-ocr duplicates characters inside long identifiers. A 36-character hexadecimal identifier with hyphens on a page header came back with several of its groups repeated, on all three page images tested. Prose on the same pages was read correctly. Tesseract and PaddleOCR-VL read the identifier correctly.

Expected: identifiers are read once.

Steps to reproduce

1. Run candle-glm-ocr on a 150 dpi PNG of a page whose header carries a long hyphenated hexadecimal identifier.
2. Compare the identifier with the tesseract output for the same image.

Relevant files and configuration

xberg v1.2.4, release build with candle-cuda, NVIDIA L4, 2026-09-18. About 12 s per page, 5.5 GB of GPU memory.