https://commons.wikimedia.org/wiki/File:Default_ocr.png

I think it could help if "short-term, debug only" images were not uploaded with a generic name to Commons (what's the educational value?) but uploaded here.

If the OCR software can read 2 columns text on english wikisource correctly, why does it merge the columns into one when it reads Marathi/ hindi pages?

Hard to say - there isn't much space between the two columns and software interprets, there might be bugs in software, etc.

It does look like Google isn't great at this. There's not much we can do about that, but have you tried using Tesseract instead (with Wikimedia OCR)? For example, does this look okay?:

https://ocr.wmcloud.org/?image=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fthumb%2Fa%2Fa2%2FTagorechiGoshti-Marathi.djvu%2Fpage5-543px-TagorechiGoshti-Marathi.djvu.jpg&engine=tesseract&langs%5B%5D=mr&psm=3

The IndicOCR gadget is using the https://indic-ocr.toolforge.org/ backend, which (if I'm remembering things correctly) is using the Google Drive API and so is a different OCR engine to the Google OCR provided by Wikimedia OCR.

Yes, this looks OK. But on Marathi Wiki-source site, I can see Indic OCR and Google OCR both of which are merging the columns incorrectly. Please add Tesseract to the site so that 2 column text can be processed.

I guess a good fix here would be for the Indic OCR gadget to add itself as an option in the 'Transcribe text' dropdown menu, rather than as a separate button. That would reduce confusion perhaps.

Also the page region OCR that's done but waiting for merge will hopefully help with this, as you can OCR each column separately if the OCR engine fails to segments the page correctly: https://phabricator.wikimedia.org/T294903