Thank you for the feedback!

I have cleared up the description to clarify your points and modified test cases a little.

entry will always consist of 8 chars in a string and should be treated as such — you do not need to validate that it forms a real date.

If an item is missing / row is deemed invalid, return an empty string "" - do not discard the entire dataset.
(That was my bad, I did forget to implement that check in the tests. Should be aware of the dataset not allowed to be fully dropped when at least one row is valid now.)

Note that you cannot assign multiple values to one item
and
If multiple possible matches exist for a field (e.g., multiple single-letter values that could be name_t), use the first one that fits according to the schema inference rules.

Hope that cleared things up a bit and I have revised it properly.
More feedback is welcome!

Yes, discipline is set to "Puzzles" and this does take a bit of working with the information given to figure out.
I have updated the description a little to be more clear.

Your task: reconstruct the original table rows from the corrupted export file.

Mind telling me what exactly you have attempted in your experiments / want a tip on how to solve it?

Thanks for the feedback!
I added 5 more example tests (now there are 7 in total) for the user to see / experiment with.

Think that is sufficient, or are more necessary?

Respect for keeping it THAT short and effective!

Hi!
Genuinely curious, how long did it take for you to write all that code?
Looks pretty advanced.