Several such systems exist, one of which is SQLite. https://www.sqlite.org/csv.html
I don't know if there's a virtual table extension for tab separated values, but SQLite imports them just fine.
It's intentionally available as a single source file:
https://www.sqlite.org/src/artifact?ci=trunk&filename=ext/mi...
For easy modification to handle TSV, TSVx, JSON, Pandas, YAML, or whatever else my heart desires.
ClickHouse and DuckDB can both do this. Pretty sure AWS has some built-in stuff for this too.
That said, I'm reading their documentation right now, and I think one or the other might slot into a different use-case I have where I do have rather large data. I'd last looked at this space quite a number of years ago, and both are way ahead of what I'd recalled.
For your model to work, what would be needed is the reverse: a system which allows me to interact with TSV files with SQL.
For 98% of uses of SQLite, performance and storage efficiency don't matter; we're talking about dozens or hundreds of records. It doesn't need to be well-implemented from that perspective. It does need to be robust. For example, databases solve a lot of problems especially around consistency in threaded systems which this would need to handle as well (although a KISS solution, such as handling all access from one thread / process, and simply blocking other requests during transactions, would be fine).
Something like that would likely eat SQLite's lunch.
q seems close. It'd need to support writes, table creation, and all the other stuff, perhaps be more actively maintained, provide documented API (not just command-line) access, and work with things like ORMs.