Parquet connector

Runs SQL over Parquet files on an embedded DuckDB — no database to stand up. Each .parquet (or .pq) file in the directory becomes a queryable table named after the file (sales.parquet → sales), via DuckDB's read_parquet. Parquet is columnar and already typed, so it's the fastest file source — no header sniffing or type inference.

# sources.yaml
main:
  type: parquet
  directory: data        # folder of .parquet/.pq files, relative to the project

Then query the table by file name:

SELECT region, SUM(amount) AS revenue
FROM sales                -- data/sales.parquet
GROUP BY region

Key	Purpose
`directory`	Folder of Parquet files (each → a table by stem).
`files`	Or an explicit `{table_name: path}` map.

Extra: none — it's in the core install (the parquet reader ships in core DuckDB). Inherits the DuckDB connector's reconnect-on-fatal resilience, and dashdown query --tables / --schema <table> work out of the box.