[SinaHKazemi]

The deployed backend runs on a 512 MB instance. With the PyPSA-Eur set it uses about 450 MB while idle and reached 497 MB after a single generation plot. A generation plot also took about 10 s. This PR fixes both causes. Results don't change.

Changes

• mga_format.read_output: to read a few points, it used a filtered dataset scan (pq.read_table(..., filters=...)). The scan found the right row groups but buffered about 200 MB more than the rows it returned. It now looks up each point's row group from the row-group statistics (the format requires one row group per point) and reads them with ParquetFile.read_row_groups. A file without one row group per point still goes through the filter. validate_output shares the lookup.
• mix_output: it checked the set's members against every row of the output (isin over 420,000 Arrow strings): 2.5 s for a few points of generation. It now checks against the distinct keys.

Measured on the PyPSA-Eur set (local Docker backend)

	Before	After
mix_output, first call, includes the read	3.22 s, peak 554 MB	0.95 s, peak 311 MB
plot_data("generation")	2.83 s, peak 615 MB	0.15 s, peak 336 MB

Testing

• New tests check that the direct read returns the same table as a filtered scan, for points in any order, repeated points, an unknown point and an empty selection. Another test covers the fallback for a file without one row group per point.
• The mga-format and mga-compass tests pass (162), and so do the backend's (45).

Merging redeploys the backend on Koyeb.

🤖 Generated with Claude Code