Replit is used a lot in this context. Their agent is good, but their circumvent-policies-to-get-something-in-front-of-execs-quickly is an amazing and mis-priced feature.
What the "right answer" is will vary widely on application and analyst. That's one reason there are so many coordinate ref systems. If you keep everything in 3857, you'll get answers in ~meters, but whether that's "right" depends on where and how large the distance or geom is and what your precision threshold is. So, really, everyone's needs are necessarily "specific."
DuckDB > geopandas, certainly for anything out of core. Though, I recently gave up on importing 70GB worth of large multipolygons (from a csv in hex wkb), and just used a postgis container. In concert with DuckDB's growth, I'd also mark the advent of geoparquet.
The big change, in my view, over the past decade in GIS software, is in compute and storage efficiency across the typical stack. DuckDB has become a part of this, but h/t to the advances from shapely, geopandas, geoparquet, and GDAL. There's a lot of overlap in that venn diagram, and credit should be spread around. QGIS is great, too, though I feel there is market opportunity to apply 90/10 to its massive feature set and move it to the web.
Zstandard (zstd), with a properly trained dictionary, could do well against larger (than alice29) texts. I doubt you'd see ~0.4 bpb, but 1-1.5 seems achievable. And it would be ~100% accurate with no GPU or other overhead necessary.
A quick experiment using a small selection of truncated texts yielded 1.8 bpb against alice29. I imagine other texts could do better.