The 30% to 94% recall claim is doing a lot of work in that table and I couldn't find how it was measured. I'd rather see a real number with the method than impressive without.

How many times do we need to solve this problem using glue and hacks?

While the idea of improving codebase search and documentation has some merit, any modern coding agent harness does not have the problem of starting each session blankly and giving generic advice. The actual issue that I have experienced is that it will waste tokens re-reading the codebase over and over again to understand your query, although you can mitigate that by proper use of AGENTS.md and documentation.

But more importantly, is it too much effort to write the readme yourself? Or at least to ask the AI to rewrite it so that it doesn't have all the tell-tale signs of AI slop?