A repository does not fit in a context window, so something gets cut: you compress the text, or you retrieve the chunks that look most like the issue. Hold the budget fixed across 300 real issues and both lose to the dullest option — work out which whole files matter and paste them in untouched. That resolved 43.7% and 46.7% of issues on two models, against 30.0% to 40.3% for the best compressor, and the reasons are mechanical: a chunk retriever often never shows the file that has to change, and a compressor shows it rewritten, so the patch the model writes no longer matches the file on disk.