The quest for performance is often also a cause for bloat: pushing something into a background thread now requires the UI thread to be aware that the background

The quest for performance is often also a cause for bloat: pushing something into a background thread now requires the UI thread to be aware that the background thread may not have the results yet - a lot of extra ifs around accesses to the data model. A fancy fast algorithm to solve a problem at larger scales faster is often more complex and performs slower with smaller inputs. So while people have added 5k or 10k LOC to the project to make fast for the occasional large input, performance for the smallest possible inputs may have reduced for an order of magnitude or so. Then there are the unavoidable features. Accessibility and localization are in that category. Suddenly you need to do extra work every step of the way. Look up translations, format strings with locale dependent rules, build a model for screen readers to process, alternate UI themes and layouts for vision impaired people... death by 1000 paper cuts, but also too important to ignore.