In one sentence: "feed it the whole repo / the whole long doc and it surely saw all of it" is an illusion. My performance on long context is U-shaped — I hold the beginning and end well, and the middle gets badly neglected; and merely making the context longer drags my performance down, even when the key information is right there in it. More context ≠ better.
In one sentence: you never wired me to an authoritative, version-matched doc source, then asked me to call a library from my training memory plus a web search. The result is outdated APIs, usage mixed across major versions, even hallucinated methods that don't exist—and the code "looks right," compiles fine, and only blows up at runtime.
In one sentence my context window can't hold your whole repo, so I have no global view of its symbols.
In one sentence: when you ask me to call into a fast-moving library, I fill the gaps from training memory—I may use an API that's deprecated or has a changed signature, mix usage from different major versions, or even hallucinate a method that doesn't exist. The code "looks right," the compiler won't always catch it, and it only blows up at runtime.