docs: name the two things a reader will notice anyway

main.rs is about 136,000 lines and model_api.rs about 68,000. Saying so, and
saying they want splitting, is better than letting someone find it and conclude
nobody looked. The note points at the library code instead, which is normal
sized and is where the work is.

Same for the 84 CLI subcommands with names like
gpu-paged-mla-fp8-splitk-ragged-full-latent-gate. They are hardware gates named
after what they gate. just --list is the curated surface and the README says so.

Author: maincode-dave

README.md

@@ -645,6 +645,26 @@ order that makes each piece land before the next one needs it.
 
 ---
 
+### Two things you'll notice
+
+`crates/mainarch-cli/src/main.rs` is about 136,000 lines, and
+`model_api.rs` is about 68,000. That is not a good shape and nobody is pretending
+otherwise. Both grew as harnesses, one holding 84 CLI subcommands that are mostly
+hardware gates, the other holding a declarative metadata surface plus the tables
+that pin it. They are wide rather than deep, so they read as a catalogue rather
+than a call graph, but the honest summary is that they want splitting and have
+not been split.
+
+The library code is where the interesting work is, and it is normal sized.
+`olmo2.rs` is about 1,600 lines, `attn.rs` about 3,500, `model.rs` about 2,100,
+`gpu.rs` about 27,000 because it is the whole device layer. Start there. The
+reading order above skips both large files deliberately.
+
+The other thing is that `mainarch --help` lists 84 subcommands with names like
+`gpu-paged-mla-fp8-splitk-ragged-full-latent-gate`. Those are hardware gates,
+each pinning one thing that was hard to get right, and they are named after what
+they gate rather than for a reader. `just --list` is the curated surface.
+
 ## Layout
 
 ```