docs: add the hero image

One drawing at the top of the README. It reads left to right: a rack, a
processor, the AQL queue as a ring, the KV cache as a tape of blocks, and a
single token coming out the far end. That is the repository in one picture, so
it earns the space.

Downsampled 2:1 from the source, which for line art doubles as antialiasing, and
written back with no ancillary chunks at all. Three chunks: header, pixels, end.
No EXIF, no XMP, no colour profile, no embedded provenance manifest. The alpha
channel was fully opaque so dropping it was lossless. 7.3 MB to 1.4 MB.

docs/art/README.md records that the drawing was art-directed by us and rendered
by an image model, and that nothing in it is load-bearing: every claim in this
repository lives in prose, a command you can run, or a gate that fails when the
claim is false.

Author: maincode-dave

README.md

@@ -1,5 +1,9 @@
 # mainarch
 
+![A rack, a processor, the AQL queue drawn as a ring, the KV cache as a tape of
+blocks, and a single token coming out the far end, in isometric line
+work](docs/art/hero.png)
+
 Running a large language model on an AMD GPU normally means going through ROCm,
 and ROCm is a lot of software. There's the HIP runtime, the HSA runtime beneath
 it, the thunk library beneath that, and only then the kernel driver that actually
@@ -712,3 +716,7 @@ publicly.
 ## License
 
 Apache License 2.0. See [LICENSE](LICENSE).
+
+The one image in this repository was art-directed by us and rendered by an image
+model. [docs/art/README.md](docs/art/README.md) says which one and why nothing in
+it is load-bearing.

docs/art/README.md

@@ -0,0 +1,29 @@
+# Art
+
+`hero.png` is the one image in this repository. It reads left to right: a rack, a
+processor, the AQL queue as a ring, the KV cache as a tape of blocks, and a
+single token coming out the far end. That is what this stack does, so it seemed
+worth showing rather than describing.
+
+## Provenance
+
+The drawing was art-directed by us and rendered by an image model from a written
+prompt specifying the isometric line style, the paper ground, the single accent
+colour, and the objects in order. It is not hand-drawn and we would rather say so
+than have someone work it out.
+
+Two things follow from how it was made, and both were deliberate. The image
+carries no words at all, because image models are unreliable at text and a
+diagram that misspells `wavefront` is worse than no diagram. And nothing in it is
+load-bearing: every claim in this repository is made in prose, in a command you
+can run, or in a gate that fails when the claim is false. The picture is there to
+orient you, not to argue.
+
+## The file
+
+1461x806, 8-bit RGB, 1.4 MB. It was downsampled 2:1 from a 2922x1612 source with
+a box filter, which for line art doubles as antialiasing. The alpha channel was
+fully opaque so dropping it was lossless.
+
+It carries no ancillary chunks at all, so no EXIF, no XMP, no colour profile and
+no embedded provenance manifest. Three chunks: header, pixels, end.