Made of little things.

# Current camera schemas and prompts

All 32 views were exported through the real soyLI editor on 2026-10-06, then
individually rendered with the built-in image generation tool. Each `*-schema.png`
is the exact 640×480 geometry projection; its `*-prompt.txt` contains the complete
generation instructions, including any targeted corrections after visual review.

[The render manifest](../room-render-manifest.json) maps every room/camera to its
JPEG, schema, prompt and hashes. [Level JSON](../../levels/demo.json) supplies the
camera bindings and geometry signatures. Twelve previous views were replaced and
twenty approach/recovery views received generated backgrounds for the first time.

Runtime images are 640×480 JPEGs in `src/art/`. High-resolution generated files,
discarded variants and the old room-art sets are not included in the repository.
The native 480×360 game render provides the semi-pixelated delivery. Generative
art can drift; inspect door positions, furniture edges and actor occlusion before
accepting a future replacement. Do not change geometry signatures to hide drift.

To pack another complete reviewed set, export fresh schemas into ignored
`.napplet-space/rerender/`, create an untracked JSON array of `{name,path,prompt}`
with names `room-id-camera-id`, and run:

```sh
node scripts/pack-room-renders.mjs <reviewed-shot-index.json>
pnpm test:levels
pnpm build
```

The packing script converts local reviewed files with ffmpeg, refreshes links and
hashes, and removes superseded backgrounds only after all views are packed. It
does not generate images or call an image service. Review the hosted result and
commit JSON, JPEGs, schemas, prompts and the manifest together.