First note on the new engine. The frontmatter for this post is validated against the Note page-type schema at build time—including that in-reply-to up there. If I typo a field name or skip pubDate, the build fails and ddc diagnostics tells me exactly which file did it.

Show notes for the trailer. The audio is a synthetic placeholder—a slow pad rather than a voice—so the pipeline has something real to master before the first episode is written. scripts/podcast-encode.rs produced it: two-pass loudnorm to -16 LUFS and -1.5 dBTP, 128 kbps stereo, ID3v2.3 tags, chapters, and an embedded cover. It then ran ffprobe over the result and wrote the duration and byte count back into the frontmatter above, so the feed has everything it needs from content/ alone.

Note the show = "podcast" field: that's a typed @link(@Show) reference. If it pointed at a page that isn't a Show, the build would fail.

A field of stars that twinkle, a comet on a slow ellipse, and a wordmark. No game engine: it is Rust compiled to WebAssembly talking to a 2D canvas through wasm-bindgen and web-sys, which makes it the smallest thing that still exercises the whole pipeline.

The whole artifact is about 58 KB of WebAssembly, 25 KB of that over the wire, plus 17 KB of glue. It streams in only when you press play, so a gallery full of toons costs nothing until you ask for one.

Two things about the build are worth knowing. dodeca content-hashes every file in static/, and wasm-bindgen resolves its own .wasm sibling through a string literal that dodeca does not rewrite, so scripts/sketch-manifest.rs publishes a small manifest.json mapping the original names to the hashed ones; the embed shell reads that instead of guessing. And the shell unhides the canvas before instantiating the module, because a hidden element has no layout, and a sketch that measures itself at startup would size its canvas to nothing.

I wrote the sizing loop wrong the first time in a way worth recording: a <canvas> with no CSS width of its own takes its layout size from its width attribute, so measuring the canvas to set the canvas doubles the box every frame until the tab runs out of memory. It measured the shell instead.