The carousel that builds something

2026-09-03. The landing page's stage read as three screenshots. A visitor could not tell there was a video in it, and the two slides that moved were a slider bending a photograph of a cat. This is what was tried on the way to slides that start on a clean board and get somewhere by hand, including the two approaches that were wrong.

Three things were wrong, and only one of them was the carousel#

/guide/* was not served in dev at all. The docs site is a separate static build (scripts/build-guide.mjs → dist/guide/) that Pages hosts beside the app; the dev server knew nothing about it. A miss under /guide fell through to vite's HTML fallback, so /guide/img/signal-path-callout.jpg came back as index.html with a 200. An <img> holding a document paints as an empty box and the network log shows nothing wrong — a 404 would have said so on the first look. vite-plugin-guide.ts serves figures straight out of docs/img (which is where the builder copies them from, so a figure is right without a guide build having been run) and pages out of dist/guide when there is one.

Since superseded. The docs site is an Astro project now (../DEVELOPMENT.md › Docs site), and its dev server renders the markdown live under /guide/, so scripts/build-guide.mjs and vite-plugin-guide.ts are both gone. The fault this paragraph describes is what that dev server exists to make impossible.

Nothing on the stage moved on arrival. It played only after a Play chip that sat in a row of visually identical tab chips. It autoplays on scroll-into-view now and the chip is gone; the tabs are the only control, and they are the way in for every reader rather than the fallback for one kind.

The header was a clip. 86K fetched before anybody had scrolled anywhere, at 55% opacity under a scrim closing to the page colour — about a fifth of it reached the screen. It is the demo flagged hero in demos.json as a still now, at 8K, and the flag is also what ogimage.mjs grounds the link preview in, so the page and the card standing in for it are the same picture.

The wash had to change with it. The old one ran top-to-bottom and held the whole width near 55%: a moving fifth of a picture reads where a still one is mud. It is the link preview's angled wash instead — opaque down the type column, clearing past it — and where it clears is the measure, not a fraction of the window. Written as a percentage it cleared at 336px on a 390px phone, straight through the middle of the sentence under the buttons. calc(50% + var(--measure) / 2) lands past the right edge below the measure, so a phone simply gets an even wash, which is the right answer where there is no "beside the type".

Two wrong turns on the slides#

The cat. The slides ran on the bundled photograph through a tape path, chosen because damage is legible on a cat. That is true and it is also the problem: a stranger watching a slider bend a photograph has been shown that the program has sliders, and nothing else.

Winding one row back. So they were moved onto the looks the gallery lists, with one row of each wound back through ?set= (which layers over ?p=) so a drag could arrive at the demo. Worse. The board was ninety per cent of the way there before the clip started and the drag took the credit. It also does not work: chroma gain on that board is inside a feedback loop, so arriving at 15.79x is not arriving at the look — the loop needs laps to breed it, and the clip cut away on a black frame with a slider reading 15.79x, which says the control does nothing.

What ships, and what the probes said#

params: {} — a bare load, which is the app's own near-blank board: stock controls, colour bars on both sources. Every value that ends up on screen was put there by the hand in the frame. The lead slide raises three rows and finishes on a field of rainbow that was not there sixteen seconds earlier; "This look · 3 off stock" in the frame is the proof.

Which three took a probe rather than a guess (?set= from clean, warm 300, shoot the window — one browser per spec, because a single Firefox does not survive a long WebGPU batch and a lost device paints a "rebuilding" toast over the shot):

• Colour bars hide chroma gain. They are already saturated, so 12x clips them back to almost the same bars. The photograph was in there for a reason.
• fbMix and fbGain multiply. A loop under unity decays to a dark wash however far the fader goes: fbMix:0.95 alone is a dim green smear. What makes a picture breed rather than fade is the round trip sitting just over 1, which is what the group's own readout says on screen while the hand is on it.
• gain is unreachable. It is fine: true, so it lives behind the group's "N fine tweaks" disclosure and there is no row to reach for.
• auto-iris hunt is the row. Non-fine, fourth in the group, and the better demonstration: winding up a servo that is metering its own output shows why the runaway happens. fbMix:0.9, fbIris:0.6, cfbMix:0.95 is a white-hot core with rainbow ripples across the frame.

Recorder bugs this turned up#

• A map box is aimed at its label now, not its group. The two loop pills carry the dotted band they ride on inside the same <g>, so the group's box is 271x38 of mostly empty band and its centre lands on a bare <path>. The press was a silent no-op that only showed up as the rest of the timeline being wrong.
• A stage opened from the map unfolds its first group. The receiver's is SYNC, so a timeline that pressed RECEIVER and reached for a deflection row found none. Seed the group instead; walking the map is the third slide's job.
• A scrollTo on a panel with nothing to scroll is a still beat, not a crash. scrollPlan returns null there and the caller read .from off it. The same timeline is recorded at two widths and a row below a phone's fold is already on screen at 1112px.
• stillAt per slide. The poster was fixed at 0.55 of the timeline, which suits a walk and is wrong for a build: the middle is a picture on the way to the one the slide is about, and it is the whole of what a reduced-motion reader sees.

Left open#

• The lead slide is 914K, the heaviest file the page has ever had. Per-slide crf was tried and removed: 36/38/40/42 is 871/782/695/618K off the same frames, and smooth gradient is where a raised quantizer bands. Length is the lever, and the timeline has about a second of slack in its holds.
• The stage no longer plays anything for a reader on prefers-reduced-motion, by instruction. They get the stills and the tabs and step through by hand. The old Play chip existed for exactly them.
• showcase is down to one demo, since only the third slide names a look. The flag is still the right shape; it is just barely used.
• Recording is ~90s a take and there are six. It wants a git worktree copy with its own vite (cacheDir and port of its own, node_modules symlinked, run node_modules/.bin/vite directly) — an HMR reload from another agent editing src/ mid-run resets the engine under the recording.