Scene floorplan — visual regression without pixels

A structural snapshot of what a scene actually draws: one record per mesh the camera can see, at its true world position and size. Comparing two of these answers "did the picture change, and where" without ever rasterising anything.

import { sceneFloorplan, floorplanDiff } from 'tosijs-3d-ensemble'

const before = sceneFloorplan(scene)
// …rebuild, edit, load a different file…
const moved = floorplanDiff(before, sceneFloorplan(scene))
if (moved.length) console.warn(moved)

Why not a screenshot

Because a screenshot compares the wrong thing. Antialiasing, font hinting, GPU driver and a shader that compiled differently all move pixels without moving anything an author cares about, so an image diff reports 3.2% of pixels changed and leaves you to work out whether that is a regression or a Tuesday. The signal is real but it arrives wrapped in noise you cannot turn off.

A structural diff has the opposite properties. A rectangle in the same place, or very nearly, versus absent — that is non-flaky signal. Tolerance is a number you set rather than a rendering artefact you fight, and a failure names the thing:

ground: moved (0, 0, 0) → (0, -12, 0)
watchtower: GONE

rather than a percentage and a heat map.

This is Tonio's observation about tosijs-floorplan applied to a renderer: floorplan turns a DOM page into {caption, bounds} records and compares those instead of images, and a scene graph already IS that data — Babylon keeps an authoritative draw list because it has to. The two are the same idea over different substrates, which is the argument in tosijs-ui#142.

What it deliberately does NOT capture

Colour, material, lighting, shadow. This says a thing is there, at that size, in that place — the layout claim, not the appearance one. Appearance still wants a human, or a vision model looking at a picture; the point is that LAYOUT no longer has to borrow the picture's flakiness to get checked.

activeOnly (the default) reads the post-frustum-culling draw list, so an object behind the camera is legitimately absent. That is usually what you want — "what does this look like" — but it makes the snapshot camera-dependent, so compare snapshots taken from the same viewpoint or pass activeOnly: false.