Asserting that something moved — the feedback fixtures

A baseline cannot say a picture moved. It says a picture is the picture it was last time, which is the opposite question. When the thing under test is a motion — ADR-0048’s transformed feedback, where an accumulation is resampled through an affine every frame — the guard has to compare frames of one run against each other, and core/tests/feedback.rs is where that shape lives. The next author of a motion test should copy its habits rather than reinvent them.

Make the figure static, so the only thing that can move is the thing under test. Its motion fixtures are all one Maurer rose at spin = 0, parked off-centre with pan_x. If the scene animated, a displacement measurement could not attribute what it found.

Measure in the coordinate the claim is about. These transforms are radial and tangential, so the guards convert every lit pixel to (radius, angle) about the frame centre in pixels and compare extents: fb_zoom must grow the radial span faster than the angular one, fb_rotate the reverse. Two details that cost real time to rediscover — an angular span has to be computed as 2π minus the widest empty gap, or a set straddling the atan2 branch cut reads as the whole circle; and “lit” is a fraction of that frame’s own peak, never an absolute byte, because the tail of a decaying trail is dim by construction and how dim depends on the tonemap, the palette and the adapter.

Assert a ratio, not a number. Every claim there is one run against itself — late frame over early frame, or one axis’ growth against the other’s. There is no pixel count to re-tune when a shader changes by a byte, and no threshold that encodes the capture size.

Say what a broken version would look like, then check the guard sees it. The shear guard rotates the figure into a closed ring and asserts its pixel bounding box is square, because a ring about the frame centre is round on screen whatever the target’s shape. It was verified by breaking it: with the transform’s aspect forced to 1.0 the box goes from 45x46 to 44x71 at the file’s 100x160 portrait target. Which is also why that target is portrait — see the file’s own header, and ADR-0047.

Two more habits from the same file, for the deposit rather than the motion:

• A convergence claim needs both ends. “The frame stopped changing” is also what a stage that never accumulated would say, so the additive guard asserts that the first window moved a lot (97 bytes) and the last did not (0) — and compares them as a ratio, since an unbounded accumulation moves the late window by about what it moved the early one.
• A one-key comparison is only controlled if the key does something. Two of these fixtures measured 0.000000 apart before they were tuned, for two different reasons that both look like a passing test: max(cur, prev * fade) over a stationary figure is exactly cur, and so is a trails stage whose tail is shorter than the scene’s own fade. Where a guard’s premise is “this key changes the picture”, assert that premise before asserting anything on top of it.