[transfix]

The flake

On macOS CI, cvcgl_ari_golden sometimes failed grid's "one more frame draws the same pixels (the UI has settled)" check on the first ctest attempt, with 8 px off by 1. --repeat until-pass:3 hid it, because the retry passed. It hit 2 of 4 macOS lanes on master (run 39610796467, 796517b), 2 of 4 on #760 (run 39612556385), and 1 of 4 here (run 39654691408).

Cause: the Apple Software Renderer, not the UI

• The first commit adds a fingerprint of what ImGui hands GL: draw lists (vertices, indices, commands, clip rects, texture ids) plus each texture's size and pixels. The settle line now reports it.
• On the failing macOS attempt (package-macos / libcvc-debug, attempt 1): 8 px differed by 1 (first at (39,61): 38,38,38 then 38,39,39), and ImGui drew the same UI both times (572 vertices, texture 512x128). The renderer drew identical input two ways.
• The differing pixels move around. The earlier runs had one column at x=117, y 230..237 (the "d" of the "Inner borders" title). This run had scattered pixels in x 39..567, y 61..107.
• It has only ever hit grid, the first document each process renders. In the same failing process, menus, scene, widgets and windows all settled at +-0. So it's a transient of the renderer's first frames that the test can't pin.
• Linux llvmpipe never shows it: 96 runs of grid, 32 at a time, with byte-identical frames and draw data.

The fix

The settle check keeps its +-0 tolerance. When the next frame differs and the UI fingerprint is identical, the later frame becomes the capture and the test renders another, up to 10 times. The log says how many frames that took and what the first difference was. A UI that changed still fails at once, and so does a rasterizer that never settles. docs/ARIADNE.md §12 describes this.

This does loosen one thing: a document may now take up to 10 extra frames to settle, but only when ImGui's draw data proves the UI didn't change. The pixel tolerance stays at +-0.

Verified locally

cvcpkg Mesa 25.3.6 llvmpipe on Xvfb, CVC_REQUIRE_RENDER=1:

• cvcgl_ari_golden 34/34; every document reports "ImGui drew the same UI both times".
• ctest ^cvcgl_ 72/72.
• A fault-injected scratch copy of the test (not committed) exercised each path:
  • +1 noise on one pixel in the first 2 or 4 renders: passes, "settled after 2 / 4 more frame(s)".
  • Noise on every render: fails, "had not settled after 10 more frame(s)".
  • A changed draw-list fingerprint: fails at once, "ImGui's draw lists changed".

Verified on CI

On 18a0734 (CI/CD run 39710264633), all 25 checks passed. cvcgl_ari_golden passed on the first attempt on all four macOS lanes.

• The "Ariadne golden-image numbers" step shows grid took 7 frames on libcvc-debug and libcvc-release-grpc-static, against the usual 6, and 6 on the other two.
• So the renderer noise fired on 2 of 4 lanes, as before, and one extra frame with the same UI absorbed it each time.
• That step prints only the frame count and the golden lines, so the "settled after 1 more frame(s)" note itself isn't in the job log.