Flaky: Test Drivers (selenium_headless, chunky_png) — one input row renders differently ~1 run in 5

[pftg]

Split out of #284 so it does not sit in the release path.

BrowserScreenshotTest#test_compare_crops_only_when_other_part_is_not_working fails
intermittently on one matrix cell, Test Drivers (selenium_headless, chunky_png):

the change spans 4378 of 16800 px (26.06% of the 200x84 image)
  changed region: [0,19,199,41] (left,top,right,bottom edges)
  max color distance: 50.9
  judged against: tolerance 0.13, color_distance_limit 40

What is established

• Base rate: 4 success / 1 failure over the last 12 Test runs on that cell. Every
  other Drivers cell is 100%.
• Not environment drift. The failing job (97473161378) and a passing job at the same
  commit (97450518382) ran on the identical runner image 20260816.277.1, same
  runner version, same Ruby.
• Not antialiasing. The changed region is 200x22 = 4400 px against 4378 px changed —
  the entire first input row and nothing else — at max colour distance 50.9. That is a
  materially different render of #first-field, not sub-pixel drift.
• The fixture (test/fixtures/app/index-without-img.html) loads no remote resources;
  the font stack is local with fallbacks, so it is not the usual web-font race.

What is not established

The actual cause. A plausible shape is leftover focus/selection state on #first-field
from the test that does fill_in "First Field:", but that is unverified — and it
cannot be verified from CI artifacts until #284 is fixed, because the failure's own
evidence is overwritten before upload.

Blocked on #284

Fix the shared screenshot name first, then let this flake reproduce once with real
artifacts. Debugging it before that means guessing.

Also worth deciding

fail-fast turns this single flaky cell into three red cells (two siblings report
cancelled), so a run reports far more damage than occurred. Same mechanism hid the JRuby
lane for 15 runs (#283).

Scheduling

Not a 2.0.0 blocker: it predates the release, and chunky_png is deprecated in 2.0 and
removed in 2.1 — so this cell disappears on its own in 2.1. Filed against 2.1.0 so the
deletion is a deliberate outcome rather than an accidental one, and so nobody mistakes the
removal for a fix.

[pftg]

Unblocked — #284 is fixed (#290, 7809756), so this can finally be diagnosed

Until now a failing run destroyed its own evidence: three tests shared index-without-img-cropped, so whichever one failed had its differing PNG overwritten by the next test's matching capture before artifacts uploaded. The artifact from the one run we caught contained three identical files (5fccb56f…), which is why every previous attempt to debug this dead-ended.

They now have distinct names, so the next occurrence will upload a real .png vs .base.png pair.

What to do on the next failure

1. Download screenshots-selenium_headless-chunky_png-diffs from the failing run.
2. Compare index-without-img-cropped.png against .base.png — they should now genuinely differ.
3. The .diff.png / .base.diff.png annotations show where, which is the thing we have never had.

Current standing, restated so it is not re-derived

• Base rate: 4 success / 1 failure over 12 runs on that one cell; every other Drivers cell 100%.
• Not environment drift: the failing and a passing job at the same commit ran on the identical runner image 20260816.277.1, same runner version, same Ruby.
• Not antialiasing: 200×22 = 4400 px region with 4378 px changed — the whole first input row — at max colour distance 50.9.
• Cause still not established. The plausible shape is leftover focus/selection state on #first-field from the test that does fill_in "First Field:", but that is unverified and should not be treated as the answer.

Note on the milestone

Still 2.1.0, and still worth fixing rather than waiting for the backend to be deleted. chunky_png goes away in 2.1, which would make this cell vanish — but a flake that disappears because its host was removed was never diagnosed, and the same render instability could exist on the vips path where nobody has looked.