How the tolerance is loosened when the cap binds, and how far.
A selection over maxPatches is not truncated at the level that overflowed.
It is retaken with cellPixels multiplied by this, and again, up to
COARSEN_STEPS times, until it fits; only past the loosest tolerance is the
tree cut off where it stands and reported saturated.
The breadth-first cut is the wrong degradation for a camera on the ground, and the Mars landing is where it showed. It takes the deepest level from everybody at once, and everybody includes the node the eye is standing in: at 3200×1800 display pixels the hover wants 1,583 patches against 1,280, and the cut tree stops at level 13 with the camera a meter off ground whose floor is 16 — three levels, eight times the cell, and it arrives in steps, 16→15→14→13, as the lens narrows through the settle. Loosened to 1.5× the same eye wants 1,011 and the ground underfoot is still level 16, because a node the eye is inside is at distance zero and refines at any tolerance. Every ring further out moves in by the step instead — a fraction of a level spread over the whole disk, which the morph carries, rather than a level taken from the one place the picture is looking.
Crack-free by construction: a coarsened walk is the same predicate with a
different constant, so the handover argument in regionSpacing holds
unchanged. The step is 1.5 rather than 2 because 2 is a whole level and a
whole level arriving in one frame is the pop this exists to remove; three
steps reach 3.4×, a level and three quarters, which covers a DPR 2 window at
the telephoto end and stops short of pretending a racked zoom fits. The
retry is a second walk on the frames it fires — the first walk stops at the
overflow and skips the balance pass, so it costs less than a full one — and
a caller that hands back coarsening pays it on the frame the cap first
binds rather than on every frame after.
It buys nothing where the tree is balance-limited, and the cut is the
answer there as before. DEFAULT_CELL_PIXELS says the graded tree has a
floor of its own; at the flight lens over 3200×1800, a meter off the
ground, the 2:1 rings are what set the count and the predicate is not what
binds — loosening 16 px to 54 px cuts under 5% of the patches across
thirty-six eyes on Mars and on Earth. The ladder is climbed and reports
saturated at the top. Where the predicate binds, the same eyes through
the landing's 34° lens, one step is the 1,583 → 1,011 above.