"Decimated heightfield cliffs render as sawtooth teeth — snap the above-gap cluster"

A decimated heightfield mesher (one that reduces a dense height grid to a coarser triangle mesh) renders carved cliffs as an alternating sawtooth — the crest reads as a row of teeth instead of a clean lip. The two obvious fixes disappoint:

• Forcing consistent diagonals cleans the mid-wall but leaves the crest serrated.
• A majority-vote corner snap (snap each mixed corner to whichever height most of its samples agree on) does not fix it either.

There are two stacked artifacts, and only probing the actual per-corner numbers separates them:

1. Mid-wall: the "pick the smaller-delta diagonal" rule flips triangle orientation block-to-block when both diagonals carry similar (large) deltas across a steep wall. The fix is to force one diagonal on wall-relief blocks so the orientation stops flipping.

2. Crest: the cliff line runs diagonally through the cell lattice, so the mixed 4-cell corner column drifts about one cell every two to three rows. Corner averaging then smears a 45° ramp along several block rows of crest — that smear is the sawtooth fringe.

A majority vote fails on the crest specifically because on a one-to-two-cell-wide carved wall the transition cells carry intermediate heights: mixed corners are mostly 2–2 splits, so the majority never fires; and where it does fire it snaps in alternating directions, which is the sawtooth all over again.

Replace the corner average with a gap-cluster snap:

1. Sort the four corner height samples.
2. Find the largest gap between consecutive sorted samples.
3. If that gap exceeds wallSlope × cellSpan, take the mean of the cluster above the gap (not the mean of all four).

This is a uniform plateau bias: it dilates the lip by at most one cell and keeps the lip polyline level, which matches a plateau-first, faceted silhouette. To avoid flicker on near-threshold graded ramps, blend the snap in over the interval [gap, 2 × gap] so a corner right at the threshold can't oscillate between "snapped" and "averaged" frame to frame.

Averaging pulls the lip toward the mean of a diagonally-drifting corner column, which is exactly what smears the 45° ramp. Taking the mean of only the cluster above the largest gap keeps every crest corner pinned to the top surface, so the lip stays a level polyline instead of a graded ramp — the teeth collapse into a clean edge.

Corollary — faceted styles only. The snap suits a faceted (flat-shaded) mesh. On a smooth-shaded mesh the residual cluster wobble on graded ramps punches a new grazing-angle sawtooth into the skyline, so gate the snap behind a per-style flag rather than applying it everywhere.