Live-session findings, three in one: (1) Gauntlet walls are 2 tiles thick —
the outer row's camera-far neighbor is the inner WALL, so its flag stayed 0
and it stood full-height behind the collapsed inner stub; the flag now scans
up to 3 tiles through consecutive KNOWN walls (unknown stops the scan — no
assumptions past the info boundary), and any newly-learned tile refreshes the
wall chain behind it. (2) Corner pillars at the screen-bottom junction never
collapsed — their interior floor sits DIAGONALLY behind, which no straight
scan line reaches; the combined diagonal joins the scan set. (3) Per-fragment
smoothstep falloff carved organic notches (the 'sphere' read) — the cut now
computes per-instance in the vertex shader from the box's tile center and
quantizes at 0.5: every wall box is either full or stub, X2-crisp.
Verified via locomotion_cutaway autopilot capture: south runs read as clean
stub rows, corners collapse, camera-far walls stand.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
jpmschweitzerandClaude Fable 5
6372537f2f