The swept-sphere probes tilt their contact normal while rolling over a
thin board's edge, which faked slope top-outs, ceiling hand-offs and
bare-rim vaults mid-face. Confirmed via temporary exit-path logging
(stripped again) and fixed at the verified causes:
- RaycastWallFace: confirm "face flattened into a slope" with a thin,
strictly horizontal ray instead of trusting the tilted sphere normal
(a vertical board still reads steep; real ramps still top out).
- Rim judging: the full-capsule span probe is the arbiter — while it
still sees a steep face anywhere along the capsule, re-adopt that true
normal (breaking the tilt self-amplification feedback loop) and keep
climbing; only a persistent rim (wallTopOutConfirm debounce) vaults.
- WallContinuesAbove: band-sampled horizontal sweep so a grazed shelf
edge above never reads as a cleared rim.
- Bare rim: parks by default (upward motion clamped); holding up vaults
over the edge via MantleOverEdge, headroom-gated.
- ProbeWallSpan sweeps strictly horizontally; center probe always
rejects down-facing hits (edge grazes), wall->ceiling hand-off only
via the explicit head-gap check against clearly down-facing surfaces.
- Slope/ground top-outs additionally gated on a genuine stand
(HasStandableTop) so a board's own top edge no longer detaches.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
adminandClaude Fable 5
78ab6e2dd6