Offender: wrld_f0d1ad30-bf56-444c-ba62-b2582e49767f

Replication: use clamped non-linear L1 solver (https://grahamhazel.com/blog/2017/12/22/converting-sh-radiance-to-irradiance/), navigate out and to the right, almost all the way down the corridor, until you see sudden change in pixel colours. I suspect values close to negative epsilon for W component.

Sample: avtr_1e0dc4a6-dadf-4001-b4f2-30d63c257372, avtr_6bfae6c2-3842-4d57-b216-74f8404c3cfa, avtr_7e61d1a4-9b1e-46c0-b8d0-4a04369d42fc etc.

Why it happens: SHAx.w is non-positive, while SHAx.xyz reports uniform directions. This results in flipped, but also irrelevant value, since irradiance can't be negative. Moreover, SHAx.xyz can only be non-zero when SHAx.w is non-zero (when there's no light, there's no dominating direction, checks out).

Current solution is to abs() the SHAx.w triplet, but that results in running extra operations in the shader (avtr_29f036dc-1fe6-459c-85e0-09b9a57c392a, leftmost ball). Instead, implement a sanity check that runs at bake-time or at load-time and clips values to [0, ∞).