render-defect-forensics

When to use. Any "the image looks wrong" bug. Why: predictions about images are usually wrong — hand camera-cone math mispredicted framing every time; one artifact survived six rounds of plausible-but-wrong theories until a single ray cast named the true object.

Procedure.

1. Reproduce with a cheap targeted render (one view, low samples). Composition questions are answered ONLY by rendering.
2. Two failed fixes → stop theorizing: cast a ray from the camera through the offending pixel and print what it hits. Fix the named object at its root (ours: a floor slab 20 mm high since four phases earlier).
3. Know the physics traps before inventing new ones: emission meshes block their own light source; small emitters need glossy/transmission ray visibility off; filmic tonemaps eat EV lifts (measure landed exposure and correct once, don't stack fixed lifts); smooth-shaded slab rims fake light leaks (flat-shade architecture); crushed shadows usually mean the sun:sky ratio, not geometry.
4. A/B render before keeping any perceptual change, and check for NEW artifact classes the change can introduce (per-channel sharpening ⇒ chromatic fringing; fix: luminance-only).
5. Done when: the fixed still is rendered and inspected (not asserted), and a mechanical probe or test now guards the defect class.

Knowledge: FIELD-KNOWLEDGE-NOTHOMEWRECKER.md §F3 (incl. USD-importer flattening, stamp-based matching, exposure targeting, de-tiling and albedo-grunge recipes).