[Bug] IBL diffuse term uses multi-bounce-compensated reflectance instead of single-bounce #1296

Context

Follow-up from PR #1292 (material shading fixes) — see review comment: #1292 (comment)

Issue

In computeIBLContribution (Sources/UntoldEngine/Shaders/LightShader.metal), the diffuse term is attenuated by the full multi-bounce-compensated reflected value returned by environmentReflectance (which includes the Kulla/Conty laterBounces energy-compensation factor), rather than the raw single-bounce reflectance (reflectedInOneBounce).

Physically, the energy "taken from" diffuse should be the single-bounce specular share at that viewing angle. The multi-bounce compensation exists to recover energy lost by the specular term's own approximation (accounting for light bouncing between a rough surface's microfacets before escaping) — it isn't a measure of how much the surface diverts away from diffuse scattering in the first place. Using the compensated value for both purposes conflates the two.

Why it hasn't caused a visible bug yet

• For non-metals, F0 = 0.04 keeps laterBounces close to 1, so the error is small.
• For metals, (1 - metallic) = 0 zeroes out the diffuse term entirely, so the discrepancy never shows up.

None of the new render tests in #1292 caught this because they don't target the specific case where it would matter: a non-metal with high roughness where laterBounces deviates furthest from 1.

Suggested fix

Use reflectedInOneBounce (already computed inside environmentReflectance) for the diffuse-energy-conservation term, and keep the fully compensated reflected value only for the specular contribution. This likely means exposing reflectedInOneBounce from environmentReflectance (e.g. as an additional return value or a second small helper) rather than just the final compensated color.

Acceptance criteria

• Diffuse attenuation in computeIBLContribution uses single-bounce reflectance, not the multi-bounce-compensated one.
• Existing EnvironmentReflectionShadingTests (especially testAWhiteSurfaceGivesBackAllTheLightWhateverItsFinish and testARoughColoredMetalStaysWithinItsColor) still pass.
• Consider adding a test for a rough, high-F0-adjacent non-metal (e.g. roughness 1.0, base color gray) to pin down the expected diffuse/specular split precisely enough to catch regressions here.