TODO

cleanup:
- guided_extend vs guided_create_path?
- vertex area measure vs dwp
- use throughput or not? we are evaling the bsdf..
- don't refit vi==1 bvh but build from scratch with real bounds immediately
- only need one lbvh for first bounce, probably no extra index array

fixes:
- knn needs to consider first bounce gaussian!
- environment maps
- check index matched ior

bugs:
- pdf cache invalidation on edges (don't copy paths for mlt?)
- jacobian for vmlt->kmlt and transmit/reflect

vol:
- light sampling should work on edge->vol and vertex->interior (dedup)
- requires some refactoring of medium interface (store flags with
  homo/hete/emissive accessible without callback)
- nee + culling
- refactor interior and vol such that mu_t mu_s mu_e are on vertex.shading
  and ior shader are on vol struct

pathspace:
- clean up propagate vs. update_throughput (propagate with mutation semantics etc
  seems like the place to leave throughput + pdf alone)
- remove consistency checks after all calls to path_project() with mutation
  mode (now done inside)

guiding:
- one bvh per regex prefix (bounce, reflection mode)
- fill distance sampling cdf from eye also in neighbour tiles
- specular or otherwise requires to sometimes sample by bsdf instead
  - need to keep regex prefix useful
- when to cluster next vertices together, when to separate?

hmc:
- multi step size sampling
- anisotropic step size sampling

regression:
- fix scenes for new geo format!
- sparse git commit hash only for big changes (in timing plot)
  also copy/pasteable

shaders:
- introduce multi-layer shader to test spec/glossy lobe interaction
- wire anisotropic roughness to the microfacet models that support it

pathspace:
- on-demand pdf/pdf_adj cache system for efficiency!

cleanup:
- include optics to tools/

displacements:
- include into prims_geo
  - depend on triangulated obj so far!
  - init prim in alloc() callback in prims_geo.h (set extra bits, function in shell.h)
  -