# FAILURE_MODES
This crate treats failure modes as first-class evidence, not cleanup notes.
## Signal Limits
- Weak residuals can make DSFB advisory rather than decisive.
- Strong host heuristics can tie or beat DSFB on some captures.
- The checked-in Unreal sample is intentionally retained even though its current Demo A classification is `heuristic_favorable`.
- Missing ROI or disocclusion exports reduce interpretability.
- Motion-vector convention errors are fatal because silent normalization mistakes would poison the evidence.
- The minimal crate-local Unreal sample currently uses metadata-derived motion vectors and replay normals because the editor-side Linux capture path did not expose numerically stable dense velocity and unit-normal exports.
- The minimal crate-local Unreal sample uses `monotonic_visualized_depth`, not overclaimed linear depth.
## Content Limits
- Transparency
- particles
- UI
- post effects
- specular-only motion
- poorly observed disocclusions
These regimes can violate the current assumptions about depth, normals, and reprojection.
## Deployment Limits
- GPU timing in the bundle is environment-specific.
- Demo B remains an allocation proxy unless a renderer-integrated sample budget path is exported.
- This crate does not prove universal renderer integration.
## Language Discipline
Allowed posture:
- supervisory layer
- trust / admissibility / intervention signals
- engine-native empirical replay
- evidence consistent with reduced temporal artifact risk
- promising insertion point for Phase I / pilot work
Not allowed posture:
- universal proof
- solved rendering
- renderer replacement
- production-ready without engine integration evidence