Add 3DGS epiphany roadmap plan

Author: AdaWorldAPI

.claude/plans/3DGS-epiphany-roadmap-plan.md

@@ -0,0 +1,172 @@
+# 3DGS Epiphany Roadmap Plan — lance-graph
+
+## Goal
+
+Collect high-leverage follow-up ideas discovered during the 3DGS planning pass.
+
+This file is intentionally exploratory. Promote items into separate plans when they become implementation-ready.
+
+## 1. CHLOD: Certified Hierarchical Level of Detail
+
+Replace heuristic-only LOD with certified decisions.
+
+```text
+HLOD + error certificate + query relevance = CHLOD
+```
+
+Key idea:
+
+- parent tile can be accepted only if error certificate passes
+- child refinement is triggered by certified error, not just geometric error
+- stored certificates allow repeatable render decisions
+
+Potential crate:
+
+```text
+crates/geo-chlod
+```
+
+## 2. 3DGS change detection with error envelopes
+
+Use certificates to distinguish real scene change from approximation artifact.
+
+Pipeline:
+
+```text
+scan A
+scan B
+  -> align / normalize
+  -> compare splat blocks and features
+  -> subtract known error envelopes
+  -> emit changed regions above confidence threshold
+```
+
+Useful for:
+
+- construction progress
+- infrastructure inspection
+- forestry
+- archaeology
+- digital twins
+
+## 3. Query-aware foveated traversal
+
+Classic foveated rendering cares about camera gaze.
+
+Query-aware traversal also cares about semantic focus:
+
+```text
+camera focus + query focus + certificate budget -> tile priority
+```
+
+Example:
+
+- user selects pipe network
+- renderer refines splats near pipe assets
+- unrelated tiles stay coarse even if visually nearby
+
+## 4. ArcGIS-compatible output mode
+
+Goal: export enough 3D Tiles / FeatureServer-like resources to interoperate with GIS tools.
+
+Do not reimplement full ArcGIS Server first.
+
+Start with:
+
+- static 3D Tiles export
+- GeoJSON / vector overlay export
+- simple FeatureServer-like query endpoint for local data
+
+## 5. Lance-backed implicit subtree server
+
+Use Lance tables as the source of truth for implicit tiling.
+
+Endpoint generates subtree files from table rows and availability bitstreams.
+
+Advantages:
+
+- no giant JSON tile trees
+- fast updates
+- compact change sets
+- queryable availability
+
+## 6. 3DGS tile compiler
+
+Build a compiler:
+
+```text
+raw splats / glTF / SPZ / photogrammetry output
+        ->
+local frames
+        ->
+HHTL partitioning
+        ->
+3D Tiles + Lance sidecar
+        ->
+certificates
+```
+
+This is likely the first impressive demo path.
+
+## 7. Scientific/medical sibling path
+
+The same certified 3DGS machinery may work for volumetric ultrasound / medical field rendering.
+
+Shared concepts:
+
+- splat-like anisotropic kernels
+- covariance projection
+- certified uncertainty
+- queryable metadata
+- CPU-first previews
+
+Keep this as a sibling idea, not in the core geospatial path.
+
+## 8. Tile intelligence as graph
+
+Represent the world as:
+
+```text
+Tile -> Content -> SplatBlock -> Feature -> Asset -> Observation -> Certificate
+```
+
+This makes visual data queryable and auditable.
+
+The viewer becomes a spatial intelligence engine, not just a renderer.
+
+## 9. External compatibility envelope, internal columnar core
+
+Recommended strategy:
+
+```text
+external envelope: 3D Tiles / glTF / ArcGIS-compatible outputs
+internal core: Lance / Arrow / ndarray SIMD / certificates
+```
+
+This avoids fighting standards while still gaining performance and auditability.
+
+## 10. First demo target
+
+Do not start with the whole planet.
+
+Start with:
+
+```text
+one small 3DGS scene
+one tileset.json
+one Lance sidecar
+one camera traversal request
+one certified decision report
+one CPU projection preview
+```
+
+Then grow from there.
+
+## Promotion criteria
+
+Promote an epiphany into its own plan when:
+
+- it has a clear crate owner
+- it has a testable input/output boundary
+- it can be built without blocking the base 3D Tiles runtime
+- it improves either compatibility, performance, or auditability