Add 3DGS PR-X12 cross-pollination capstone

Author: AdaWorldAPI

.claude/plans/3DGS-PRX12-crosspollination-capstone.md

@@ -0,0 +1,283 @@
+# 3DGS / PR-X12 Cross-Pollination Capstone — lance-graph
+
+## Purpose
+
+Capture the wider expansion discovered by connecting the Cesium / maps / 3DGS exploration with the older PR-X12 codec canon.
+
+The key realization:
+
+```text
+Cesium-style 3D tile traversal
+PR-X12 Skip/Merge/Delta/Escape block grammar
+GGUF/safetensors tensor compression
+HHTL datalake traversal
+SplatShaderBlas / BLASGraph execution
+```
+
+are not separate tracks. They are different faces of the same substrate:
+
+```text
+hierarchical local fields
+  -> approximate block reconstruction
+  -> certified skip / refine / hydrate / render / decode decisions
+```
+
+## Capstone thesis
+
+```text
+The maps path proves the traversal.
+The codec path proves the block grammar.
+The GGUF/safetensors path proves tensor deployment value.
+The datalake path proves generality.
+```
+
+Together they define a universal execution pattern:
+
+```text
+structured source
+  -> hierarchical block graph
+  -> compressed/certified local kernels
+  -> query/render/decode/hydrate schedule
+  -> exact work only where the certificate demands it
+```
+
+## Cross-pollination map
+
+| Source line | Core idea | What it contributes |
+|---|---|---|
+| Cesium / 3D Tiles | HLOD, tile hierarchy, implicit tiling, screen-space error | traversal grammar |
+| ArcGIS | enterprise geospatial service model, FeatureServer/MapServer/Scene layers | data-source compatibility |
+| 3DGS | anisotropic kernels, EWA projection, scene anchors | local-field reconstruction |
+| PR-X12 x265 | Skip/Merge/Delta/Escape, BLAS/GEMM codec loops | compressed block grammar |
+| PR-X12 x266 | EWA splat as `Basis<T>` | basis-swappable reconstruction |
+| PR-X12 GGUF | tensor CTUs, activation-aware RDO, decode-during-GEMM | model deployment path |
+| PR-X12 anti-neural | frozen lookup tables instead of NN hot loops | deterministic low-latency runtime |
+| SplatShaderBlas | tiered bitpacked / palette / 3DGS execution | unified execution naming |
+| Datalake HHTL | certified block pruning and hydration | non-visual generalization |
+
+## General substrate shape
+
+```text
+Source domain
+  -> adapter
+  -> block hierarchy
+  -> local kernel summary
+  -> certificate
+  -> action
+```
+
+Actions should stay common:
+
+```text
+skip
+keep approximate
+refine
+hydrate metadata
+hydrate exact payload
+render
+decode
+fallback
+reject
+```
+
+## Maps as the first visible body
+
+The maps path starts with:
+
+```text
+3D Tiles / Cesium / ArcGIS source
+  -> tiles / contents / features / splat blocks
+  -> Lance/Arrow sidecar
+  -> ndarray 3DGS EWA/SYRK kernels
+  -> certified tile decision
+```
+
+What it enables:
+
+```text
+offline or local-first 3D worlds
+queryable digital twins
+certified tile skipping/refinement
+3DGS scene overlays
+feature-aware rendering
+change detection with error envelopes
+```
+
+But the same structure extends beyond maps.
+
+## Datalake as the first non-visual body
+
+The datalake path maps tiles to fragments:
+
+```text
+tile                 -> fragment / row group / page
+bounding volume      -> stats / bloom / centroid / schema domain
+screen-space error   -> query relevance / confidence error
+refine               -> hydrate deeper metadata or exact rows
+render               -> execute query / scan rows / return vectors
+```
+
+What it enables:
+
+```text
+metadata-only query planning
+certified shard skipping
+vector + SQL + graph hybrid traversal
+EXPLAIN output with skip/refine reasons
+partial hydration of exact rows only when needed
+```
+
+## GGUF / safetensors as the tensor body
+
+The tensor-container path maps codec CTUs to model-weight blocks:
+
+```text
+video CTU
+  -> tensor CTU
+  -> activation-aware RDO
+  -> basin/codebook pointer
+  -> residual tail
+  -> decode-during-GEMM
+```
+
+What it enables:
+
+```text
+GGUF Q_PRX12 experimental quant type
+safetensors sidecar compression
+Lance tensor chunk storage
+partial model loading
+cross-layer / cross-head Merge
+model-weight lakehouse search
+smaller distribution artifacts
+```
+
+## Anti-neural rule
+
+The runtime discipline should be explicit:
+
+```text
+NNs may train tables.
+NNs do not run in the hot loop.
+```
+
+Preferred inner-loop primitives:
+
+```text
+lookup table
+popcount
+palette distance
+BLAS / GEMM / SYRK
+EWA sandwich
+rANS
+HHTL traversal
+```
+
+This is what keeps the substrate deterministic, explainable, portable, and cheap.
+
+## Unifying graph model
+
+Represent all domains with the same graph skeleton:
+
+```text
+Source
+  -> Frame / Tensor / Dataset / Tileset
+  -> Block
+  -> KernelSummary
+  -> Certificate
+  -> Decision
+  -> Output / Payload
+```
+
+Examples:
+
+```text
+Tileset -> Tile -> SplatBlock -> Certificate -> RenderDecision
+Dataset -> Fragment -> RowGroup -> Certificate -> HydrationDecision
+GGUF -> Tensor -> TensorBlock -> Certificate -> DecodeDecision
+Safetensors -> TensorShard -> TensorBlock -> Certificate -> LanceChunk
+Ultrasound -> Frame -> PSFBlock -> Certificate -> FusionDecision
+```
+
+## What this enables by domain
+
+```text
+geospatial:
+  certified 3D maps, digital twins, ArcGIS/Cesium interop
+
+datalake:
+  HHTL query planning, explainable skipping, vector/SQL/graph fusion
+
+LLM deployment:
+  GGUF/safetensors compression, decode-during-GEMM, partial loading
+
+RAG:
+  chunk-family traversal with provenance and retrieval confidence
+
+observability:
+  incident-root traversal over logs/traces/metrics/deploy events
+
+robotics:
+  sensor-field updates with map-change certificates
+
+ultrasound:
+  RF/IQ/Doppler frame fusion into certified splat volumes
+
+genetics:
+  motif/gene/pathway block traversal with uncertainty envelopes
+
+neuronal/cognitive fields:
+  activation/edge/time block comparison with 4x4 carrier summaries
+```
+
+## Implementation priority
+
+Do not implement every domain. Stabilize the substrate through two hard bodies:
+
+```text
+1. 3DGS geospatial tile runtime
+2. HHTL datalake traversal
+```
+
+Then attach tensor-container work:
+
+```text
+3. PR-X12 GGUF/safetensors tensor block prototype
+4. Lance/Arrow tensor chunk storage
+5. decode-during-GEMM microbench
+```
+
+Exploratory domains remain adapters until the core proves itself.
+
+## Required references across plan set
+
+This capstone should be kept linked with:
+
+```text
+3DGS-Cesium-feature-mapping-plan.md
+3DGS-HHTL-datalake-traversal-plan.md
+3DGS-SplatShaderBlas-BLASGraph-crosspollination-plan.md
+3DGS-blast-radius-application-map.md
+3DGS-domain-adapter-strategy-plan.md
+3DGS-certified-query-render-plan.md
+```
+
+The ndarray sibling capstone is:
+
+```text
+PR-X12-tensor-container-expansion-capstone.md
+```
+
+## Acceptance criteria
+
+- The plan set no longer frames 3DGS as only geospatial.
+- PR-X12, GGUF/safetensors, and datalake traversal are explicitly connected.
+- SplatShaderBlas / BLASGraph is treated as the execution bridge.
+- Domain adapters remain separate from the core substrate.
+- The first implementation path remains small and testable.
+
+## Wall sentence
+
+```text
+Cesium gave us the world-tree, PR-X12 gave us the block grammar, GGUF/safetensors give us tensor deployment, and HHTL turns all of it into certified traversal.
+```