docs: plans-alignment-triage — rank ChatGPT 48h corpus against ground truth (W1 #1)

.claude/knowledge/crate-registry.md

@@ -0,0 +1,126 @@
+# Crate registry + naming convention (3dgs-* family)
+
+> **AUTHORITATIVE policy** (a decision, not inspiration — unlike the plans triage).
+> Anticipates a battery of splat/tiles crates; this is the index + the rules that
+> keep them tracked and stop the battery from becoming empty shells.
+
+## Posture (read this first)
+
+**We do not *use* ArcGIS / Cesium.** No SDK, no service integration, no
+API/wire compatibility, no runtime dependency. The relationship is exactly two
+things, and nothing else:
+
+- **Reinvent** — build native CAM SoA from the data's first principles. This is
+  the goal.
+- **Reverse-engineer** — one-way extraction of data out of their static files,
+  rebuilt as CAM SoA. This is a *donkey-bridge*: transitional grounding,
+  acceptable **only** when a native refactor isn't yet feasible, retired the
+  moment it is.
+
+Using-ArcGIS vs reinventing-and-reverse-engineering is a completely different
+animal. We are the latter, never the former.
+
+**Optional grounding pillar (modesty clause).** We keep *one* optional,
+feature-gated `cesium` **reference** crate — never in the production hotpath,
+never a runtime dependency — whose only job is to run the real Cesium/ArcGIS
+path and emit reference output we diff against. This is the *runnable* form of
+the legacy pillar: it keeps "we're at parity / better" honest by **measuring**
+against a live oracle instead of asserting it from a capability matrix. It is
+the institutional humility against reinvention-hubris — the same "measure,
+don't assert" discipline we apply everywhere, turned on our own reinvention
+claim. It reads their JSON because it *is* the reference path; that's fine —
+it's dev/reference, walled off from production, so rule 3 still holds.
+
+## Hard rules
+
+1. **Single production crate family: `3dgs-*`.** No `cesium-*` *production*
+   family, no `ada-*` (deprecated). Naming is `3dgs-<role>` (`3dgs-tiles`,
+   `3dgs-ewa-syrk`, future `3dgs-render`, `3dgs-codec`, …). **One exception:** a
+   single *optional, feature-gated* `cesium` **reference** crate — the runnable
+   legacy pillar / parity oracle (see Posture's modesty clause), never in the
+   production hotpath, never a runtime dependency.
+
+2. **CAM SoA is the mandatory representation.** CAM SoA =
+   - **content-addressable** — addressed by `cam_pq` (CAM-PQ) codebook index, not by position;
+   - **structure-of-arrays** — `simd_soa::MultiLaneColumn` / `cognitive-shader-driver::bindspace` columns / lane carriers;
+   - **binary planes** — `AwarenessPlane16K` (16384-bit), q8 lanes, `#[repr(C)]`.
+
+   Whatever a source contains is **reverse-engineered** into CAM SoA — and
+   **ONLY** reverse-engineered: extract and rebuild, never depend on the source
+   format as a library, never round-trip it, never emit it. No AoS, no
+   source-native structs, no intermediate format survives past the import
+   boundary.
+
+   **Refactor > reverse-engineer.** The CAM SoA layout is *refactored natively*
+   from first principles, not derived from the source's structure.
+   Reverse-engineering is the donkey-bridge (see Posture) — a crutch, never the
+   design. Never let an imported format's structure calcify into ours.
+
+3. **NO JSON IN THE HOTPATH. EVER.** Restates the existing W1a invariant
+   (`no serde`, "no JSON in types", `#[repr(C)]` — see
+   `lance-graph-contract::splat`). JSON is permitted **only** at the cold,
+   one-time, read-only *import boundary* (reverse-engineering someone else's
+   file); it is rebuilt into CAM SoA immediately and **never enters the hotpath,
+   is never persisted, and is never emitted**.
+
+4. **Import adapters are read-once, one-way, output-CAM-SoA.** The 3D Tiles
+   reader (in `3dgs-tiles`) reads `tileset.json` + `KHR_gaussian_splatting` glTF
+   + `ESRI_crs` once, reverse-engineers into CAM SoA, and discards the source.
+   No `cesium-*` family; nothing emits or serves their wire format.
+
+5. **Rethink clause.** Our CAM SoA is canonical. An external format is
+   reconsidered **only** on falsifiable proof it is strictly better — not by
+   default, not for interop convenience. The optional `cesium` reference crate
+   (Posture / rule 1 exception) is the instrument that *produces* that proof.
+
+6. **Creation gate.** A crate exists **only with ≥1 real consumer** (≥2 for a
+   generic abstraction — the `certified-field-kernel-substrate` self-fence).
+   Until then it stays a **module inside an existing crate**. A `3dgs-*` crate
+   with no consumer is the code version of a markdown-only plan.
+
+## Registry (seed — verified first-party only)
+
+`status`: shipped / planned / module-until-earned. Only crates read first-party
+this session carry a description; the rest are the catalogue checklist below.
+
+| Crate / module | Repo | Family | Status | Purpose (verified) | Lane |
+|---|---|---|---|---|---|
+| `3dgs-tiles` | lance-graph | `3dgs-*` | planned | 3D Tiles reader. **Reverse-engineers** KHR_gaussian_splatting glTF + ESRI_crs → CAM SoA (donkey-bridge, transitional); JSON read-once at the cold boundary only | Geo |
+| `3dgs-ewa-syrk` | ndarray (`benches/`) | `3dgs-*` | planned | EWA-SYRK crossover bench (`project_batch` vs cblas-batched sandwich) | Kernel |
+| `cesium` (reference) | tbd | **optional/dev** | planned | runnable **legacy pillar / parity oracle** — runs the real Cesium/ArcGIS path, emits reference output to diff against. Feature-gated, never hotpath, never a runtime dep. The grounding + modesty instrument | tbd |
+| `splat3d` (module) | ndarray `src/hpc/` | — (pre-conv.) | shipped | CPU-SIMD forward renderer (`spd3`/`project`/`raster`/`sh`/`tile`/`gaussian`) | — |
+| `cam_pq` (module) | ndarray `src/hpc/` | — | shipped | 6×256 CAM-PQ codebook — the **CAM** in CAM SoA | — |
+| `simd_soa` (module) | ndarray `src/` | — | shipped | `MultiLaneColumn` lane carrier — the **SoA** in CAM SoA | — |
+| `bgz17` | lance-graph | — (bare) | shipped | LOSSY golden-step recoverable sampling (BASE_DIM=17, GOLDEN_STEP=11); `ScalarCsr::spmv_min_plus` | — |
+| `blasgraph` (module) | lance-graph `…/graph/` | — | shipped | **canonical bit-exact** 16384-bit GraphBLAS substrate (7 HDR semirings) | — |
+| `jc` | lance-graph | — (bare) | shipped | 12-pillar proof-in-code (Pillar 10 `pflug`, 11 `hambly_lyons`) | — |
+| `sigker` | lance-graph | — (bare) | shipped | path-signature kernels (`signature_truncated`, `signature_kernel_pde`) | — |
+| `cognitive-shader-driver` | lance-graph | — (bare) | shipped | BindSpace **SoA** / mailbox / token_agreement / codec bridge | — |
+| `lance-graph-contract` | lance-graph | — | shipped | SPLAT-1 CAM SoA contract types (`CamPlaneSplat`/`SplatPlaneSet`/`CamSplatCertificate`/`AwarenessPlane16K`) | — |
+
+**Convention note:** existing bare/`lance-graph-*` crates predate the `3dgs-*`
+convention and are **not force-renamed** (churn). The convention binds **new**
+splat-pipeline crates. CAM SoA (rule 2) and the no-JSON-hotpath rule (rule 3)
+bind **everyone** — old and new.
+
+## Catalogue checklist (not yet first-party read — do not guess purposes)
+
+lance-graph crates to add rows for, from each `lib.rs`, by lane owners:
+`bge-m3`, `bgz-tensor`, `causal-edge`, `deepnsm`, `highheelbgz`, `holograph`,
+`lance-graph-archetype`, `lance-graph-benches`, `lance-graph-callcenter`,
+`lance-graph-catalog`, `lance-graph-codec-research`, `lance-graph-cognitive`,
+`lance-graph-consumer-conformance`, `lance-graph-ontology`, `lance-graph-osint`,
+`lance-graph-planner`, `lance-graph-python`, `lance-graph-rbac`,
+`lance-graph-supervisor`, `lance-graph` (main), `learning`, `neural-debug`,
+`p64-bridge`, `reader-lm`, `sigma-tier-router`, `thinking-engine`.
+
+## Why this exists (governance parallel)
+
+Same shape as #200 Tier-4, one layer down: docs proliferated → needed "2
+canonical sources + perspective-docs-aren't-evidence"; crates will proliferate →
+need "single `3dgs-*` family + registry + creation gate + CAM SoA as the one
+representation." Naming + registry *track* the battery; the creation gate
+*prevents* the part that shouldn't exist; **CAM SoA + no-JSON-hotpath keep every
+crate speaking the same content-addressed SoA substrate instead of a battery of
+private formats — and the reinvent/reverse-engineer posture keeps us from ever
+becoming an ArcGIS/Cesium consumer.**