danielcamposramos · danielcamposramos · Apr 18, 2026 · Apr 18, 2026 · Apr 18, 2026 · Apr 18, 2026
diff --git a/.gitignore b/.gitignore
@@ -57,3 +57,16 @@ rpn_kernel_profile.*
 
 # Runtime data (generated per session)
 $ROOT/
+
+# Ingestion staging — knowledge working state (lives in local folder, not repo)
+scripts/ingestion/staging/
+
+# Compiled binaries (rebuild locally)
+scripts/ingestion/d3/build/
+
+# Benchmark runtime outputs (regenerated per run)
+data/benchmarks/sovereign_bitnet_attention_*.json
+
+# Runtime dedup digests (regenerated per pipeline run)
+TEMP/D3_duplicate_cluster_digest_*.jsonl
+
diff --git a/AGENTS.md b/AGENTS.md
@@ -4,15 +4,18 @@ This repository uses AI partners (both external AI assistants and internal Synth
 
 ## MCP Infrastructure — USE THIS FIRST (All Agents)
 
-**Two MCP servers run locally. Query them BEFORE reading spec files from disk — save your context.**
+**Three MCP servers run locally. Query them BEFORE reading spec files from disk — save your context.**
 
-- **`k3d-knowledge`** — Qdrant semantic search over all 35 `docs/vocabulary/*.md` specs (1319 chunked points). Tool: `mcp__k3d-knowledge__qdrant-find`. Pattern: `qdrant-find("What does the spec say about X?")` → returns relevant excerpts + source paths.
-- **`ollama-specialists`** — Delegate implementation, planning, research, and multi-angle analysis to local Ollama models. Tools include `kimi_swarm`, `ask_coder`, `plan_task`, `flesh_out_code`, `extract_facts`, `summarize`, `web_search`, `route_specialist`, `ask_cloud`, `mvcic`, `memory_harvest`.
+- **`k3d-knowledge`** — Qdrant semantic search over all 35 `docs/vocabulary/*.md` specs (1319 chunked points), served on `:8501`. Tool: `mcp__k3d-knowledge__qdrant-find`. Pattern: `qdrant-find("What does the spec say about X?")` → returns relevant excerpts + source paths.
+- **`k3d-ptx`** — Qdrant semantic search over PTX ISA + CUDA C Guide + Inline PTX reference (11,298 points), served on `:8503`. Tool: `mcp__k3d-ptx__qdrant-find`. Use this before touching `.cu`, `.ptx`, or CUDA bridge code.
+- **`ollama-specialists`** — Delegate implementation, planning, research, and multi-angle analysis to local Ollama models, served on `:8502`. Tools include `kimi_swarm`, `ask_coder`, `plan_task`, `flesh_out_code`, `extract_facts`, `summarize`, `web_search`, `route_specialist`, `ask_cloud`, `mvcic`, `memory_harvest`.
 - **Standing directive from Daniel**: "Always dispatch ollama specialists instead of burning your tokens."
 
-**Workflow**: `qdrant-find` spec lookups → `plan_task` / `ask_coder` for implementation → read full files only if MCP results are insufficient.
+**Rule of three**: `qdrant-find` specs → `qdrant-find` PTX (if kernel-touching) → `plan_task`. Then code.
 
-Both MCP servers are available to Claude Code, Codex, and Cline via streamable-http transport (ports 8501/8502). Config lives in `/home/daniel/.claude/settings.json`, `/home/daniel/.codex/config.toml`, and Cline's `cline_mcp_settings.json`. Launch script: `/home/daniel/.claude/launch_mcp_containers.sh`.
+**Timeout rule**: for `kimi_swarm` and deep `ask_cloud`, use timeouts of at least `240000 ms`.
+
+All three MCP servers are available to Claude Code, Codex, and Cline via streamable-http transport (ports `8501/8502/8503`). Config lives in `/home/daniel/.claude/settings.json`, `/home/daniel/.codex/config.toml`, and Cline's `cline_mcp_settings.json`. Launch script: `/home/daniel/.claude/launch_mcp_containers.sh`.
 
 ---
 

diff --git a/ATTRIBUTIONS.md b/ATTRIBUTIONS.md
diff --git a/CODEX.md b/CODEX.md
@@ -544,6 +544,42 @@ def answer_query(query):
 
 ---
 
+## Minecraft-for-Cognition Lanes (handoff 2026-04-20, Claude-pilot session)
+
+**Context:** Codex was limit-locked; per memory [`Claude Runs Pilots When Codex Is Limit-Locked`](../../home/daniel/.claude/projects/-K3D-GitHub-Knowledge3D/memory/feedback_claude_pilots_when_codex_limit_locked.md), Claude piloted the Document-Galaxy-as-Symlinks + Texture Forge + Image→Procedural + Image→3D + ARC3 live-screen + Memory-as-Image architecture pass. Full spec: [TEMP/CLAUDE_TEXTURE_FORGE_IMAGE_TO_3D_ARC3_SCREEN_04.20.2026.md](TEMP/CLAUDE_TEXTURE_FORGE_IMAGE_TO_3D_ARC3_SCREEN_04.20.2026.md).
+
+### What Claude-pilot landed
+
+1. **Registry reservations** — [docs/vocabulary/RPN_DOMAIN_OPCODE_REGISTRY.md §11.2](docs/vocabulary/RPN_DOMAIN_OPCODE_REGISTRY.md) extended with 7 new reserved blocks: 0x280 (Texture Forge), 0x290 (Image→3D), 0x2A0 (ARC3 screen), 0x2B0 (Memory-as-Image), 0x2C0 (MVCIC extensions), 0x2D0 (Document Galaxy), 0x2E0 (headroom).
+2. **73 opcode constants** — [knowledge3d/cranium/ptx_runtime/rpn_opcodes.py](knowledge3d/cranium/ptx_runtime/rpn_opcodes.py) + `__all__` export updated.
+3. **Token registrations** — [knowledge3d/cranium/ptx_runtime/modular_rpn_engine.py](knowledge3d/cranium/ptx_runtime/modular_rpn_engine.py) OPCODES dict grew 277→398.
+4. **ARC3 bridge kernel** — [knowledge3d/cranium/codecs/kernels/arc3_screen_bridge.cu](knowledge3d/cranium/codecs/kernels/arc3_screen_bridge.cu) compiled clean to PTX (sm_86, zero spills, 5 kernel entries).
+5. **Document Galaxy vocab spec** — [docs/vocabulary/DOCUMENT_GALAXY_SYMLINK_SPECIFICATION.md](docs/vocabulary/DOCUMENT_GALAXY_SYMLINK_SPECIFICATION.md) (NEW, normative). Core principle: a document = star + ordered list of Word Galaxy symlinks; chars/glyphs/meanings stored once across corpus.
+6. **§13 Collision Resolution** — documents the relocation from 0x1D0–0x1FF (already reserved for VIRTUAL_PAGE_*) to 0x280–0x2DF.
+
+### Open P0 for Codex (ordered by critical path)
+
+1. **ARC3 bridge Python launcher** — ctypes wrapper for [arc3_screen_bridge.cu](knowledge3d/cranium/codecs/kernels/arc3_screen_bridge.cu). Zero numpy. Palette upload via `cuMemcpyToSymbol(c_arc3_palette)`. Wire into `knowledge3d/tablet/wine/game2d_wine.py` replacing Python frame processing.
+2. **Texture Forge kernel** — 0x280–0x28F (splat, kuwahara, voronoi, normal-from-height, …). Spec: [TEMP/CLAUDE_TEXTURE_FORGE_IMAGE_TO_3D_ARC3_SCREEN_04.20.2026.md §3.1](TEMP/CLAUDE_TEXTURE_FORGE_IMAGE_TO_3D_ARC3_SCREEN_04.20.2026.md).
+3. **3D-opcode GPU promotion** — 0x29C/0x29D/0x29E replace the Python host-fallback path currently fronting 0x170–0x176.
+4. **Inverse-fit Lane B kernel** — 0x28A/0x28B (RPN program inversion from raster target).
+5. **Memory-as-Image kernel** — 0x2B0–0x2B5 (DotMap bake of reasoning trace → 4-way-addressable memory cells).
+6. **Document Galaxy kernel** — 0x2D0–0x2DB per vocab spec §4 opcode surface table.
+7. **Ingestion integration** — document stars emit via proceduralizer; text bytes never reach VRAM (only Word Galaxy symlinks).
+
+### Sovereignty gates (MUST hold before merge)
+
+- `grep -r "numpy\|cupy\|scipy" knowledge3d/cranium/codecs/` returns zero hits on hot-path files
+- Every new opcode has a §11 registry row before minting (see §13.3 failure-mode note)
+- `nvcc -arch=sm_86 -O3 -Xptxas -warn-spills` reports **zero spills** for every new kernel
+- Document stars contain zero raw text bytes — only Word/Character/Meaning star IDs
+
+### MVCIC failure-mode lesson (2026-04-20)
+
+The 6-partner MVCIC chain recommended opcodes 0x1D0–0x1FF — **already reserved** since April 18 for VIRTUAL_PAGE_*. Caught via grep of §11 before minting; relocated to 0x280+. Same failure as the 2026-04-19 incident. **Rule:** always grep [RPN_DOMAIN_OPCODE_REGISTRY.md §11](docs/vocabulary/RPN_DOMAIN_OPCODE_REGISTRY.md) before minting anything from MVCIC output.
+
+---
+
 ## Codex's Mandate
 
 **Wire ALL GRE specialist kernels into the composed head pipeline. Expand benchmark coverage. Shrink Python orchestration toward ~200 lines. Keep ARC 10/10 and Math 20/20 pinned.**