docs(onnx) S0: convergence amendment (HOLD, competing capstones #1744/#1750)

.claude/plans/d-onnx-relalg-cap-convergence-1745-v1.md

@@ -0,0 +1,80 @@
+# D-ONNX-RELALG-CAP — unified architecture and evidence gate
+
+> **Path note (2026-10-23, merge of #1745).** #1744, #1745 and #1750 each wrote `.claude/plans/d-onnx-relalg-cap-v1.md`. The operator merged #1744 first, so its text is canonical at that path, and then ordered #1745–#1754 merged. This document is #1745's version, unchanged, moved here so it is not a competing canonical file: an amendment carrying the Tensor / DR / Hybrid convergence. #1750's version is at `d-onnx-relalg-cap-capstone-1750-v1.md`. Where they disagree, the native-first correction (`d-native-first-dependency-correction-v1.md`, #1753) governs.
+**Date:** 2026-10-23  
+**Status:** PROPOSED / documentation only. No implementation, ABI, operator coverage, speedup or adoption is asserted.
+
+## Mission
+Reuse ONNX model/graph/function representation and versioned operator vocabulary rather than creating a competing computation graph. Use existing lance-graph-relalg as a mathematical *polyfill/planning* landing zone under the D-RA-2 decision: canonical single-table set algebra and rewrite laws belong to lance-graph-contract, while relalg retains relational planning and lowering.
+
+Preserve ClassView, existing DR64/DR256, CausalEdge64, 64K SoA, Palette256 typed methods, Quack, R2IL, mask-risc, SIMD, evidence/certification and Thought-Git ownership. No additional general executor, new global identity layout, implicit cross-table positional gather or mandatory tensor/cube allocation.
+
+## Convergence thesis
+**One declared mathematical contract; three competing physical plans:** Tensor (dense GEMM/Conv/Attention-style kernels), DR (deterministic Boolean support/address selection + typed numeric folds), Hybrid (DR-selected support with numerical microkernels). Physical plan selection is not a mathematical rewrite certificate. Performance wins must be measured; eligibility is established before cost selection.
+
+ONNX semantic adapter → existing contract/RelAlg mathematics → optional Global-to-Local focus/bind → {Tensor, DR, Hybrid} → existing physical backends → independent oracles and appropriately scoped certification. Avoid requiring every ONNX node to create a DR view.
+
+The canonical holographic addressing inversion is Thought(t)[Table(a)[i] : Table(b)[i]]: operand tables are named first, a compatible coordinate selects their rows, and per-table presence controls inclusion. Not equivalent to arbitrary sparse execution, ontology equivalence or value-steered foreign gathers. Preparation-time attested same-table seeks remain subject to ADMISSION B; cross-table positional gathers remain prohibited.
+
+Separate **LogicalShape**, **FocusWindow**, **PhysicalView**, and **ResultDemand** (Exists/Count/Rows/Tensor) conceptually. They are not new ABI types without justification. Every supported operator requires type, shape, domain, semiring, multiplicity, reduction, effect, numerical/error and binding/authority premises.
+
+## Geometry and epistemic non-equivalence
+| Logical Boolean plane | Cells | Dense bit equivalent |
+|---|---:|---:|
+| DR64² | 4,096 | 512 B |
+| DR256² | 65,536 | 8 KiB |
+| DR64³ (PROPOSED) | 262,144 | 32 KiB |
+| DR256³ (PROPOSED) | 16,777,216 | 2 MiB |
+
+DR64² / DR256² = **1/16**; full DR64³ / DR256³ = **1/64**. A 1/16 cubic-window claim needs a specifically defined two-axis projection or other explicit geometry. Logical shapes do not imply allocations or cache residency.
+
+The existing 64-thought occupancy model M[address,thought] (65,536 × 64 bits) is **not** DR64³. Identity, property, relationship, thought and tensor axis are distinct typed dimensions unless a witnessed mapping proves otherwise.
+
+## Architectural owners and source crosswalk
+- D-RELALG-UNIFIED / PR #1641: current capabilities are partial. No general μ/preimage, universal ONNX mapper or common executor is implied.
+- D-QR-CAP / PR #1663: ClassView and WideFieldMask select *structural fields*, not causal/relational identity; optional DR geometry is separate.
+- D-QSP/D-RFT / PR #1667; PR #1685: lab Boolean/Distinct and Bag composition, including competing physical strategies, not dense numerical ONNX MatMul proof.
+- D-DRMC / PRs #1700, #1707, #1738: narrowly typed Boolean/Distinct associativity certificates exist. Independent proof checker, general law coverage and R2IL/Quack production integration remain separate.
+- PR #1740: bounded fresh-frame certificate reuse demonstrated in the lab for its admitted law family; not general rebinding.
+- PR #1741: inspect latest closed/merge disposition before describing its Julia/egglog map as open or superseded; ONNX does not replace its unification/proof topics.
+- Canonical: .claude/plans/dr-mathematical-compiler-v1.md, d-relalg-unified-v1.md, d-qr-classview-binding-isa-capstone-v1.md and .claude/architecture/cartesian-epistemic-self-model.md.
+- D-RA-2, ADMISSION B, frame/focus/scope as three distinct guards are binding constraints.
+
+## Stage map (separate docs-only draft PRs)
+- **S0** this architecture/recovery map.
+- **S1–S2** ONNX intake, typed semantic preservation, minimal RelAlg polyfills and independent oracles.
+- **S3a–S3c** typed Cartesian DR64³ probe; holographic focus and expansion; optional DR256³ validation hypothesis.
+- **S3d** three-arm Tensor/DR/Hybrid differential and cost benchmark.
+- **S4–S5** typed Palette256/cognitive-method probe, ONNX export, cold certification and native specialization evaluation.
+
+Stages S3a and S3b may be developed independently once common contracts exist. S3d need not wait for a DR256³ production path; S4 activation import need not wait for cubic validation. These are reviewable plans, **not implementation authorization**.
+
+## Research ownership and workflow
+Haiku: bounded inventories and evidence links. Sonnet: specifications, matrices, documentation/board proposals. Opus: formal semantic choices, adversarial counterexamples and review. Final authorization remains with operator and current code/board owners. EXO-50 for cross-agent coordination; GitHub board ledgers for durable findings.
+
+## Open points, not presumed resolved
+- Recover original DR64³ 12-stage plan: no dedicated source confirmed by this limited main/open-PR search; full branch/ledger search remains **OPEN**.
+- Verify available ONNX importer(s), official current opset schemas, external-data security, operator families and effect semantics before selecting dependencies.
+- Identify actual minimum RelAlg additions by counterexample; do not create an independent AST by reflex.
+- Establish projection dependency closure and out-of-focus behavior: exact-local / exact-expandable / approximate / invalid / unknown.
+- Establish whether DR256³ adds validation value versus independent scalar or sparse references; no theorem can follow from larger sample alone.
+- Distinguish data parity, mathematical proof, binding validity and measured hardware cost.
+- Independent proof checker; general snapshot rebinding beyond #1740; numerical operators and production runtime remain OPEN.
+
+## S0 acceptance gate
+Documentation-only, no shared owned file overwritten, no board closures or autogenerated indexes manually changed. Opus falsifier review must confirm no second executor, no loss of SQL K3/Bag semantics, no false ONNX compatibility claims, no unwitnessed gathers, no 1/16 cubic misstatement and no unverified speedup. Stage implementation requires separate permission.
+---
+
+## 2026-10-23 source-backed correction: existing BGZ/Helix/ndarray, triple S0 collision
+
+**DO NOT MERGE THIS S0 AS A COMPETING CANONICAL FILE.** #1744, #1745 (this branch) and #1750 all add `.claude/plans/d-onnx-relalg-cap-v1.md`; their board and scope differ. #1744 has 14 OP-ONNX issues and source-backed inventories, #1750 has 19 open points and CC2 board reconciliation. Choose one owner/canonical S0 through EXO-50. Preserve the relevant evidence, then incorporate our Tensor/DR/Hybrid convergence as an owner-approved amendment. No rebases, forced overwrites, auto-closes or silent merges.
+
+**Corrected existing-capability facts:**
+- `crates/bgz-tensor` already provides WeightPalette, FisherZTable/FamilyGamma, attention and composition tables; `crates/bgz17` provides earlier palette/distance layers.
+- `crates/helix` already implements RollingFloor and 256×256 DistanceLut, including an exhaustive 256³ circular-metric check. Its KNOWLEDGE.md retains an Overlap & Consolidation plan.
+- `AdaWorldAPI/ndarray` already implements SIMD f32 tanh, ARM NEON, Burn `CompiledAttention` (PR #42) and `CompiledLinear` (PR #100) with conventional GEMM fallback. These are implementation assets, not generic ONNX certification.
+- **P0 binding limitation:** Burn caches the compiled attention by d_head and compiled linear by matrix shape. Their optimized paths do not validate original operand weight/model identity; an ONNX adapter must not silently select them on shape alone. The attention table produces precomputed distance-derived scores, not the numeric ONNX MatMul result by construction; compiled centroid linear is an approximation with a known uncorrected zero-point term.
+- `bgz-tensor::fisher_z` documents ρ ≥ 0.999 across 21 Qwen3-TTS roles (5,000 sampled pairs each). This is scoped empirical quality evidence. #1726/#1731 show particular multi-hop palette composition laws fail; EXO-139 records silent Fisher-Z saturation; EXO-137 records ungated BGZ integration tests. Neither all-Palette256 success nor all-Palette256 failure follows.
+- Therefore **S4 is REUSE/INTEGRATE/ADMIT**, not a codec construction stage. #1749 owns the proposed identity-safe numerical adapter research; #1748 must compare the existing compiled palette paths under explicit quality policies, not silently treat them as exactly equivalent Tensor/DR/Hybrid variants.
+
+This amendment does not claim a new S0 owner ruling, completed import, new tests, a production cache fix or a mathematical proof. Refer to #1746–#1749 and Linear EXO-50, EXO-137, EXO-139, EXO-129 and EXO-132.