Compile native dyadic blocks: 766 → 698 scalar allowance (no new κ)

research/dyadic-block-compiler/.gitattributes

@@ -0,0 +1 @@
+* -text

research/dyadic-block-compiler/ATTRIBUTION.md

@@ -0,0 +1,11 @@
+# Attribution and overlap
+
+The retained native source and scalar pricing are pinned to canonical main `2a13eb1827784206202db90e5dc8327cf8a250db`. The source is zxv's native/compiler lineage, retaining its earlier contributors and notices. The cleanup windows are already present in that source; they are not new data or a new provider here. Credit Jacob Alan Sussman's mixed-target compiler and Cinna Davis's padding, query-thinning, retiming and joint-boundary work in #4359, #4416, #4440, #4441, #4447 and #4470.
+
+The immediate predecessor is Dugongue's #4489, head `67d1ded63edc604b3d9e6506a4f5d3689bdc9f41`, which proves the rank-one shared bridge and records no κ improvement. This package adds an alphabet-preserving arbitrary-rank factorization algorithm, executable emitted programs, direct source-window binding and exact local allowance reductions. It does not copy #4489's graph or repeat its chart searches. Integer unimodular reduction and low-rank linear-map factorization are classical; no literature priority or absence-of-overlap claim is made.
+
+Discovery consulted the shared MathProof DAG and `Meta/TITensorRank20260902.lean` for transported rank-one decompositions, and `Abstraction/VPVNPFixedGateVisiblePricing20260826.lean` for `linear_response_top_sum`. These are source-only precedents for moving a linear response through a finite sum, not freshly audited Lean imports or new proof dependencies. No Smith-normal-form theorem from that tree is asserted. The canonical DAG and Lean files remain unchanged.
+
+Main, recent PRs and the newest visible handoffs were checked. API title/body search required authentication unavailable to this worker; public web searches did not establish complete coverage. Unknown coverage is not evidence of novelty. Neighboring tensor-error and uniform-rank compiler PRs #4480–#4488 concern different proof obligations; this package makes no competing bound claim.
+
+Prepared for Dugongue with substantial OpenAI assistance in mathematics, discovery, implementation and verification. New work is Apache-2.0; existing notices remain. No independent human review or proof-assistant acceptance is claimed.