[petalas]

Summary

unchanged PR2586 word, all 5,398 events and native prefixes
  104 independent Y inputs:  4 B9
   78 independent N inputs:  2 B9
  176 other public inputs:     B9
  eight owned private banks: exact zero erasure

Certifies this sharp conditional entrance among independent principal-ideal Gaussian Newton product modules. Every restricted source has an attaining coefficient, a separate backward witness check, and a literal failure when weakened by one pi. Exact correlations let N2 and N4 use 2 B9 where minimum-path propagation asks for 4 B9.

The symbolic theorem explicitly retains the native recipe/change-of-chart identities. The complete literal fixture does not independently prove those identities for every vector. This changes the global input contract; it does not admit the actual caller, implement a consumer, supply a physical saving or improve kappa.

Only research/p104-newton-source-contract/ is added. Credit Clear1nsight/PR3382, ywz65/PR2586, Tekalign/PR3025, Ross Richardson/PR3068 and the full retained P104/native/precision lineage. The active actual-caller investigation studies the first collector's 58-bank cone and named consumers. The distinct global scope is recorded in the advance claim.

Evidence

• Before: one collector had a Newton entrance contract; the complete all-role product criterion was missing. After: exact forward rational propagation checks 102,787 coefficients and all final logical restoration rows; 182 reverse-adjoint traces check every positive maximum.
• Complete literal replay covers 2,070 native occurrences, 2,490 wrapper records, 7,682 native wrapper/stage checks, 5,390 non-erasure event outputs, all 358 dense signed public inputs/outputs and eight private erasures. All actual pairs agree with the signed-floor formula. All 182 one-pi-weakened actual source fixtures reject.
• Registered normal/optimized full replays and seven behavioral test groups pass on Python 3.12.15. Ruff diagnostics, exact committed-byte/source-history checks, and the optimized Git-free committed export pass. The export remains byte-identical, and corrupted/missing scientific inputs fail through the actual CLI. These are local checks, without a hosted-CI or human-review claim.
• Replay: python3.12 -I -Werror -B scripts/research_package.py --package research/p104-newton-source-contract. Standard library only. Proof, complete receipt, exact sources, limitations and open questions are included.

Merge Danger

Door: two-way. One isolated eight-file research package.

Blast radius: package. Native covariance remains inherited. Actual preparation, physical ownership/routes/control, fast recursive pricing, consumer cost, tensor-depth/all-size extension and supplier admission remain open. The selected frontier and frozen predecessor inputs are unchanged.