research: bind the fixed kernel entrance to native replay

[Clear1nsight]

Problem and scope

PR #1179's fixed kernel entrance had an explicitly unrun native five-stage check. PR #3273's public XOR normaliser and PR #3326's native audit were tested on the original zero-start word and its scatter variants, so their acceptance does not by itself admit that changed entrance. Native selection metadata also needs an explicit connection to the program it describes.

This PR adds one finite, independently reviewable package, research/kernel-entrance-native-boundary/, on frozen base 6f100d4bb5083f8c35d96fd95a539811c62f245c. It applies the actual public normaliser to the zero-start source, proves an exact three-file export bridge to the constructor input, runs the authenticated PR #1179 construction with pivot 444, partner 470 and inherited frame 79, then checks the original exported program with unchanged native C++ checkers and genuine Boost 1.85.

The public normaliser still rejects the transformed nonzero-start candidate. Its wrapper suppresses the checker's stdout diagnostic, so the receipt records both the wrapper rejection and a separate unchanged-checker run identifying target/helper 444 does not start at frame 0. This is an explicit changed finite contract.

Published head: eb50e48ee908f99d89223abf08967ab5e1fd729e.

Result and added evidence

The committed RECEIPT.json reports PASS_KERNEL_ENTRANCE_NATIVE_COMPOSITION. Its SHA256 is 8b8d47ffaddf8e71e8ff20ad0bb8d8408a7e9993ae810cb2b6e046cd0b8fdcf4.

• Both the nine-copy kernel fixture and the ten-copy actively used alias fixture pass native legality and all 883 formal columns under the correct copy-count contract. Both native checkers reject the opposite copy count.
• Independent rational frame analysis and all 661 local symbolic columns establish the fixed kernel relation, complete word binding, restored dirty helpers, COPY/ERASE lifetime and local invoice. The pure word maps commute, N(G(w)) = G(N(w)).
• The active alias uses both retained slots 0 and 9 for helper 472. It preserves the operator and adds the full rank-seven COPY charge. Local totals are 24,252 ADDs and raw rank mass 5,128 or5,135. These reproduce the credited predecessor saving rather than claim a new optimisation.
• Declaring partner 223 instead of 470 while preserving the valid program passes native legality with the identical receipt. The independent model rejects the declaration and proves that the falsely declared pair fails the kernel relation at 53 target rows, with the first witness at target 111. Native execution acceptance alone does not establish the declared kernel identity.
• Removing the final restoring shear passes geometry but fails the native column check, with the exact sole local defect output 470 += input444. Wrong dimension and synthetic source-owner/donor collisions are rejected by legality. Each positive native column replay also rejects all six omitted-bridge controls.
• A cancelling duplicate pair produces byte-identical public exports and the identical transformed word. Its native receipt is reused only under that exact equality.

All14 immutable source files authenticate by commit, Git blob, SHA256 and byte count. All182 compiler-observed Boost headers match the official archive. The verifier ran under python3 -I -O -B, rechecked its three implementation files and retained source snapshot, and completed without altering frozen predecessors.

Reproduction

The README gives the complete Git and Git-free source-export commands, official Boost archive pin, environment, exact mathematical scope and next questions. The dedicated native command is:

python3 -I -O -B research/kernel-entrance-native-boundary/verify.py \
  --git-sources --source-root . \
  --boost-include /path/to/boost_1_85_0 \
  --boost-archive /path/to/boost_1_85_0.tar.bz2 \
  --workdir /tmp/kernel-native-review \
  --output /tmp/kernel-native-review-receipt.json

The work directory and output receipt must be new. This package uses a dedicated native replay and is not registered in the generic Python package workflow. No hosted-CI result is claimed.

A second complete replay from the independently checked Git-free source archive produced a byte-identical receipt. The companion focused patch also reconstructed the exact entire published Git tree through a private index. The archive contains the eight contribution files, fourteen pinned inputs and seven inherited notices, plus reproduction instructions and its inventory.

Limits, coordination and credit

The positive selection is nonempty, but its identity role labels are local and the physical source-owner/donor manifests remain empty. No physical hosting, complete bank invoice, fresh bank nondegeneracy audit, entry conversion, native 9/11-stage check, selected assembly, precision/tape or all-size transfer is established. Retaining the old width-eight host charges back the local one-unit rank saving. There is no new global kappa, supplier admission or selected-frontier change.

Task supplier-boundary-audit has scope_status=changed, with the original card, family narrowing, checkpoint coverage and remaining obligations in TASK.json. The advance claim and released predecessor handoff are in Discussion #2976. Required baseline orientation was attempted but could not complete its selected-source bindings in the sparse checkout; the exact missing-source result is retained and is not a selected-construction failure.

Adam Gall owns the PR #1179 gauge, census and local saving, at a09f799fdb312e049a81f59b44f1339b5a053bbe. Philipp/mightyOil's PR #1253, at 21dd0227f0c0d1b2a91c01c23cb0a16a3d1934be, supplies the fixed-word common-entrance classification. Nick Petalas owns PR #3273 at 2e0408a03e770c85a7e59b9ade547dcbbaea7e73; this continues Clear1nsight/OpenAI PR #3326 at b142a984f724197c28e410f076dc914627b147f5. The README credits the earlier kernel, selected-word, exporter, native checker and invoice work. Archived or partially reviewed predecessors are not treated as scientific admission.

Prepared for Clear1nsight with substantial OpenAI assistance from four concurrent agents. New work is Apache-2.0; predecessor sources and notices are preserved. Independent human review remains open.

[Clear1nsight]

Git-free reproduction bundle for head eb50e48ee908f99d89223abf08967ab5e1fd729e.

The ZIP contains the eight committed package files, fourteen pinned source files, seven inherited notices, reproduction instructions and a complete byte inventory. A fresh replay from this archive produced the exact committed receipt, SHA256 8b8d47ffaddf8e71e8ff20ad0bb8d8408a7e9993ae810cb2b6e046cd0b8fdcf4.

Archive SHA256 f6004fcfa8d067ed8229b8ba9c96ad228399698a024ee41807919b059d081146 (317,691 bytes). Boost and the compiler remain separately supplied dependencies as documented.

kernel-entrance-native-boundary-source.zip