Skip to content

Certify a 2425-bank minimum for the fixed native snapshot schedule - #7681

Open
aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-025
Open

aliromman wants to merge 1 commit into
CrocSwap:mainfrom
aliromman:codex/consecutive-result-025

Conversation

@aliromman

@aliromman aliromman commented Nov 18, 2026 •

Copy link
Copy Markdown
Contributor

Result and scope

The banked native DAG of PR #7674 can use 2430 private homes instead of 15290, with the same row order, scalar operations, paid calls and reset allowance. This field-level construction uses 2425 snapshot banks, one shared paid-row accumulator and four scalar scratch fields. Sixteen disjoint occurrences reserve 38880 homes instead of 244640. Public fields, guides, guards, child/adapter workspace and compiler backing remain additional.

The snapshot pool size 2425 is minimum for the stated fixed-order contract. A bank is reserved before its row starts; every donor stays through the complete row and child return; all terminal outputs stay until installation; distinct live tokens cannot share a bank. Event 5755, row token 98194, has 2425 simultaneous owners. A complete interval allocation attains that lower bound. This is not a global minimum over schedules, in-place methods, recomputation or equal-value sharing. The total 2430 is a sufficient reservation, not a proved minimum for all scratch.

Read the proof, all 9144 allocation entries and the peak witness, receipt, verifier, source inventory and scope record. This completes the storage claim under the source's review request. The source graph constructor is reused; the allocation and owner replay are new.

No new kappa, supplier admission or selected-assumption discharge is claimed. All five assumptions and selected PR #7000 remain unchanged. This is a smaller conditional field reservation, not a measured runtime or memory result for a complete multiplier. It applies only to the pinned graph, including all fifteen pivots.

Evidence and cost

Normal and optimized CPython 3.12.16 replay passed on macOS arm64. A per-slot interval check and a separate difference sweep verify the allocation and lower witness. A second execution uses live owners and remaining-consumer counts. It evaluates all 8533666 scalar coefficient products, checks every division and donor chart, and verifies all 1501 public endpoints. It ends with every private owner retired. Six controls reject a live-bank collision, early retirement, lost outputs, insufficient capacity, a missing token and a terminal sign fault.

There are still 15290 logical lifetimes and 30580 full initialization/retirement services per native occurrence. Each reuse is initialized and retired in full. Terminal rows accumulate directly in their pool slots, so no extra copy or temporary lifetime is added. The source's conservative raw-caller reset allowance remains 11816384. All 6142 paid calls and the productive invoice are unchanged. View setup, selection, scans, returns and suspended backing remain charged under the source's physical primitives.

Metadata, committed bytes, executable modes, source ancestry, the official 26-file export and its integrity check passed. A read-only optimized export replay passed without Git or a usable PATH; all 27 payload/manifest hashes and modes stayed unchanged. All 15 retained predecessor files match their original Git blobs and SHA-256 values. The proof and README were checked through GitHub Markdown rendering.

From the repository root, with CPython 3.12 and its standard library:

python3.12 -I -W error -B research/banked-native-snapshot-reuse/verify.py
python3.12 -I -W error -B -O research/banked-native-snapshot-reuse/verify.py
python3.12 -B scripts/research_package.py --package research/banked-native-snapshot-reuse
python3.12 -B scripts/check_research_commit.py --package research/banked-native-snapshot-reuse --check-source-history --check-modes

Expected status markers:

PASS_BANKED_NATIVE_SNAPSHOT_REUSE
PASS_REGISTERED_RESEARCH_PACKAGE
PASS_RESEARCH_COMMIT_BYTES

Hosted CI did not certify this package. The plan job failed in the shared test_current_ci_section_matches_selected_record: the base guide names frontier6030 while the selected record requires make frontier7000-verify. The log reports 760 tests, one failure and 24 skipped; package replay was skipped. This package changes no shared CI file. Local replay remains the scientific evidence. Predecessor scientific suites, physical native operators, the signed caller, selected cold replay, Lean and Docker were not run. A concrete compiled fixed-tape lowering, independent human review and formal-kernel review remain open.

Sources and credit

Prepared by @aliromman with substantial OpenAI Codex assistance. The graph, banked realization, original 15290-home reservation and physical primitives come from PR #7674, source proof and pinned constructor. Its scientific commit is 26b6c469aa6c4aa04322ce25a22150f90e5fb14e; head 28b2af322a1fab01b3734365988930920f9dda48 adds the shared CI repair. @SYLforge supplied PR #7669. The original parent credit and source notices are unchanged. The interval-allocation method is standard; this contribution supplies the certificate, sharp witness and complete ownership/reset argument for the current pinned graph. New work is Apache-2.0.

The separate precision result PR #7680 is related work, not a dependency of this source-grid replay. A newer role-230 banked handoff changes the graph and is not covered. Base: canonical main e54523517d0f03e58352e647c403a5af3e525a57, priority revision 8; card supplier-boundary-audit refined to the explicit storage request. Open/closed PR and Discussion overlap was refreshed; sparse local notes and unpublished work remain coverage limits. No source PR or selected artifact is edited.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant