Verify a 12-helper co-retiming improvement on pinned PR #299

[maximussthegreat]

Summary

Add a standalone verifier, construction note, and exact receipts for a small conditional improvement to PR #299, pinned at 5f6bd3fbc0e6796dd31263cef1f06dc968186f64.

The construction introduces 12 early helper restorations, explicitly undoes six overlapping old restorations, and co-retimes their donor, blocker and partner deliveries. The complete measured local histogram delta is {1:+18, 2:-12}, with residual-family delta {2:+12, 3:-18, 4:+6}.

Under the same retained source/compiler/all-size interfaces:

• Conditional kappa: 749906929903449/10^18 = 0.000749906929903449, versus PR Conditional κ = 7.50237961741364·10⁻⁴: kernels, early restorations, terminal sinks, a post-sink retiming and addition reordering on the gen5b word (+0.473% over #296) #299's 0.000749899215319498.
• Gain: 0.000000007714583951, approximately 0.00103% relative.
• Bit coarse saving: 750469712644613/10^18.
• Literal stock: 1,229,735 (15 fewer); unreplicated paid calls: 482,665; rank mass: 2,455,070; unchanged deficit: 4,400.

The partner co-retiming is essential. Omitting it creates a forbidden rank-23 to rank-22 target-frame descent. The final ledger charges the extra partner-carrier split. An earlier optimistic root-only ledger is superseded.

Dependency and diff scope

PR #299 remains an unmerged dependency. This PR adds only research/gen5b-coretime12/; it does not copy PR #299's changes into the diff against main, modify PR #294, or assume a main checkout already contains PR #299's inputs.

A small data-only fetcher downloads the 11 exact immutable inputs listed in inputs.json and verifies every byte count, SHA-256 and Git blob hash. Existing downloaded inputs can be reused. No large upstream word/frame/selection bundles, original executables, private research notes, or generated full scalar/bank tables are committed.

Verification

All 49 targeted tests pass on Python 3.12.14, and run_checks.py --check reproduces every committed receipt exactly as JSON data. The aggregate check also passes from an unrelated working directory. Only newly authored code executes.

Coverage includes:

• Both complete 55,985-addition suffixes and all 18,954 independent entry columns, forward/inverse, over F2 and the integers.
• Exact absolute-coefficient majorant equality, plus an independent 172,944-commuting-swap certificate and conservative scalar-norm transport.
• All 60 affected frame paths, 212 exact rational containments, and 12 verified post-descent partner setups.
• All 18 changed endpoint charts, at most 142 factors, within retained 576/815 chart/normalizer allowances.
• Explicit assignments for all 15,427 helper roles, 925,620 role/replica occurrences per stage and 4,628,100 across five disjoint stage namespaces.
• Full-width packing, exact moment enclosures, three bootstrap steps, all 47 strict outer constraints, and rejection of adjacent coarse/kappa grid points.
• Changed displayed finite coefficient 90386288803484701 < 2^80 and conservative 103-bit payload below the 104-bit cap.
• Negative controls for missing restoration, old-restoration undo, partner promotion, bank capacity, duplicated role, noncommuting scalar swaps, malformed arithmetic and changed source bytes.

From the new directory:

python3 -B fetch_inputs.py --destination /tmp/coretime12-inputs
export CORETIME_SOURCE_DIR=/tmp/coretime12-inputs
python3 -B -m unittest discover -s . -p 'test_*.py' -v
python3 -B run_checks.py --check

Use run_checks.py --output-dir /tmp/coretime12-results --write-streams --check to emit full scalar streams, the compact role-address table and exact chart factor programs.

Evidence boundary

No additional open changed-stage construction obligation is identified by these checks. The original prefix and source admission, unchanged finite compiler, weighted selectors/charts, restored-row, routing, prime, recovery, complex-supplier and analytic/all-size interfaces remain inherited hypotheses. They have not been rerun or reproved here. The bit payload remains F2; the stronger integer suffix identity is not a claim that the full parity-filtered supplier is an integer decoder.

The displayed finite-bill calculation does not assign a value to the retained theorem's unknown C_full. Neither make verify nor the upstream complete admission suites were run. This is a conditional research improvement, not an unconditional multiplication bound or full machine-checked theorem. Main's selected certificate, production guards and manuscript patches remain unchanged.

Attribution and assistance

The source package is Chafik Boukhalfa's PR #299, prepared with substantial Anthropic Claude assistance. This contribution preserves the inherited lineage, including DreamingOfClouds, Rohan Arun, eumemic, utcorvusvolat-dotcom, Dugongue, Evan McKinney, Henry Grant, Jacob Sussman, ikeboy, icekylinx, James Chang, sennemmi and their credited dependencies. The complete immutable source notice is linked in NOTICE.md.

New material is Apache-2.0 and was prepared with substantial OpenAI assistance. Separate AI-assisted checks are not human peer review, an OpenAI endorsement or proof-assistant verification. No global priority or optimum claim is made.