[CubeJerry]

Summary

The proposed QUERY-only caller has a conditional 843,431,488-unit ordinary
invoice after its rounded-precision surcharge. The new literal arithmetic
kernels have physical tape/flat-table bills. The source already warns
that those units have not been numerically connected.

This package sharpens that obligation with an all-width mandatory lower
cost vector for the actual compact ADD constructor, an exact source
census, maximal legitimate released ADD credit, and a necessary numeric
conversion threshold. No new kappa, physical supplier admission or
unconditional runtime/cap obstruction is claimed.

Exact result

For reduced coefficient a/b, let L=max(bitlength(|a|),bitlength(b)) and
delta=[oddpart(b)>1]. Every successful retained compact ADD on B public
numeric data bits requires at least B*(250+73L+28delta) raw physical
reads/writes/moves. With fetched flat rows, the lower bill is
B*(288+86L+33delta). Initialization, mandatory Horner rewinds, staging,
installation, head returns and current private-body retirement are
included; omitted setup/extra-width/history costs are nonnegative.

The actual local source has 61,052 ADDs, sum L=167,955 and 4,079 odd ADDs.
At 18 joint contacts, a direct full ADD substitution has mandatory bills
497,482,686B raw or 578,910,834B with rows, assuming equal comparison volumes.
The separate COPY is excluded, and this audit does not install a complete
source registry or build all 4,462 coefficient tables.

Removing every old ADD and ADD-rounding charge releases 47,484,288. The
rest of the frozen model costs 795,947,200, leaving 204,052,800 for the
replacement ADD services. A converter permitting rho raw actions per
tariff unit per comparison data bit therefore necessarily requires
rho>=82913781/34008800. With flat rows included it requires
rho>=96485139/34008800.

At explicitly assumed equal units rho=1, the fixed replacement ledger
already has lower endpoint 1,293,429,886 (raw) or 1,374,858,034 (with rows).
This rules out that numeric substitution certificate. It does not prove
the actual full program exceeds the cap: other entries remain inherited
upper/model charges, and neither the canonical O(V) lemmas nor the source
defines rho=1. Different implementations, repricing or conversions remain
possible. A concrete common-volume/clock bridge and complete upper bill
are still required.

Source pins and novelty

Canonical base: 7c49fe32c55de02c2c873655c804b16e4c3fef0a.
DreamingOfClouds' #6936 word is pinned at tested
head 03626137d0bd57f09b0bf9597376183ce6fa9683; later head d360b11 changes
only VERIFY timeout. The QUERY-only caller is public 6179034, from
#6937 at a98df73; precision is 56bc938; compact
ADD is 38cd824; lifecycle is ab239cb; rounded SCALE is 6476455; global LAMBDA
is 591394d. Full SHAs, paths, Git blobs and SHA256s are recorded locally.

The existing warning that tariff conversion is owed is credited, not
claimed new. New contributions are the universal forced-pass vector,
its exact source-weighted sum and the necessary replacement inequality.
The proof also corrects an illustrative quotient-position count 64 versus
59 in the compact predecessor; that predecessor's slack 128U phase upper
bound remains valid.

Reproduction and checks

python3 -I -B research/literal-tariff-conversion-audit/verify.py --suite all
python3 -I -O -B research/literal-tariff-conversion-audit/verify.py --suite all

Expected: three PASS_LITERAL_TARIFF_* markers. 19 authenticated immutable
dependencies; standard library; no runtime Git/network/solver.

The checker independently reconstructs both full model invoices and the
complete source ADD multiplicities. 36 unchanged-executor zero cases use
independent hardware meters, heterogeneous widths through 65, all three
actual maximal-coefficient witnesses, dirty/displaced work, public/source
preservation and cleanup. Nine exact width-increment controls isolate
the symbolic coefficient. Controls detect COPY contamination, omitted
18-fold expansion, omitted residue rewinds, wrong coded-read units and
the fourth quotient move. Finite checks support the control-flow proof;
no full numerical caller is replayed.

Limitations and credit

Global foundational assumptions remain premises, and the selected
certificate is unchanged. All LAMBDA/native/SCALE/compiler, rounding,
source access, owned history and recursive costs still need a complete
physical upper invoice. No improvement in accepted kappa is asserted.

Prepared for @CubeJerry, with substantial OpenAI assistance.
Credit DreamingOfClouds' #6936 and #6937, the
canonical stream/finite-precompilation authors, and the originating
CubeJerry lifecycle, compact ADD, QUERY-only, precision and rounded
service packages. Separate AI review is not human or proof-kernel acceptance.