Native centre sharing: zero first transcript, later query 1/4

[navinor]

Result

Scoped native counterexample; no new supplier or κ. Answers open question 3
of #2199, head
30ec048eedcaca970f94bbec0fdd3c52eacf9fdc.

The scalar D-only identity backward1 centre = forward0 centre / 4 does not
lift to the actual cover-owner action for selected replica 0, calls 0 and 4.
The counterexample rules out more than identical raw carriers:

• At physical cover g=rho_1(port0), put payload-constant D-field differences
  (1,-1,-1,-1) on ports (0,1,6,7); all other differences are zero.
• All eight first-call centre fields vanish on the entire cover, before
  and after transport. The complete first native word leaves those four
  constants on D and adds them to T, with zero helper/private differences.
• At the later common owner I, the actual stage-1 routes expose only port0.
  Its channel-0 native input is exactly 1/4; the post-transport amplitude at
  address 0 is i/4, and inverse carrier normalization returns 1/4.

Thus no linear decoder of those eight saved full-cover fields alone, even
with arbitrary owner mixing and native transport, recovers this later query.
This is not a recursive-call tariff, a full native rank calculation, an
exclusion of additional observations/changed contracts, or an output-quotient
bound for overlapping scatters.

Scope and evidence

Adds only research/selected-centre-quarter-native/ (seven files).
The README gives the explicit eight-by-four column relation and kernel proof.
Each local orientation's 24,767 literal events is executed with unpacked
exact Gaussian-dyadic statevectors on 512 local addresses, retaining native
children, residual/endpoint adapters, scalar signs and private clears.

The payload is constant on all 73 physical axes/tensor/spectator factors at
one cover owner; it is not an address-basis column. A legal restricted input
can falsify an all-input identity. No equality is inferred from constant-input
tests. Distinct arbitrary cover-owner arrays are retained, never identified.
Initial target/helper dirt is the same arbitrary baseline in both executions;
linearity justifies zero differences without assuming clean workspace.

Actual stage-0/1 owner/chart/translation recipes are bound and reconstructed.
Calls 1–3 touch none of the live witness families, and the first five scheduled
events are local sweeps, so no bridge or idle event is silently dropped.
Both complete local endpoints and pre/post carrier values are checked.

Positive controls reproduce PR 2199's scalar quarter-factor on all 8×112 D
coefficients and the cancellation on cover-invariant inputs. Seven elementary
controls cover exact roots, phases, inverses and typed receipts. Four native
controls cover the cover-invariant case, wrong owner location, wrong backward
global phase and wrong gather coefficient. CLI tests generate a fresh receipt
and reject corrupted/wrong-typed responses under -I -O -B.

No native executor code is imported from pending PRs. The pinned Clifford
kernel supplies canonical circuits only for inverse normalization; the literal
word evaluator is new code with the retained conventions. Six selected archive
members are hashed before decoding, using unique full paths. Source Git blobs,
digests, all 73 columns of g, route digests and responses are retained.
No dependency installation, frozen-source edit or selected-record change.

Separate read-only AI review independently reconstructed the four-column
relation and all four 73-column routes without importing this verifier, and
obtained 1/4 from a separate backward causal-slice replay. A separate packed
evaluator checked the used centre transport and eight active endpoint adapters
on all 512 columns, phase included. No critical or important defect was found.
The review requested an explicit coordinate-covariance argument for the
nonconstant intermediate states at owner g; the README now includes it.
The global lift uses the retained native local-view embedding, not a newly
rebuilt physical address/stream compiler.

Verification

python3.12 -B scripts/research_package.py --package research/selected-centre-quarter-native --check-only
python3.12 -B scripts/research_package.py --package research/selected-centre-quarter-native

Local Windows / Python 3.12:

PASS_PACKAGE_METADATA_AND_PINS (not a mathematical acceptance check)
PASS_SELECTED_CENTRE_QUARTER_NATIVE_COUNTEREXAMPLE first_transcript=0 later_query=1/4 no_new_kappa
Ran 2 tests in 24.940s
OK
PASS_REGISTERED_RESEARCH_PACKAGE research/selected-centre-quarter-native outcome=completed

No hosted CI pass, independent human review or formal-kernel execution is
claimed. No full cold replay, global hosting/stream/compiler/precision invoice,
outer proof, supplier adoption or κ change. All inherited obligations remain.

Assignment and overlap

Aaron Wang / navinor, with help from chatgpt (substantial research, math,
implementation, writing and separate read-only AI review).

• Recursive-interfaces, card centre-path-sharing, scope refined to one
  necessary information-reuse implication, not a sharing construction.
• OS nonce 56c03933c3a237bef454593979c3b883, catalogue key
  184695337cedab03152e, priority revision 4.
• Base/current canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2;
  selected record hash 10b9e340fd8e354a37a0c4813c90e5de1dd42044b5f6aca4a3fc5a42b685fa32.
• Claim posted before computation:
  Ten-instance independent research session: interface and verifier audits #1884 (comment).
• Canonical main fetched 2026-10-15T07:08:24Z; instruction/priority/selected
  compatibility diff empty. Text-based overlap refreshed 07:10:09Z, complete
  pagination: 588 open PRs / 1116 discussion records. Only our claim and PR 2199's
  handoff matched; no novelty guarantee. Both closest PR heads were rechecked.

The prescribed ceiling and ideal lever table were replayed: baseline
a_cap≈0.001690054109089, ideal two-call sharing about 1.90×. These are prior
planning numbers. This contribution removes no centre/copy/restoration charge
and proposes no changed moment. Any repair still owes additional information,
storage/transport, hosting, restoration, precision and a complete invoice.

Credit and handoff

Sagar Thakrar / #2199:
scalar identity and missing native implication; Adam Gall /
#1918: composed scalar
projection and transcript context; commdata2338 /
#2048
(a0e55f357344f1742f72942f94ae2a6f13ff7065): disjoint-forward-pair native lift
and unpacked statevector method. That pair is not rechecked or superseded here.
Retain Sagar Thakrar's exact Clifford kernel, the bridge/cover-action proof,
Rohan Arun/Jeremy Kothe/Hao Yan selected lineage, icekylinx's compiler/emitter
and DreamingOfClouds' frame proof lineage. New work Apache-2.0.

Next: test one explicitly priced observation against this four-port witness,
then prove complete ordered recovery and native access if it distinguishes it.
Alternatively, PR 2199 question 2's demanded final-output quotient is a different
contract; it still needs physical bridge/idle and overlapping-panel semantics.

Workflow feedback: live handoffs avoided repeating the newly excluded QQ16
dense-B0 repair and the actively claimed 66-port test. The useful delta here is
the cover-owner action; a mere difference in raw carrier frames would not have
established missing information. Cleanup follows publication at the donor's
request.