[Arham-Tahir64]

The QQ16 producer handoff leaves dirty-helper magnitudes outside its completed-stage error theorem. This audit supplies one exact, restoration-preserving fault witness in the reviewed F0 common-chart scalar projection: at literal event 24438 the operands are (-4T,T), so one admissible completed relative ADD defect 8uT leaves 4uT maximum final data error. All 480 helpers restore and all eight private outputs remain zero on this witness family. Therefore local relative ADD inequalities and exact restoration alone cannot imply a bound depending only on original data size. The derived necessary reserve for the stated F0 tolerance is u <= (H/T) 2^(-beta-10).

The package pins current main 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11 (priority revision 6), includes the task card, full 52-field response, forward rational checker, separate reverse-adjoint checker and public CLI controls. This closes one precisely scoped scalar implication of Nick Petalas's PR2965 handoff; it does not contradict that theorem. Claim and scope.

The physical cutwise gauge binding remains unverified, as do native-child interiors, a particular rounding implementation, actual dirty ranges/grids, physical cost and the F17 consumer. Exact grid-preserving ADDs escape this witness. No selected frontier, source input or shared verifier changes; no physical failure or new kappa claim. README/TASK identify the next native binding and stronger numerical-contract questions. Family decision: narrow and stop.

Validation: registered CPython 3.12.14 replay passes; independent reverse checks cover all 712 suffix coordinates and two prefix rows. Six behavioral groups cover source/receipt corruption, altered cancellation sign/location, tiny relative defects, absolute-error/zero-dirty controls, optimization and a read-only Git-free export. Optimized CPython 3.14.7 tests also pass. Selected baseline/source pins, metadata, committed-byte preflight (15 package/input files) and diff checks pass. Live overlap refreshed across 237 open PRs, 2,440 closed PRs and 1,703 discussion records; canonical main compatibility checked. Full selected replay, physical binding, hosted CI and formal-kernel review are not claimed.

Credit Nick Petalas/PR2965, Ali Romman and the QQ16 team, ywz65/PR2540/2707, Arham Tahir/PR2681, and the native/geometry/supplier lineage retained in the source package, including starship-s, Sagar Thakrar, Ross Richardson and Douglas Colkitt. Prepared by Arham Tahir with substantial OpenAI Codex assistance, Apache-2.0. The two checking algorithms are by the same assisted author, not independent human review.