[zxv]

Coupled provider #5402 currently returns eight helper outputs that cannot receive ordinary Gaussian credit. This tests their first two actual FULL/QUERY consumers and identifies a concrete, charged exit contract rather than decoding automatically at recursive return.

The exact trace retains both query responses and the second helper's cross-contamination. The proposed eight-mode-only feed into the next FULL would require W3 T = D1 W1; a phase-correct 17-bit residual Gram has rank eight, excluding that equation even for arbitrary linear T. The all-zero address passes and is retained as a false-positive control. The same fourteen consumer instructions are checked against #5398's successor caller.

A constructive native-return exit needs the versioned eight-field response b=ZQY. If actually supplied, it restores all original outputs with 8 H1 calls, 456 ADDs and 240 monomial field maps, preserving the borrowed response. Buying a full QY response costs 120 H1 calls instead. Applied to the sixteen unwrapped backward calls, that purchase changes rank surplus 752 to -16 and fails the exact complex-saving .004 moment screen, including against #5413's declared neutral profile. Every existing thin query remains paid. The local exit alone does not repair a consumer that has already run.

This is a scoped partial result, with no new kappa, admitted recursive family, universal decoder lower bound or selected-frontier change. The provider endpoint is a pinned premise; complete observation embedding, late correction, precision, tape/storage, compiler and outer funding remain open. The eight-dimensional witness is not a paid Gaussian volume lower bound.

Validation at 39d6f18d53f073fcae75b7fd98055060cb648e73: registered Python 3.12.8 normal and isolated optimized replays, all three literal exit programs and semantic controls, committed-byte checks, official 16-file export/integrity, and both read-only exported replays with no Git on PATH pass. The optional ancestry preflight stopped at the shallow-history guard; its audit did not run. No full native/provider, selected cold, Lean, Docker or outer replay was run. Hosted CI is separate.

python3.12 -B scripts/research_package.py --package research/coupled-query-transport

Only research/coupled-query-transport/ is added. PROOF.md, RESULT.json, literal-program generator, exact origins, validation limits and Open next questions are included. The next useful task is an actual funded eight-field ZQY embedding, or a larger coupled type with its complete volume and eventual exit bill; more seeds inside the rejected feed equation cannot change its result.

Credit Dillon / #5402 and #4933; Ali Romman / #5383, #5398 and #5411; Chris R / #5407, #5400, #5413 and #5414; Cao Yifan, Dugongue, Cinna Davis, Ryanson Jonathan, Jacob Alan Sussman, zxv and the retained native/Clifford/compiler lineage. Full provider attribution is retained verbatim. New work by zxv with substantial OpenAI Codex assistance, no subagents; Apache-2.0.

Canonical main was refreshed and remains 1ec1119972c9aefe42525172d07b332fcff81cb4, priority revision 7. The final overlap refresh covered 363 open PRs and 4701 Discussion records; sparse local notes and unpublished work remain limits. Discussion #4930 remains the broader investigation. This complements the active caller search and does not repeat its three-call program or claim its inverse as new.