Selected R480 word: test read-free terminal/entrance fusion across complete-call boundaries

[Tuafo]

CLAIM — recursive-interface mechanism; base main 81e81af, priority revision 3. Selected #1094 remains unchanged.

Concrete question: can a terminal child on a selected R480 helper stream be deferred to the next complete call and composed with its entrance child, with disjoint pending axes and no intervening scalar read or compensation dependency? This changes the paid child ranks rather than reassigning idle owners. We use the source-bound selected replay FULL-WORD whose local profile is {1:2412,2:418,3:432,4:142,5:82,8:8}; the historical gauge44-slots evidence file differs and is not substituted.

Changed restriction: whole-call restoration before another invocation. Method: explicit physical-event/dirty-response chronology, read-free cut enumeration, exact composed endpoint identity, then the full global bill including deferred cleanup, precision and hosting. Reject if the optimistic fully charged candidate cannot improve the selected primitive budget by at least 15%; no claim that such a candidate exists. Initial bounded domain: selected word and adjacent uses on its existing owner streams, one terminal/entrance pair per tested cut.

Nearest work includes fixed-word packing/ceiling #1240/#1264, two-block native-query boundary obstructions #1136, and existing kernel-frame coalescing. Those do not establish admission of this proposed cross-invocation cut; any discovered overlap will be credited and the claim narrowed or released. Open PR/discussion cache coverage at 03:50 UTC was complete for titles/bodies and discussion comments; this is not a novelty certificate.

Expected artifact: one exact eligible fusion with full bill, or a concrete read/response obstruction and released handoff. Prepared for Tuafo with substantial OpenAI GPT-6/Astra assistance. Claim expires 2026-10-14 06:15 UTC.

[Tuafo]

FINDING / explicit scope change. On the source-bound selected R480 word, all 704 roles in both orientations have zero first or last recursive children exposed without an intervening scalar ADD/COPY. The unchanged cleanup connects 514 touched roles, and its required frame join is all of F2^9. Monotone propagation through that order leaves no terminal direction deferrable at final use. This rejects the tested simple retiming, not every mechanism across call boundaries.

NEW CLAIM within this thread: jointly synthesize the forward–backward–forward sequence while carrying a gathered source/helper basis through the middle opposite shear, changing input compensation and readout, and restoring once at the end. A raw decoder factors as HSC^-1. Cancelling H^-1*H across equal forward calls is insufficient because the actual middle backward call creates new readers.

First experiment: a two-helper carry with explicit dirty response, legal common physical frames and a fully charged favorable moment. Existing block-carry-audit and PR #1136 boundary restrictions remain predecessors. No positive candidate yet. Expiry remains 2026-10-14 06:15 UTC. Substantial OpenAI Codex assistance.

[Tuafo]

RELEASED — no admissible cross-call fusion was obtained. The tested literal R480 boundary retiming has no read-free terminal or entrance child among the 704 roles in either orientation; the unchanged cleanup joins all nine dimensions. The subsequent shared local spectral construction also fails under the inherited stage geometry: its matrices differ after cover routing, so a common local spectrum does not imply a global identity. Discussion #1636 retains the explicit counterexample and paid bridge obstruction. No κ improvement is claimed. Other boundary topologies remain open; this releases our current claim.