RHCERT-v0.1 Advances From Format Experiment to Repeatable Mass Production, cell68 Promoted to Publication-Candidate: Publication-Candidates Reach 5/77, cell75/76 Await Only Certificate Bytes

Round 35 (AMRAL-LUC-FC-R35, 2026-09-20) carries RHCERT-v0.1 — first implemented in Round 34 on two cells — from a format experiment into repeatable mass production: cell68 passes bytes-only replay across all 9255 of its 9255 leaves with 0 replay failures, on a raw RHCERT file of 35,940,305 bytes covering 2,126,336 rational vertices, a minimum exact margin of 2.2329×10⁻⁸>0, and an MPFR dominance audit of 1565 unique states with 0 failures — and because the bytes-only verifier performs no candidate generation at all, checking only decode, directed membership, exact convexity, and exact integer shoelace, cell68 is formally promoted to PUBLICATION-CANDIDATE-SHARD, raising the publication-candidate tier from Round 34's 4/77 to 5/77 (≈6.49%, {47,48,68,69,70}). In the same round, cell75 (10948/10948 lower PASS, marker independent upper bound and MPFR upper bound both PASS, dominance audit of 1690 states with 0 failures, minimum lower margin 6.2518×10⁻⁸, marker outer value 0.8349344874879304<0.835) and cell76 (10932/10932 lower PASS, the same class of upper-bound and dominance audits at 1688 states with 0 failures, minimum lower margin 2.0477×10⁻⁸, marker outer value 0.8349600023952688<0.835) have both closed their mathematics and cross-backend audits completely, but the document states plainly that both are currently missing only their concrete RHCERT bytes — so they remain at the "cross-backend audited, RHCERT-pending" tier rather than being counted as publication-candidates. Together with cell68, the seven cells of the Depth-30 batch — {47,48,68,69,70,75,76} — have now all completed finite topology, marker upper bounds, and full rational lower migration, contributing an arithmetic-closed-or-better tier of 7/77 (≈9.09%), while geometry-complete remains unchanged at 66/77. This round also identifies a previously unrecorded geometry-durability gap: the true pending-box states of earlier geometry continuations existed only in runtime-local checkpoints, while the durable artifacts retained only summary statistics — count, volume, depth — so a summary of the form (#pending, V_d) does not by itself imply that the frontier can be reconstructed; Round 35 states explicitly that it refuses to guess box state from a summary, and so it has not fabricated, and does not fabricate, any new geometry closure. As a remedy, this round specifies the DFRONT-v0.1 format: each pending witness box is persisted as three dyadic pairs relative to the root interval — (k_φ,m_φ), (k_x,m_x), (k_y,m_y), denoting an interval of the form [m/2^k,(m+1)/2^k] — independent of historical split order, and stored together with the parent checkpoint hash and the root dependency hash to support lossless resumption; but this round only defines the format itself — actually using DFRONT-v0.1 for real frontier recovery is work reserved for the next round (Round 36). Round 35 also lists Round 36's follow-up items: emit RHCERT for cell75 and cell76 and complete their bytes-only replay, promoting both to publication-candidate if they pass; restore the old geometry frontier bytes, or reconstruct them from the nearest durable certified ancestor; and switch all new geometry checkpoints to DFRONT-v0.1 from then on. The global bound a_Leb≥0.835 remains NOT CERTIFIED, and this round does not change that. Research direction is chosen and led by Neo.K; this round's execution was carried out by Aletheia / GPT-5.6 Sol.