GL gesture final socket ordering — 2026-11-20

The strict GL group now takes its pre-DOWN whole-frame capture before native prepare. Its transport owns the single final socket comparison. The strict controller supplies a synchronous publication callback: after that comparison and its synchronous evidence writes, the callback checks cancellation, the original native deadline and the independent release window, then schedules UP relative to actual DOWN publication. The group no longer performs a duplicate socket capture. Four distinct independent UP participants remain READY and armed before the first DOWN; their original finite fallback is unchanged.

The strict controller requires native-whole-frame-equivalence-v1. It does not select the Hex manual-context branch. The caller must pass the publication callback through dispatch(edge, publish) to equivalence.fence(handle, edge, image, publish) at the actual socket send, without an intervening await. Omitting that callback closes the controller. Private proof validation, one-use retirement, identity checks and complete exact RGBA comparisons remain in force. Source 30s, review 15s, fresh 5s, native 500ms and maximum hold 1500ms remain unchanged.

Run: scratch/runs/20261120-gltron-final-socket-release-order-native-round/. The exact loader retained Git-pinned 7bca66241 common equivalence and Hex bootstrap helpers, with the owned GL strict-controller overlay. It served the same 10bdae2c… 2,418,815-byte module and 34c794d5… graph with committed 72dc77da / cf181231 capture overlays. Preflight validated the actual dynamic controller/socket joins, 26 helper closure, five compiled export signatures, actual bootstrap lazy body and 92 media pins before browser launch. Current Hex, GPS and held-region changes were not served.

Actual native partial result: Space DOWN socket at 1795175065640, native USER return CONSUMED at 1795175065655, UP due 1795175065710, trusted UP at 1795175065713, ACK at 1795175065716. Exactly one final socket comparison preceded that DOWN. This repairs the previous run's UP-before-DOWN ordering. The next pre-DOWN comparison refused whole-frame RGBA differs at 1795175065879, before a second prepare or DOWN. The scope closed with one of four edges completed, zero subsequent ordinary requests or socket sends, no replay and no conditional original launch. Four-edge native qualification, visible trail/heading, pause/resume, round and ordinary Quit remain unproved. The unchanged strict whole-frame contract cannot authorize later edges after resume changes the reviewed image.

Evidence limitation: the inherited run adapter did not install the optional synchronous capture evidence sink. It retained the reviewed complete PNG and fresh capture digests, but not the actual final-capture PNG or the rejected pixels. No post-UNKNOWN capture was requested. The precise changed pixels cannot be diagnosed from this run. This is partial ordering evidence, not complete native/game qualification.

Focused tests exercise publication-relative UP, delayed final capture, release race, finite fallback, cancellation, wrong target and mixed prepare ACK refusal, four private proofs, one-use retirement, complete-frame mismatch, actual child cancellation and independent UP joins. Authored endpoint tests do not substitute for native USER qualification. Required full build and publication receipts are retained in the run. All owned runtime participants joined and were nonlive before verified evidence retention and tmpfs cleanup. No remote box was created: ID null, state NOT_CREATED, expiry null, stop receipt null.