swayfx-mosaic/docs/design/STEP20C_GPU_FULL_FIELD_FRESHNESS_FOLLOWUP.md

STEP20C — GPU full-field freshness follow-up

EXPERIMENTAL / INCONCLUSIVE — 2026-10-14. One distinct G1 diagnostic run, no retry. The source design/preflight was accepted; its only raw epoch was rejected before any accepted GPU receipt or analysis. Final dossier review pending; documentary publication PREPARED. STEP20C NEXT; production multi-output NOT YET SUPPORTED.

Ticket: SWAYFX_MOSAIC_STEP20C_MINIMAL_CORE_BUILD_RUNTIME_AND_GPU_FIELD_GATE_V1. LAB: /home/fy59/swayfx-mosaic-minimal-core-2l7riqt_/, private and outside Git. The historical ROI report remains immutable: its108 native ROI reads support local outside-fresh-domain history, first sampled fresh divergence down-2, with whole-field causality INCONCLUSIVE. Those observations do not qualify the new core or G1.

Source question and distinct identity

LAB/gpu-design/HANDOFF.md records a source-only prospective design. Full physical 800×600 ping/pong textures remain allocated across down/up passes; fresh rasterized regions change by pass, rather than resizing those textures. The audit distinguishes geometry regions, viewport/scissor state, shader qualifiers, actual buffer formats, copy/cache and arithmetic precision. It does not treat a mediump source declaration or a precision query as a measurement of effective GPU rounding. The intended experiment extends local support observations to A/B/A fields with prospective readback/data bounds and explicit GL/FBO/pack/restoration checks. It does not modify functional shaders, clears, clamps or samplers.

The separately built/executed diagnostic has SHA 256 1905a20fa72ebcc7ec47d1584a4653654280288096a20005199eb75673ac0383, Build ID 1505e31c3dc08d2a7799c2ee9e1b2705eb61f5fe. It is not core SHA 29baf31. Source gate SHA 256 58535958c0c55b3bee758bfda413df594302122f9a9da2e1a4aedde3889b9f5f and independent source/runtime preflights are ACCEPTED for their scoped inputs. Actual maps/logs identify the diagnostic and wlroots SHA 256 7edca9a7d65855addda68e29ba5db94d57c803eb5fa7d0f86db3136b90ca9837, without a dynamic SceneFX mapping. Logs identify EGL Mesa/radeonsi, OpenGL ES3.2 and AMD Radeon780M, with renderD128 descriptors observed. Two EGL_BAD_ALLOC messages occur during eglQueryDevicesEXT before successful renderer startup; no causal connection to the later metadata rejection is demonstrated.

Sole run and immediate rejection

The recorded command python -B gpu-field/runner.py returns1 after 4.743s, Rejected: dmabuf fields. It spends one run and three output captures: two H2 references and one H1 witness. Only epoch 1/A is armed. The native log emits GPU_FIELD end epoch=1 points=9 complete=1; these are publication markers for raw data, not an accepted readback result. No B/epoch 2 or A-return/epoch 3, accepted epoch receipt or full-field analysis is produced.

Raw file	Size	SHA 256
LAB/gpu-field/runtime/epoch-1.meta	19,686 bytes	c8ca51bca694db9e21471356a4260064ed81ab7013a88d4edb4b8572044a825c
LAB/gpu-field/runtime/epoch-1.rgba	8,312,048 bytes	a6dc7fd873e1ab7e52e04625bacb3251ba8e1aad2e766811efebde7ae7932a99

The metadata has 24 records:1 begin,9 samples,11 writes,2 copies,1 end. read_success=NOT_PROVEN, GL_ERROR_QUEUE=UNOBSERVED, effective precision UNCERTAIN. File sizes/hashes establish retained integrity, not glReadPixels success, synchronization neutrality or independently complete GL restoration.

The read-only diagnosis in LAB/audit/gpu-rejection/G1-rejection.md compares the writer and frozen validator. The validator requires every DMA-BUF plane stride to be at least800×4=3200. All nine samples report DMA-BUF available, size 800×600 and planes [[3584,0],[1024,2293760]]. The first stride passes; the second1024 fails that first demonstrated rejecting predicate. Role formats match the validator: main XR24/875713112, effects AB24/875708993, cache XB24/875709016. Neither absent DMA-BUF nor a wrong role format explains this immediate rejection. This initial rejection-only audit establishes no GPU failure, driver defect or specific DCC explanation; raw pixels remain rejected.

A subsequent read-only modifier audit, separately recorded in LAB/audit/gpu-rejection/modifier-layout-audit.json, observes the same modifier 020000001046bb04 in all nine objects: AMD vendor 2, GFX11 tile version 4, DCC 1, RETILE 0, PIPE_ALIGN 1. The local libdrm header and the Linux DRM format modifier header describe main image surface in plane0 and DCC surface in plane1 for this non-retiled layout. Requiring the image stride width*4 on every plane is therefore not justified for this modifier: a fixture predicate defect is established. This interpretation neither proves global DMA-BUF validity nor GL read success and changes no raw gate. No validator repair or replay occurs.

Cleanup, status and next evidence

Compositor exits 0; background, filler0 and glass exit1 while the other tracked processes exit0. Scientific cleanup remains INCONCLUSIVE; mechanical cleanup passes with the private runtime removed and identified processes reaped. LAB/gpu-execution/G1-summary.json indexes49 artefacts with tree manifest SHA 256 ccd45c0a56c1ee72ff7bcd84711090745012911a7ecc3e0d890aa52ccb715329. The scout's read-only audit verifies these49 hashes; it does not rerun the validator or reconstruct unavailable data. No diagnostic or core retry, functional shader correction, system-library change or historical replay follows.

GPU_FULL_FIELD_DESIGN_REVIEW=PASS denotes the accepted design review receipt. GPU_FULL_FIELD_DIAGNOSTIC=INCONCLUSIVE and GPU_WHOLE_FIELD_CAUSALITY=INCONCLUSIVE denote actual evidence. A new proposed gate must first specify and review valid per-plane metadata semantics and read-success/error/restoration observability, using primary format evidence; only a separately accepted scope/budget may acquire another diagnostic matrix. No remaining slot is consumed by this documentary suggestion. The integration dossier retains PRIVATE_CORE_NEEDS_TARGETED_EVIDENCE; this rejected first epoch closes neither core graphics nor whole-field causality.