swayfx-mosaic/docs/design/STEP20C_MINIMAL_CORE_GRAPHICS_QUALIFICATION.md

STEP20C — minimal private core graphics qualification

EXPERIMENTAL — 2026-10-14. Ticket SWAYFX_MOSAIC_STEP20C_MINIMAL_CORE_BUILD_RUNTIME_AND_GPU_FIELD_GATE_V1. Eleven E runs executed once; measurement, raw runner exit, terminal IPC and client closure remain separate. Evidence delta review ACCEPTED for experimental fidelity; final dossier review pending at this edition. Documentary publication PREPARED. STEP20C NEXT; production multi-output NOT YET SUPPORTED.

LAB is /home/fy59/swayfx-mosaic-minimal-core-2l7riqt_/, outside Git. Paths below identify private local evidence, not a self-contained public distribution. The build/C report records the exact parents, private repair, unit tests and build. The integration decision records aggregate limits and remaining gates. Historical reports remain immutable.

Exact identities and preflight

The core is LAB/mosaic-build/sway/swayfx-mosaic, SHA 256 29baf31aa3eeed259825c87d0799fa5d3dc83710afb967154604eb1c6657f965, Build ID 1b42e557e38db3440ae8dcf7ac967e4132223365. The observed wlroots library is /usr/lib/libwlroots-0.20.so, SHA 256 7edca9a7d65855addda68e29ba5db94d57c803eb5fa7d0f86db3136b90ca9837. Core observations do not qualify A/G, historical diagnostics or the distinct new GPU diagnostic. E0-pristine uses the pristine control, not the core.

Initial preflight accepted 1,095 files, manifest 9c75d2249a45d2a9cbdbdb3e0cbff91db82bb89b14d6f580ca12a9ed1ed0b705. All three initial attempts failed before captures on AttributeError: 'NoneType' object has no attribute 'endswith' while inspecting an ancestor's floating field. They spent their reservations and were not retried. Offline fixture repairs and reviews before the remaining cases produced the R2 manifest 5e49909975ea6776a677280a4e444d5014699bf36b68c68fba3345626b196fd3 with 1,097 files and review 0e7525b69fdd3e3f75b5c62fb0c3c833557024c6153c3ee033c2d9eebe9d66d0 ACCEPTED. This is fixture repair before new reservations, not a core algorithm change or a rerun of E0/vertical.

Private headless runtime directories, Wayland displays and explicit IPC sockets were used. Actual logs identify the EGL/Mesa/radeonsi renderer; a requested renderer environment variable alone is not treated as proof. Each scenario has its own config, startup/maps/output observations, event log, receipt and raw stdout/stderr/exit under LAB/graphics/runs/<case>/run/ and the execution folders.

Executed matrix

Capture counts are output-capture debits. The fractional and R90 cases each also retain one marker-source PNG, excluded from capture counts.

Case	Raw runner exit	Measurement outcome	Output captures	Retained boundary
E0-pristine	1	INCONCLUSIVE	0	Ancestor fixture failure before comparable standard control
E0-core	1	INCONCLUSIVE	0	Same failure before standard-layout measurement
E2-vertical	1	INCONCLUSIVE	0	Same failure before vertical pixels/input
E1	1	INCONCLUSIVE	6	First chain sampled; second ancestry fails before standard return
E2-fractional	0	PASS, sampled	9	Scale1.25 grabs/pixels/input/dismissal; marker teardown timeout retained
E2-r90	0	PASS, sampled	9	Scale1.25/R90 grabs/pixels/input/dismissal; same teardown limit
E3	0	PASS, sampled	17	Resize/clamp and same-owner workspace migration/input
E4	1	INCONCLUSIVE	12	Actual current fullscreen ancestor; comparable standard control absent
E5	1	NOT_EXERCISED for full scenario	10	Hidden-owner pixels/input partial; same-chain return absent after dismissal
E6	1	INCONCLUSIVE	8	Natural timeout observed; frozen held-serial validator rejects receipt
E7	0	PARTIALLY_FEASIBLE	14	Classic blur exercised; whole-H1 return FAIL; shadow unexercised

The E matrix spends 11 runs and 85 output captures, retains 87 PNG files, and does not contain a successful new pristine standard control. All E terminal IPC commands return 1 while the compositor exits 0. Raw R2 runner exits are four 1 / four 0, in addition to the three initial exits 1. The separate LAB/gpu-execution/r2-count-discrepancy.json corrects an earlier aggregate progress statement of three/five; per-run raw files and historical summaries remain unchanged. A runner exit 0 is not a universal terminal-IPC PASS.

Popup geometry, input and remaining transitions

E1 uses two horizontal headless outputs with the Mosaic owner on H1 and a whole-H2 stable reference. Its first TOPLEVEL→A→B chain has real wl_pointer.button grab serials 31 and 40, native A350×170 and B300×130. A is at (-218,105,350,170), B at (-118,165,300,130), both actually cross x=0. Sampled H1/H2 color counts are owner 142,794/0, A24,080/0, B15,340/0. Outside input yields popup_done and destruction B then A, with no invisible popup button delivered. The second owner opens on workspace 2/H2/splith rather than the intended Mosaic ancestry; validation stops before the standard-return path. That transition, B-close/A-survival and final standard input are NOT EXERCISED in this core campaign. First-chain evidence is PARTIALLY_FEASIBLE; the global E1 result remains INCONCLUSIVE.

E2-fractional has H1 physical 1000×750, scale 1.25, logical 800×600 at x=-800; H2 is800×600 scale1. E2-r90 has H1 physical 750×1000, scale 1.25, transform90, the same logical output and negative origin, with asymmetric marker evidence. Each has real grab serials 33/42 and the same logical A/B crossing geometry. H1/H2 color counts are owner 223,038/0, A37,650/0, B23,961/0. Visible input, forbidden popup input and sampled outside dismissal pass; reference button pairs are retained. These samples do not promise universal click-through.

E3 resizes H1 logically800→960→800 and moves the same workspace 6 and owner 10 H1→H2→H1. The clamp is960 after relative request1920 with three children and gaps0; the existing 640 minimum is an experimental source fact, not final UX sizing. Four input checks pass at recorded global points (-273,168), (-275,168), (366,168) and (-434,168). Native grab serials 40/49 persist in the sampled chain. No universal transaction atomicity follows.

E4 observes owner 9 with fullscreen0 beneath non-leaf container 10 with current fullscreen1; workspace 6 also reports a fullscreen marker. The non-leaf actual ancestor, rather than that workspace marker alone, establishes the exercised ascendance. Real grab serials 49/58 and nested captures/input are retained. Without a comparable standard fullscreen control, exemption/cross-output policy remains INCONCLUSIVE; an absent clip is not classified as a leak by fiat.

E5 samples hidden-owner masking and input, then the outside click legitimately causes popup_done and destroys A/B before camera restoration. The same-chain return is NOT EXERCISED. This does not settle hide/dismiss/reposition UX. E6's timeout and E7's rendering limits are detailed in the lifetime/damage report.

Closure, review and reproduction boundary

Every E runner receipt labels scientific_cleanup=PASS and records private runtime removal. That label is the runner's process-recovery gate, not proof that every client closed normally with Wayland error 0. E1/E6 normal client closure is not acquired. E2 owners report error 0, but reference/filler helpers report error 32 after compositor stop; each marker swaybg reaches a three-second cleanup timeout. E3/E4/E5/E7 normal closure is attested for the relevant clients; E7's three FX clients terminate error 0 normally before compositor shutdown. Mechanical cleanup and protocol/normal closure therefore remain separate.

LAB/graphics-execution/summary.json, LAB/graphics-execution-R2/summary.json and their per-run manifests preserve hashes and raw observations. Existing commands are recorded as python -B graphics/run.py <case>; they are provenance, not instructions to rerun a consumed reservation. Any future experiment needs its own accepted scope, fixture gate and budget. The evidence delta review checks 319 hashes without divergences and accepts fidelity with E/F qualification PARTIAL. Final independent dossier review is pending; publication is PREPARED at this edition, with actual publication identities reserved for the private post-push receipt and final handoff. No production source, UX or package change.