swayfx-mosaic/tooling/probes/step20c-optimized-blur-edge/README.md

Optimized blur edge causality fixture

EXPERIMENTAL / PRESERVED EVIDENCE; historical campaign closed. Read the pre-registered protocol. The historical V1 cache campaign is closed; never launch or analyze into it. This fixture is a new four-slot comparative campaign, not a renderer fix.

From a full project clone containing the pinned project commits, with Meson, Ninja, C17, Wayland scanner, compatible wlroots 0.20 build dependencies, Python numpy/Pillow, grim and swaymsg, use:

PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-optimized-blur-edge/test-budget.py
PYTHONDONTWRITEBYTECODE=1 python tooling/probes/step20c-optimized-blur-edge/prepare.py

Preparation builds only, creates a new LAB outside Git and retrieves pinned SceneFX. It reads the published V1 XML/patches and older preserved patch chain without modifying them. Compiler logs and all generated files stay in LAB. Offline callback, source, AST, ledger, patch, build and linker preflight must pass before any run. The client ACKs background configures without drawing; only explicit A/B/A commands can submit its three 800×600 images.

Execute run.py LAB R0, then R1, R2 and R3 exactly once each, with PYTHONDONTWRITEBYTECODE=1. The persistent locked ~/swayfx-mosaic-optimized-edge-v1-budget.json binds all four slots to one LAB. Failed launch/action reservations count; no refund, reset or alternate LAB. Each run has 3 images, 1 scoped blur cycle, 12 PNG maximum and a 150 s watchdog. All sockets/IPC are explicitly private; cleanup targets recorded children only. R2/R3 share binary/client/library/config; OPT_TRACE absent versus 1 toggles new cache traces. Old private traces remain enabled in both. No GPU readback, forced clearing, shader, algorithm, production, package or desktop changes.

Keep exact executed sources, configurations, binary/library identities, client events, logs, PNGs, exit codes and manifests outside Git. Analyze all RGB pixels without excluding edge bands. Model/source sampling possibilities are not native causality. Results and final offline validation are recorded in the protocol after the four attempts. Fixture preservation is authorized; no new graphical run is authorized.

Closed V1 campaign

LAB /home/fy59/swayfx-mosaic-optimized-edge-a55yjb8n consumed all four runs, 12 images,48 captures,4 cycles and two initial Mosaic corrections; no rerun is allowed. All controls completed. R0/R1 full RGB and R2/R3 full RGB match in six states, while the edge anomaly persists in pristine and without OPT. Result PARTIALLY_FEASIBLE, architecture UNDECIDED; native FBO cause unresolved. The next justified distinct gate is ABI/package audit with this limitation. A proposed selective GPU observation is documented but not executed.

Offline analysis in this new LAB is analyze.py LAB then model.py LAB. model.py reuses exact V1 functions through AST selection without running the historical writer. Native rebuild schedule exists only for R3. preflight.py affects offline files only and must precede all graphical runs; no re-execution of the exhausted run slots is authorized. Source/evidence manifest creation is separate from analysis to preserve final inventories. Generated outputs remain outside Git. All runtime/socket directories were removed and owned children reaped; compositor exit=0, clients retain controlled-shutdown EPIPE.

seal.py LAB creates each final manifest once and refuses replacement; verify its entries rather than regenerating after sealing. Optional plot.py LAB requires matplotlib and produces a standalone SVG from complete column/row arrays; it is an offline figure, not another captured PNG. Its cache stays in LAB. Run all derived analysis before sealing.

Freeze verification (2026-10-13)

All 1,020 source and 189 evidence entries were independently reverified before publication. Their original manifest bytes remain unchanged. The historical manifest entry for this README (SHA-256 e9b4ea06512f1cbcef6107eecaa356d02767f01bb66b0a6cb3b797d8094b7d4c) must now be checked against README.pre-freeze.md, the exact pre-publication snapshot; every other entry uses its recorded path. Source manifest SHA-256: 044f6ab4e0d9ced000b12daa75c7cf30f234aa32fed14dc248baa9e3e11dcdff. Evidence manifest SHA-256: 7554388b89fab03b5f3ad0d15c792680fcb403ee04bce8a347c55257440c73d8. Verification must never reseal, rewrite results or rerun the exhausted campaign.