swayfx-mosaic/tooling/probes/step20c-effects-damage/README.md

STEP20C effects and damage source fixture

EXPERIMENTAL / PARTIALLY_FEASIBLE for the measured matrix. This fixture only builds and runs disposable private headless copies. It does not install or modify SwayFX, SceneFX, wlroots, the daily IPC socket, the cumulative Mosaic production patch or PKGBUILD.

Pinned inputs and private adaptation

Start from a full clone of this repository containing commit 542d9b9c945b22cf9134a4a643c8dc763b34607d. prepare.sh delegates to the preserved temporal-damage preparer, which pins private Mosaic cd550382b311efdbfe3ed892f4e1d281d852e754, SceneFX 37ccd723bef49e6891156ffafce8f549f01446cc and wlroots 0.20.2. The preserved presentation-identity patch then observes native output commit/present callbacks in the private Mosaic copy. It then applies effect-glass.patch only to the lab copy of the pre-existing nested Wayland SHM client and compiles it as C17. That patch changes the top buffer to premultiplied ARGB 0x88442010 so the underlying 64×64 wallpaper's 8-pixel #f5f5f5 / #0f192d stripes remain visible. The opaque buffer path stays unchanged. The patch does not touch compositor rendering or scheduling. The normal private clip and damage patches are those preserved by the earlier fixtures. For a separate passive-node trace, set STEP20C_EFFECT_TRACE=1 when calling prepare.sh. This applies effect-trace.patch only to the private SceneFX copy and recompiles it. Its FX_TRACE records node type, enabled flag, geometry, draw region and optimized dirty flag. Compare its outputs with the default untraced run before using it as visual evidence.

Reproduction

From the full clone, with user-local build dependencies already available:

sh tooling/probes/step20c-effects-damage/prepare.sh
python tooling/probes/step20c-effects-damage/run.py "$LAB" control --camera-pairs 0
python tooling/probes/step20c-effects-damage/run.py "$LAB" private_standard --camera-pairs 0
python tooling/probes/step20c-effects-damage/run.py "$LAB" private_mosaic --camera-pairs 1
python tooling/probes/step20c-effects-damage/analyze.py "$LAB" \
  "$LAB"/runs/control-effects-* \
  "$LAB"/runs/private_standard-effects-* \
  "$LAB"/runs/private_mosaic-effects-*

Read the exact LAB= path printed by prepare.sh; the code block uses it as a shell variable for brevity. Pass only the chosen three run directories to the analyzer when exploratory runs also exist. The runner uses the lab's private runtime, private discovered SWAYSOCK, 120-second deadline, 40-capture and 35-MB limits. It uses ordinary IPC commands; it never addresses the desktop session. run.py builds a local adaptation of the preserved applied-state runner inside the lab; the generated Python file is an artifact, not a tracked source.

The geometry is HEADLESS-1 (-800,0), HEADLESS-2 (0,0), both 800×600, scale 1, normal transform. HEADLESS-2 has a floated wev reference. The owner is the same nested SHM client in all three modes. gaps inner 24 is applied in all three modes. The standard layouts move the owner to the right edge, so ordinary shadow overflow enters HEADLESS-2. Mosaic uses scroll_x 400, then at most three 400→1280→400 camera pairs. Captures are lossless PNG and all coordinates in them are output-local. The reference keeps its natural floating size at (250,250); an earlier late-resize variant produced inconsistent reference rendering and is retained as a fixture limitation. --camera-pairs 0 allows native-event or instrumentation controls without repeating the three camera pairs. --reference-resize reproduces that retained variant, including its known limitation. The original camera campaign used it; the final native event supplement kept the natural reference size and used zero pairs. The final runner counts camera pairs across the lab's retained runs and locks the campaign against concurrent runners. A fresh lab permits at most three pairs in total; subsequent trace/native controls must use --camera-pairs 0. The executed ticket predates this guard and actually ran 15 pairs over reproduction/trace variants, exceeding the requested three-pair bound. All such runs are retained and manifested; this protocol deviation limits the campaign qualification and is not hidden by selecting a three-pair subset.

FX0 disables shadow and blur. FX1 enables shadow with shadow_blur_radius 16, shadow_offset 30 0, and shadow_color #000000A0. FX2 disables the shadow and enables blur with radius 5 and three passes. FX3 additionally enables blur_xray; this is only an optimized-blur candidate until a dirty/cache transition is observed. FX4 enables both effects. T6 toggles and closes the ordinary owner; T7 exits Mosaic through layout splith.

The analyzer records per-capture SHA-256, 800×600 dimensions, monotonic capture windows, output-local fixed ROIs, whole-image and ROI differences, per-transition pixel/damage comparison, and a complete raw-evidence SHA-256 manifest in $LAB/effects-damage-results/. A control without private damage instrumentation records damage coverage as unknown. Its manifest covers all files in the selected run directories plus the wallpaper PNG; source hashes are recorded separately in the report. Old exploratory runs are not silently substituted for selected comparable controls. --include-all-raw-runs includes every retained run in the evidence manifest while measuring only the explicitly chosen comparable runs.

Measured result and limits

In the selected three-run matrix, both standard controls show a 27,594-pixel shadow overflow in the first 46 columns of HEADLESS-2. Mosaic changes zero pixels there. FX1 is invisible on the sampled Mosaic interior, even after adding gaps, so its permitted interior rendering is INCONCLUSIVE. FX2 produces 112,308 changed pixels on the standard HEADLESS-1 images and 319,734 on Mosaic; contrast of the eight-pixel stripe edges falls markedly, establishing a real blur beyond alpha compositing. HEADLESS-2 stays unchanged by that Mosaic blur. Three camera returns exactly reproduce the pre-hide FX4 PNGs. Private damage extents cover the pixels changed between selected settled captures, but are not proof of minimum damage or every intermediate frame. The tiled FX2 path can already select optimized blur when containers cannot overlap; FX3's blur_xray therefore does not necessarily choose a different path. Optimized nodes and dirty=1/0 entries were observed, but independent cache invalidation correctness, T3 out-of-viewport shader sampling, popup effects and saved buffers remain NOT TESTED. The report also retains an instrumentation-neutrality limitation: enabled shadow (type 3), optimized blur (type 4) and regular blur (type 5) entries were traced. Standard-layout PNGs matched the untraced run, but traced Mosaic omitted the visible wev reference despite its mapped IPC tree entry. The primary pixel claims use the untraced matrix. The report contains the numerical evidence and the exact lab and manifest hashes.

Freeze V1 independent replay

The 2026-10-12 audit rebuilt from /tmp/step20c-effects-freeze-v1-clone at the pinned repository HEAD into /tmp/swayfx-mosaic-nested-grab-xbAPurbe. The commands above consumed one pair total, with 18/18/24 captures and no fixture failure. Historical evidence remains separate and nonconforming. The reference-geometry audit found 640×480 versus 300×250 in the selected untraced/traced Mosaic runs; instrumentation causality remains unknown. See the report for verified historical and new manifest hashes and limits.