Lake sampler experiment: retain the artwork, reject the patterned result

The wider source sampler is not selected. Direct inspection of eight normal-client desktop/phone Region/Hex frames still shows similar bright wave bands; the broader sampling makes clusters denser rather than solving the visible pattern. The current lake study WOG-A-009279 and both originals are retained unchanged. No source is admitted and no renderer is activated.

The experiment preserves source-coordinate slope, flow direction, macro span and native artist bytes. It doubles blend lattice frequency and halves each patch footprint and its protective edge margin, broadening randomized centers without the extra clamping a simple margin-only change would introduce. This mathematical property does not prove visual quality. The same single GPU cache, working pixel budget and source-memory bounds remain; full hardware, memory, water-contact, state and lifecycle qualification is still open.

The first implementation changed 1,055 default ocean pixels and was rejected. After preserving the original default fragment arithmetic, all 19 current terrain inputs pass 95 GPU comparisons across 243,081,117 pixels. Seven CPU behavioral checks pass. Real WebGL context-loss recovery discards an already completed GPU tile, publishes no partial surface, and matches an independent constant-frequency CPU reference across 3,387,488 pixels.

The default control reproduces three of four previous whole-map hashes. The phone Hex crop diagnostic finds 410 differences of at most one channel level, confined to the visible top-right Hex-control area. Full-map pixels were compared in memory and not retained, so this bounded crop cannot establish the cause of every full-map difference. Preserve that unresolved composite gate.

The desktop context/lake-before-after-1440.png and phone context/lake-before-after-390.png boards show default4 on the left and rejected wide8 on the right, using native screenshot crops without resampling. These two bounded witnesses document the failed visual gate. No raw pixel caches or new artist-image variants were created. See the source-bound review visual-review.json, default GPU proof default-pixel-validation-v2.json, context-loss recovery context-loss-validation.json, CPU checks cpu-behavior-validation.json, actual client observations context/results.json, phone control diagnostic phone-hex-control-crop-diagnostic.json, and rejected uniform-arithmetic receipt uniform-default-regression.json.

Finite-domain density check: the natural lake component containing cell 262807 spans 568 connected lake cells and about 9,193 by 1,100 world units. At the current source density, a single image covering its bounding rectangle would need at least 74,112×8,869 pixels. Stretching the current 1,536×1,024 painting across that component would lose native detail. Do not use that shortcut or allocate such a giant bitmap as a supposed performance fix. Further native artwork/mapping studies must retain bounded working surfaces and native grain. This is a read-only natural-kernel analysis; effective-world contact and state still need independent checks. See the exact density analysis finite-domain-density-analysis.json.