Native wetland bank and recovery checks

WOG-A-003578 remains the exact original 1254×1254 damp-earth PNG, SHA-256 276fdaf467c7b23361d8aa2a110b939faa0ce3c5347636c01313827176925c7c. The bank comparison changes only its private sampling frequency from 4 to 8 at the same 1.25-hex logical span. Source pixels, normal controls, lake water, source permissions and canonical geography remain unchanged.

The natural site is lake cell 1047295 beside wetland cell 1047294. Both are verified against the effective disposable QA world using a read-only engine probe. Sixteen desktop/phone Region/Hex views cover GPU and actual main-page WebGL2-unavailable CPU fallback, with current and wider sampling in each backend. All 12 candidate pan-return/shared-view reconnect maps are pixel-exact. Each backend has four source/cohort map pairs with actual changed pixels outside UI; material requests alone are not used as appearance evidence.

Direct inspection of both native sheets shows the same continuous bank, with fine painted mineral, moss and stem detail. The wider sampler also exposes conspicuous leaf/moss groupings in Hex. The earlier strata-site preference therefore does not establish a general wetland improvement. Keep the original and current configuration preferred; wetness cues, physical grain, reed/ground contact, all states and perceived nonrepetition remain open. Earlier raster discrepancies and failed alternatives remain in the individual histories.

The eight complete GPU/CPU map comparisons stream 16,761,344 pixels through memory. 8,828,693 pixels differ, with maximum channel changes of 3 to 8. These are not exact backend-parity results. Two unresampled native contact sheets are retained; full maps and RGBA buffers are not saved.

A separate material-API test uses the actual native wetland image rather than a synthetic pattern. It completes one GPU tile, prepares the next, then forces real WebGL context loss. No partial output is published. The GPU tile is discarded and the complete 1916×1768 output recovers through CPU. All 3,387,488 pixels match an independent literal-frequency-8 CPU reference exactly; source image bytes and decoded pixels remain unchanged. The test proves this source/extent recovery case, not whole-client context-loss behavior, all hardware or performance acceptance.

No original artwork, selected final, renderer, source contract, gameplay rule or production save is changed. The previous three ground findings stay queued, with the wetland record now carrying this newer source-bound evidence. No complete game-ready approval is added.

See the visual decision visual-review.json, full backend comparisons backend-pair.json, native context-loss result context-loss-validation.json, GPU observations gpu/results.json, CPU observations cpu/results.json, GPU derivation gpu/driver-derivation.json, context-loss derivation context-loss-driver-derivation.json and registration registration.json.