Private native-material crop coverage investigation

The current local sampler selects crop centres from only the central 17% of each image dimension. A pure in-memory prototype doubles blend-lattice frequency, halves source coordinate coefficients and halves the safety margin. This expands crop-centre coverage to 58.5% while keeping the analytic within-patch native sampling derivative equal to 1. The paintings, public renderer and worker files remain exact.

Twelve isolated cases compare approved meadow, original Golden Veld and the fine-grain study at scales 1 and 2. All stationary/pan-return cases are exact; CPU/GPU differences reach 3 channel levels with unchanged alpha. The complete comparison board was directly inspected. Coherent diagonal banding is reduced in the meadow and new study, while the original retains painted directional strokes. Some recurring clumps and blend structure may remain. Numerical sampling density does not prove equal perceived detail, palette or nonrepetition.

A separate 20-frame private-renderer world comparison completed at the exact existing field-profile site used by the 40-frame baseline. It reuses the sole owned legitimate QA world, without cloning or fabricating state. This is not production activation or canonical source admission. Contact/state, cost/lifecycle, other biome roles, full-world pressure/return and actual production appearance remain open.

Evidence: component results component-results.json, direct component review component-visual-review.json, bounded comparison board component-comparison.png. Columns: approved meadow, Golden Veld original, new fine-grain study. Rows: current/private sampler at scale 1, current/private sampler at scale 2. Raw buffers stay in memory; no native artist variants or rollback snapshots are created. The previously blocked disposable-world cleanup is not retried or bypassed.

All four Region returns have identical geometry and exact canvases; all four Hex permalink reloads have identical bounds and exact canvases. Direct desktop and phone review shows reduced coherent diagonal bands but does not accept recurring small clumps, perceived grain/palette fidelity, whole-mesh nonrepetition or other-biome/contact/state/cost/lifecycle quality. See world results world-context.json, direct scene review world-visual-review.json and private HTTP identity private-renderer-http.json. The owned server exited normally and every context/browser closed.

Forty-four further isolated cases exercise all eleven current primary profile contracts at their own spans, two scales and both samplers. All stationary and pan-return cases are exact; CPU/GPU differences reach 3 channel levels, with unchanged alpha. Both complete boards were directly inspected and eleven role-specific findings all-profile-components/visual-review.json were authored. These include separate hill and grassland uses of the same native source. Wetlands retain directional woven-looking strokes; no wetland quality acceptance is granted. Actual ocean/lake rendering is not observed by a material window. Public promotion and full perceived fidelity, recurrence, contact/state/cost/lifecycle gates remain open.

Full-native Wetland source WOG-A-003578 was also directly inspected at 1254 × 1254. Its angular stones, stems and moss/soil pockets retain crisp irregular detail. The woven-looking component display is therefore a sampling/filtering/physical-grain question to isolate; it is not sufficient reason to repaint or downscale the original. Its exact SHA-256 is recorded in the profile finding.