Local portrait sampling depends on cache history

The exact current local regionalDisplayRecord, PixelCache, place display planner and painter produce different output for the same final Region draw demand after different cache histories. The initial observation is a real-browser component result; bounded normal local map verification is recorded below. It is not a reproduction of the earlier Hex navigation failures.

The input is the individually source-approved Gabled Court, WOG-A-003363, preserved at 1536 × 1024 with SHA-256 7ba2c669cf81289dce6d88c29ac7097693ac6babe7fe70992f49c52d6d5eb3fc. Its artist bytes remain unchanged.

At the final Region zoom of 1.16 and renderer DPR 2, the requested physical width is 510.4 pixels. A cold cache prepares a 513 × 342 portrait. Warming the same cache class at zoom 1.14 first creates a 510 × 340 portrait; the final request then falls back to the 1536 × 1024 native image. The same final demand therefore uses different sampled inputs. The final opaque-backed frames differ in 80,997 RGB pixels, with maximum channel difference 83. These framebuffer measurements do not claim that source alpha changed.

The private candidate returns the original native image when an aspect-preserving rounded portrait would exceed its cache class. Both cold and warm final outputs then match the native reference exactly, with zero pixel differences. The public renderer is unchanged. This candidate preserves the existing cache budgets; its full application cost is not yet qualified.

A matrix covers all 51 currently accepted local canonical-place sources, seven stated zooms and two DPR values: 714 cases. The candidate changes 32 plans to native input and leaves 682 plans exact. No tested input resolution is reduced, and native aspect ratios are preserved. This is display-plan coverage, not 714 rendered world frames or physical-device certification.

The deployed modules at commit 74e7569 differ from the local modules; the deployed place module has no displayPlan method. The component observation does not establish a production bug. The current local Hex route also bypasses this Region display cache, so this result does not explain either the prior 32-pixel/max-245 wetland ground failure or the larger Hex pressure-return failure.

The initial stale-queue hypothesis remains unproven. The inspected landscape painter reacquires its record immediately before painting, and the canopy band path already draws immediately to avoid queued closed canvases. Cache closure alone is insufficient causal evidence. The normal-client cold/warm Region and native-reference step is now completed for the bounded scope below. Visible compositing,navigation,full-cost,pressure and lifecycle checks remain necessary. No geography, source selection or gameplay state should be fabricated for that test.

No source approval, admission, world placement, production activation, performance acceptance or whole-asset qualification is granted. The complete asset goal and every prior failure remain open.

Retention keeps the current Region-consistent control and private candidate, their compact measurements, one failure witness and one passing witness. The superseded exploratory scenario and its duplicate-style witness are retired with exact hashes and a replacement mapping. No native artist copy or raw pixel readback is retained.

Evidence: Region control region-probe.json, private candidate candidate-probe.json, plan matrix plan-matrix.json, release distinction release-boundary-audit.json, retention scenario-retirement.json.

Normal local world verification — 4 November 2026

Read-only canonical selection and the stock disposable QA world identify an existing,unoccupied grasslands town at cell248797. The ordinary selector chooses the exact approved Gabled Court source without a source alias,forced feature,world command or source admission. The read-only engine snapshot is exact before/after inspection. This study records52 normal-client observations:24 initial navigation frames and28 corrected cold/warm/Hex/return/reload frames,including26 private-planner frames. All page errors and external HTTP requests are zero. Public renderer and artist bytes are unchanged.

The first native wheel journey stopped at zoom1.15 rather than1.16,so its cold/warm comparison is not qualified. A stricter run then failed on the phone camera equality assertion: WheelEvent client coordinates are integral while the canvas height is fractional. The failure is retained. Fourteen completed desktop case measurements from that aborted run were not retained in a final result and are not counted in the52 retained observations. The failed phone captures remain separately inspectable. This is not an artwork failure waiver.

The corrected run uses a synthetic DOM wheel event through the unchanged normal zoom handler and an ordinary full-precision map link to reproduce the resulting camera on a fresh cold page. Both final zoom and actual six-element canvas transform are exact in all four matched pairs. Device emulation is not physical-phone certification.

The isolated replay uses the actual normal-client paint arguments,transform,canvas size,smoothing and alpha. Its RGBA comparison includes transparent-edge colour values; maximum differences must not be described as entirely visible or opaque changes. In the private candidate,ten measured final/cold/return/reload source replays match the native image exactly. The earlier1.14 warm-up still uses an intermediate portrait and is not claimed native-exact. Full foreground/base/static layers are measured separately,and any differences or unequal auxiliary surface sizes remain in the result. A source replay match alone does not clear whole-world fidelity.

Rounded map-link coordinates can produce a different actual transform despite matching rounded camera datasets. These reload comparisons are retained with the mismatch labelled; this observation does not prove quantization is the sole cause of every changed pixel. Exact transform matching is required for the cold/warm claim.

The four retained initial context witnesses were directly viewed. The complete Gabled Court roofs,civic tower,connected courtyard and narrow feet remain recognisable on the selected grasslands. The private cold view preserves fine masonry and roof detail without repainting or changing scale. The selected phone portrait is visible within the map; other neighbouring portraits reach the screen edge and controls,so this is not general neighbour/control clearance. This finite portrait must never become a repeating mesh background. Wider biome/contact/state,default-budget pressure,complete cost/lifecycle and production applicability remain open. No new approval or whole-asset qualification is granted.

Evidence: compact normal summary normal-context-summary.json,52-frame source records normal-world-results.json,28 corrected frames normal-matched-world-results.json,strict phone failure normal-exact-world-failure.json. No new screenshot was saved for the corrected run,and all raw comparisons stayed in memory. The same sole disposable world is retained; every owned browser context and server is closed.

Verified Atlas publication

Releaseee6d764f7326d900fc67c9eed7459f6e2b1d78bc deployed automatically as898c45bf-48ab-41af-8bb9-db106fae9d62. Exact metadata and9,206-member inventory,four selected native artwork responses including both Citadel hamlets,two desktop/phone Atlas finding views and five generated wiki pages passed live verification. The single metadata row is synchronized. This release changes no public renderer,source approval,artist byte or game-world placement. Visible-composite and default-budget pressure continuation remains separate until its own result and registration. All5,634 qualification records and prior Hex failures remain open.