Native Region fidelity: exact image retention under qualification

The preferred private method preserves the original decoded image without a preparatory raster copy. Four desktop/phone Region/Hex views retain the whole native 1254-square ruin, including the roof, blue flag, broken stone walls, snow and rubble. The source remains a separate unapproved study with every original RGB pixel preserved. The public renderer and its 322 source contracts are unchanged.

The holder checks permission before access and retains the original decoded HTML image under a version/URL key in the existing 32 MiB display LRU. The full native pixel area is charged; oversized sources use the existing native record. A holder hit precedes original-cache lookup, so native-cache eviction does not force a duplicate decode. Holder release clears its own reference only. No source resize, crop, lossy recompression, repaint or source-ID exception is used.

Actual selected-paint omission changes 30,799 desktop/30,901 phone pixels, with zero changes outside the complete projected source rectangle. Omission restoration, native-reference policy restoration and display-cache clear/rebuild all restore the full map exactly. Loading 64 other approved native images per layout causes 35 desktop/54 phone actual native-cache evictions and removes the original ruin record. The retained holder still reproduces the complete map exactly. Measured native peaks are 264,225,820/132,115,080 bytes within 256/128 MiB configured budgets; display charges stay below 32 MiB. These logical cache charges do not measure total unique browser/GPU memory.

Eight native-holder initial/mode-return/pan-return/reconnect states are each compared with direct original-native rendering at that same actual pose and interaction. All maps match byte for byte. The production 512 control differs in all 8 pairs, but restores its own policy exactly each time. The generic reference changes all visible canonical-place preparation; whole-map differences are not assigned solely to the selected ruin. Raw initial-versus-return differences remain retained in both cohorts. The holder has same-pose mode/desktop-pan differences and a changed phone pan transform. Current-state fidelity does not erase those observations or prove their cause.

The first combined holder run completed 2 desktop views and the desktop contribution/cache-pressure probes, then hit a 30-second phone startup wait caused by an inherited Playwright argument error. The corrected separate phone run completed 2 views/3 lifecycle checks; its 990 ms observed startup is not a performance acceptance criterion. Its inherited descriptive header said 4 views/6 checks; the completed frame/terminal records prove 2/3. The interrupted first production-control attempt retains its partial captures and unknown exit codes. Authoritative OS absence, an unreachable owned listener, and exact preservation of 18 durable disposable-world files precede a named successful recovery run. Runtime automatic writer-claim recovery remains disabled.

The readback-free cost study compares the published 512 policy with native holders in fresh browsers, using two reversed orders, three emulated cache tiers and both DPR 2 layouts. It measures navigation, warm redraw requests, pointer pan and repeated Region/Hex switches. The first attempt timed out on a repeated Hex readiness check after Region phases; no complete cost case or performance acceptance is claimed from it. A named diagnostic attempt retains compact per-phase state and timings. Inspect its actual final results and source/paint binding before interpreting latency. No diagnostic screenshots, canvas pixel readers or retained source-image collector run in timed operations.

A complete 2,252,800-cell natural-world selector scan covers all 39 published canonical-place sources plus the private ruin across 114 declared existing-state profile/feature contracts.110 contexts have natural sites. Four—tundra floating stones, town gabled court, cold ruin arch, and drumkeep—have no natural selection in this seed and remain explicit alternate-seed work. The plan is not rendering coverage, source admission or artwork approval. Fresh effective-world/occupancy checks, normal native Region/Hex observations and appropriate regression remain required for all affected contracts. The source dimensions total about 240 MiB decoded; the 32 MiB display budget deliberately retains only a bounded working set.

The parent ruin's wall-detail and quiet-ground companions were not painted by these parent-only tests. Preserve their originals and preferred studies without granting them the parent's fidelity result. Complete finite nonrepeating hierarchy, role/state, contact, neighbor occlusion, raw lifecycle attribution, broader travel/cache contention, complete memory/performance, actual devices, source admission and production appearance remain open. Every other pending full-library asset stays in scope.

Preferred native context native-holder-post-state/WOG-A-009284-tundra-context.png. Desktop actual contribution native-image-holder/actual-contribution-1440.png. Phone actual contribution native-image-holder-phone/actual-contribution-390.png. Complete generic regression plan canonical-regression/plan.json. Native current-state results native-holder-post-state/results.json. Published control results published-control-state-v2/results.json.

The corrected readback-free study is complete: 24 cases / 12 pairs, two reversed orders, three emulated cache tiers and both layouts. Every measured phase stays within its logical native/display budgets. The observer now keeps plain source-bound metadata by canvas and accepts the ready, visible cached Hex portrait; it does not demand a redundant paint call. Both earlier bounded failures remain retained.

Across the 12 pairs, native-to-published warm-request wall-time ratio has median 0.997 (range 0.917–1.012); browser CPU ratio has median 0.940 (range 0.577–1.114). Renderer-reported frame p50 ratio has median 0.987 (range 0.727–1.257); p95 ratio has median 1.008 (range 0.760–1.760). Ratios describe this selected scene and desktop host. Request cadence includes animation-frame scheduling; process CPU can be a lower bound for exited processes. Sampled working sets can share pages, and GPU readings are checkpoints rather than peaks. These paired results do not approve broad travel, total renderer memory, physical phones or the remaining 114-contract regression. See the compact analysis analysis.json, completed case index readback-free-cost-v3/results.json and observer correction cost-readiness-binding-correction.json.