Adaptive native-cache fidelity and instrumented cost observations

The unchanged native artwork was observed in 204 completed normal local Region/Hex frames across 12 cohorts (0 failed cohorts retained separately). The client selects its own 128/192/256 MiB native cache defaults from emulated 2/4/8 GiB deviceMemory hints. This is not a physical RAM cap or physical-phone certification. Each completed cohort uses the existing natural Gabled Court town and four existing arctic/desert/forest/wetland sites, with exact source transforms and actual source paints checked. No world command, geography change or source admission is used.

The private planner has 12 exact displayed-map pairs and 42 native-exact isolated town replays. This is bounded fidelity evidence for the measured cohorts, not whole-artwork acceptance. Gabled Court contributes 108 new own-source frames; the four primary materials contribute 96 new frames. Their invocation and coverage do not isolate final material pixels or approve artistic quality, nonrepetition, grain, contact or state. Cumulative completed client observations are 324; the town's own-source count is 196.

Raw terrain differences retain bounds, changed RGBA samples and, for pressure comparisons, CSS coordinate projections against the map panel. Coordinate measurements alone do not establish clipping/occlusion or waive raw-buffer stability. Rounded map-link reloads may change actual transforms; those results remain distinct from exact-transform pairs. The earlier control-phone four-pixel/max44 finding and prior wetland/Hex discrepancies remain open even if a later repetition is exact. No result is generalized to the differing deployed renderer.

Owned browser/server process CPU counters survive document resets. Private commit and working-set samples and a GPU checkpoint are retained for each measured navigation. Working sets can share pages, exited-process CPU is a lower bound, and the GPU checkpoint is outside timing and is not a peak. Earlier diagnostic canvases, pixel reads and PNG encodes can influence later allocations or garbage collection despite occurring outside timed operations. These costs are instrumented observations, not a clean performance benchmark. A separate fresh-browser, no-readback, counterbalanced cost study is required. Full performance, aggregate memory, physical hardware, crowded contexts and lifecycle remain unaccepted.

The initial interrupted attempt and a restart rejected by an abandoned local writer claim are preserved separately; no incomplete frame is added to these totals. Recovery preserved all 18 durable disposable-world files exactly and changed no production save or locking rule. All owned resources from the completed run are terminal. Only proven duplicate progress/current-case payloads were compacted into pointers to the exact final report. Native artist files and public modules remain exact, no raw pixel files are retained, and no approval gate is cleared.

Evidence: summary validation-summary.json, all completed measurements and failures world-results.json, retention retention.json, owned-resource terminal world-owned-resources-terminal.json.

Verified live Atlas publication

Commit 349e1acecf19a919e9d01e2132cc93eed024469f deployed automatically as a7cbf1b6-1a08-4555-b159-2c8f3f48ce50. Exact live metadata and the 9,206-member inventory, eight native image identities including both Citadel hamlets, ten desktop/phone Atlas finding views and seven generated wiki pages passed with zero browser errors. All five metadata rows are synchronized. Original artwork, source approvals, the public renderer, production geography and saves remain unchanged. All 5,634 qualifications and earlier terrain/Hex failures remain open. The separate cost-only study does not gain acceptance from this publication.