Generic native-image regression across existing canonical places

The complete plan contains 39 published canonical-place sources and the one private original-preserving tundra ruin, with 114 explicit profile/feature existing-state contracts. The earlier full natural-world scan checked all 2,252,800 cells. The new read-only effective/occupancy preflight checked all 509 retained candidate cells in the sole existing disposable local world. Every batch preserved the engine snapshot exactly, created no actors and issued no world commands. All 110 default-seed contexts remain eligible across 39 sources, with zero additional stored-candidate blockers.

In the private 40-source pool, four contexts lack a natural default-seed selection: tundra floating stones, tundra gabled town, tundra cold ruin arch and tundra drumkeep. The drumkeep WOG-A-003439 has no eligible context anywhere in this seed. Preserve all four declarations and the exact native artwork for legitimate alternate-seed checks; do not rename source IDs, force selection or replace canonical geography to manufacture a result.

The first diverse batch selects ten distinct source assets from the full eligible plan, covering multiple existing-place features and display profiles. It uses the generic private native-image holder, exact SHA-bound source bytes, the complete source pool, normal camera links and normal Region/Hex controls. It plans 40 native context views, 80 same-current-state native-reference pairs and 60 raw initial-return comparisons. These counts are a plan, not completed evidence. Each final result must bind the exact source, actual current pose, normal selected cell and complete native dimensions. Every unselected context stays in selection.json and the full effective-sites.json.

Pixel comparisons remain in memory. Keep compact per-source findings, exact source/output hashes, a bounded native context witness per tested source and actual owned-resource terminal records. No raw RGBA buffers, repeated catalog snapshots, generation inputs or artist-byte edits are introduced. The production renderer and its 322-source whitelist remain unchanged. Current-state native-reference equivalence, raw lifecycle return equivalence, artwork approval, source admission and actual production-world appearance remain separate decisions. Full contact/occlusion, family/state suitability, broad travel/cache contention, complete renderer memory/performance and actual-device checks remain open.

Read-only effective-site results effective-sites.json. Read-only server terminal owned-resources-terminal.json. First ten selection and unselected contexts first10/selection.json. First ten driver binding first10/driver-binding.json. Prior native fidelity and measured costs.

The first village source passed native desktop Region/Hex isolated replay but its phone scene failed to settle, recording 13,991 native evictions under the unchanged 128/32 MiB budgets. The retained failed map and exact source-bound controls supersede any inference that the single-ruin holder results qualify the generic policy. Published reduced preparation and direct full-native preparation both settle at the same village phone pose, with 19/24 evictions respectively. No source pixel is changed. Prefer direct-native preparation for the next complete generic regression; keep the holder experiment and its failure. Controls and limits village-cache-controls/analysis.json.

The holder batch is terminal with six complete view sets (24 frames), four failed sets (ten partial frames), 66 exact current-state native-reference pairs, and 49 raw return comparisons of which 17 are exact. The four failed sources are village crescent, farm barn, desert courtyard city, and courtyard castle. Full source-bound diagnostics are retained for the latter three; the overwritten first village diagnostic is explicitly represented by the earlier observed excerpt and its first failed map. No full source approval follows. The farm direct-native control completes all five readiness/lifecycle states. The other controls lost their already-shared server when the main cohort shut it down normally; they are incomplete, not direct-native policy failures. A fresh direct-native batch now tests the same ten source/site choices, with unique per-source failure paths and a single bounded transaction checkpoint. Holder summary holder-batch-summary.json. Direct batch driver direct-first10/driver-binding.json.

Completed direct-native cohort and production selector

The private direct-native candidate performs the existing permission checks and returns the existing native image record for canonical-place sources. It makes no intermediate raster, resize, crop or holder charge in the 32 MiB raster display LRU; other roles retain their previous preparation. Ten original sources complete 40 actual desktop/phone Region/Hex views and isolated native replays, plus 80 exact same-current-state complete-map native-reference/restoration pairs. The village, farm, desert city and courtyard castle all complete after the earlier holder policy failed their readiness. This proves the tested preparation fidelity and readiness, not broad performance or full source suitability. 21 of 60 raw return maps are exact; all remaining differences and current-pose comparisons stay explicit. The exact production pool remains 39 canonical sources / 322 permission records / 113 profile-feature contracts. The separate private candidate pool remains 40 / 323 / 114. All 39 existing descriptors, permissions and native hashes agree; 80 of 81 choice vectors are identical. Only tundra/ruins changes: the actual production pool selects cold-arch-v1 at 2035674, while the private pool selects the proposed matte. Fresh read-only occupancy/effective-state probes of all 509 retained candidate cells find 110 eligible production contexts and three natural absences (tundra floating stones, gabled town and drumkeep). The private pool has four absences including cold arch. Both full scopes remain explicit; selected-source equivalence does not establish complete-map production-pool equivalence. Continue the remaining contexts with the actual 39-source pool, including the original cold arch, and retain the private candidate separately. No public renderer, permissions, geometry, artist bytes or gameplay rules change.

Completed direct results direct-first10/results.json. Source-bound native visual review visual-review.json. Analysis and raw lifecycle differences analysis.json. Actual-production selector bridge production-selector-bridge/results.json. The next distinct-source batch uses the exact production source/permission pool, including the naturally selected original cold arch. These are planned next-batch counts until its final results and owned-resource terminal record exist.