Raw and stable native-map return findings

The completed 34 source-context view sets contain 204 raw Region return comparisons: 78 are pixel-exact and 126 differ. These are 23 actual-production-pool contexts and eleven separate private-pool contexts. The active 90-contract run is excluded. No existing raw failure or asset approval is rewritten by this retrospective analysis.

Each raw comparison is bound to its exact source, cell, viewport, initial PNG hash and the later stable current-state reference capture. Every raw expected-map hash matches that context's stable initial-reference hash. The later stable capture waits for native loads and three matching screenshot hashes before the reference-policy toggle. Its state and exact whole-map PNG hash are therefore useful additional evidence, without requiring another image copy or claiming the cause of a mismatch.

The 27 later exact hashes establish an observed stable recovery for those particular captures. They do not turn the earlier raw capture into a pass or prove that no visible transient occurred. The 34 pose/interaction changes require camera and presentation interpretation before assigning a rendering fault. The 65 same-pose differences remain unresolved: matching recorded geometry, interaction and presentation do not prove all internal renderer state is identical.

In those 65 records, the captured scalar scene-cache state differs only in revision and renderedRevision; cache center, zoom, backing dimensions, margins, valid state and rebuild count match. This rules out a recorded scene-center shift in that subset. It does not identify the changed pixels as source artwork, background, a later overlay, lighting or cache backing. No native image should be softened, cropped, repainted or downscaled on the strength of these whole-map differences.

All 272 current-state native-reference/restoration pairs in the five terminal cohorts remain exact. The direct-native candidate already uses the original image path for canonical places, so this policy parity is not an independent proof of final source contribution, initial-return equality or complete lifecycle/performance readiness. The retained isolated source replay and native-input hashes provide the separate source-fidelity evidence.

The next bounded diagnostic must replay representative same-pose failures through ordinary Region/Hex, pan-return and reconnect controls in the owned default-seed QA world after its active writer finishes. Capture raw and stable maps in memory, record the complete relevant visual and renderer state at each capture, and compare the selected-source-present and selected-source-omitted compositions at the same state. Attribute differences to opaque source detail, translucent source contact, background or later overlays before choosing a fix. Preserve focus/pointer and camera observations rather than normalizing failures away. A successful diagnostic subset must then be followed by the complete production/private contract regression and relevant cost checks; it cannot approve the full library on its own.

No artwork, public renderer, source permission, production geography or save was changed. Exact classifications and source receipts analysis.json bind all 204 comparisons; the original terminal result files remain authoritative for their raw outcomes.

The Atlas package contains 31 individual original-asset findings derived from these 34 existing contexts. Each finding binds its source hash, prior authored visual witness, exact cohort/pool/profile/cell, per-context counts and next diagnostic. These are new metadata findings, with zero new normal views or artwork edits. Registration replaces only matching assets' older queued evidence leases; the current release can still close against its exact published source/evidence leases while preserving the newer classifications. No existing approval, ledger goal action, raw result or deferred runtime declaration is changed.