Lake study across natural shore biomes

WOG-A-009280 remains the selected native short-wavelet study. Its original 1536×1024 PNG SHA-256 is ba4dc923e1209f2ff75918721bd9626f5a5a79f13742d7b9b5ffd552da1888ab. The original WOG-A-003575 and prior lake studies remain unchanged. No public renderer permission or complete game-ready approval is granted by this review.

The exhaustive canonical-kernel scan examined 2,252,800 cells, including 34,988 lake cells. It found seven naturally adjacent land biome types: arctic, desert, fields, forest, mountain, tundra and wetlands. One representative lake/land pair per type was verified against the effective disposable QA world using a read-only engine probe. The scan covers the biome inventory, not every position, bank orientation, occupied state, season or physical device.

The first matrix completed 56 actual original/selected comparison views: seven sites, two native sources, desktop and phone, Region and Hex. All 28 source pairs have matching displayed geometry, scene bounds and crop coordinates. Fourteen real pointer pan/return comparisons are pixel-exact. Complete maps were compared in memory; seven native contact sheets are retained, with no per-frame full-map PNG or RGBA cache.

Direct inspection supports continuous banks in the seven tested contexts and visible shorter painted highlights in the useful water crops. Region ponds are quieter at the small displayed scale. The field pond needs an interior detail witness. The mountain desktop Hex contact crop contains land only, so it cannot prove desktop Hex water contact. Those evidence gaps remain explicit rather than being inferred from material requests.

The initial 28 reconnect comparisons are invalid as stability evidence. The harness checked semantic view state before initialization completed and clicked a view button, changing the camera; Hex reloads also reused an initial Region URL. The original observations and exact driver are preserved. A corrected private test uses the existing Share view link, captures a settled baseline at that link's pose, reloads, waits for the expected view, then compares the full camera and scene bounds before measuring pixels. It does not repeat the already-valid shore pan-return checks.

The disposable QA world uses existing geography and normal controls with explicit private image-response substitutions; no world commands are sent. All owned browsers and the server from the first matrix closed normally. Field GPU/CPU comparisons and the mountain witness follow serially on that owned QA directory. Source admission, broader perceived nonrepetition, full state/lifecycle coverage, physical-device performance and the complete library goal remain unresolved.

See natural domain inventory natural-matrix.json, initial observations normal/results.json, initial lifecycle audit initial-lifecycle-audit.json, direct native sheet reviews visual-review-progress.json, and the previous live release closeout ../lake-short-wavelets-native-20261104/closeout.json.

Completed reconnect and missing-witness follow-up

All 28 corrected same-pose reconnect comparisons are pixel-exact. The actual full-precision camera, selected cell, scene bounds and world instance match. The invalid first 28 reload observations remain excluded. Added field GPU/CPU tests each provide eight original/study desktop/phone Region/Hex frames and six exact pan/reconnect comparisons. Eight full-map backend comparisons streamed 16,761,344 pixels through memory;9,211,580 differ,with maximum channel changes from 2 to 8. Both sources show that backend difference. It is not a pixel-parity pass or an isolated lake failure.

The retained native source-contribution crops show maxima of 2,6,2,3 channel levels for desktop Region/Hex and phone Region/Hex,with selected one-pixel luminance-change ratios of about 0.986,1.001,0.995,1.001. Direct GPU and CPU inspection still shows fine painted water marks and continuous banks. The crop mask is inferred from original/study differences and is not a full-map semantic water mask. These findings support the tested field context without closing universal fallback,exact parity or performance gates.

The corrected mountain lake-centered witness adds eight source views and four exact reconnect comparisons. Desktop Hex now visibly exposes water meeting the rock/bank edge;the earlier land-only crop is retained with its stated limitation. Region and phone Hex also expose the actual contact. Both missing Hex witness gaps are resolved at these sites. The field pond remains small and quiet in Region.

Across this package there are 108 base source/context views and 58 valid pixel-exact lifecycle comparisons,plus 28 explicitly invalid initial reload observations. No world commands,source admissions,production-world frames or artist-file changes occur. Ten bounded native contact sheets are retained;no raw pixels or per-frame full-map caches are written. See the corrected reconnect results reconnect-corrected/results.json,complete GPU/CPU map comparison fields-backend-pair.json,native backend diagnostics fields-native-backend-diagnostics.json,mountain observations mountain-interior-gpu/results.json and final visual decision final-visual-review.json.

Descriptive metadata index follow-up

The initial catalog run passed 53 checks and failed one freshness assertion:the local descriptive index lacked the two prior lake studies,9279and9280. Exactly two planned-study entries are now added;all4,535existing local entries and aliases remain exact. This is descriptive metadata only,with no source permission,artist-byte or production-index change. The existing index queue row is updated in place;the other42runtime rows remain untouched. Its broader unpublished local differences remain deferred. See the constrained index receipt metadata-index-followup.json. The initial assertion expanded both indexes into a3.85MBdiagnostic;after exact mismatch analysis,the repeated dump was compacted to a short failure receipt with the original output SHA-256.