Task 2 proof and digests supersede the historical scale-blocker status below. All four FS3 packages now bind 361c681def79cbce5549edc32f93f88360365e85; contracts/geo remain b707a0cf/0.1.4, one resolved copy. The A666 native-pair checker self-test reports 99.999999998660 m (placement XY) / 100.000000000506 m (geodesic), within 100.000 ± 0.01 m. All eight former east-scale expected failures pass as normal tests; maximum endpoint residual is 0.000814113 m under the unchanged 0.05 m tolerance.
The correction uses distinct WGS84 horizontal coefficients; measurement never uses the spherical Z scale. Numeric and browser/served-code identities are recorded in the linked evidence. A666/A673 real-GLB footprint adaptation and live placement remain worker 8’s follow-up; local relative scale does not establish survey accuracy, datum alignment or A459 release acceptance.
Primary references are now the seven full expected GeoReferences in M6’s checksum-verified mac-check-inputs.json. TEST-only copies preserve exact bytes/IDs. apps/web/src/catalogue/m6-georeferencing.test.ts validates them, compares independently projected EPSG:2953 origins to the declared WGS84 anchors within 0.001 m, verifies actual shared Mercator anchor placement within 0.001 m, preserves baked ENU/heading/local-ground semantics and checks 10 m probes within 0.05 m. The 100 m east regressions fail the same tolerance at all seven real anchors, with 0.281529–0.286361 m error; they remain explicit expected failures requiring reviewed FS3 correction. Source vertical datum/status are unknown; no absolute height conversion or survey accuracy is claimed.
Real-reference diagnosis and m6-reference-analysis.json record source/projection/matrix measurements. PROJ’s selected operation declares 1 m accuracy; numerical agreement does not claim that accuracy of physical ground truth. No authenticated model or GLB was fetched in this preparation.
The dev-only Gate C checker is prepared for exact bbox/32-29-3 counts and all seven returned GeoReference comparisons/actual GLB render views. The checklist’s 32 coverage IDs are tiles; containment is not a per-building footprint oracle. Actual house offsets/rotation remain pending M6’s oracle and the final live window. Only request IDs are persisted. The separate exact export is locked and has not run.
The prior independent-control and mocked HTTP results below remain supplementary numeric/regression evidence. Their geometry is never presented as a real M6 building.
The runtime demo implementation was removed in local commit 6b2a795. All test responses and GLB geometry are defined under tests/fixtures, intercepted in Playwright or supplied as unit test doubles. No fixture flag, mock provider or selectable data mode exists in the application.
The test-only model has the real contract shape: authorized_real, EPSG:2953 / NAD83(CSRS) source metadata, easting/northing metres, CGVD2013, ENU Z-up metre GLB, local ground, WGS84 longitude/latitude anchor, conversion receipt and real-source membership. It is still an HTTP test fixture, not a surveyed building or a model served by M6.
GeoNB’s official High Precision Network supplies Moncton monument 941004 and Saint John 941005, their native EPSG:2953 coordinates and meridian convergence. Test anchors use the service geometry explicitly requested with outSR=4326. The request did not specify a datum transformation or epoch. This verifies consumption of a published WGS84-shaped render coordinate, not an independent NAD83-to-WGS84 producer conversion or surveyed absolute accuracy.
The NRCan Ottawa CGSN 9910-1997 station publishes scaled horizontal coordinates with ±50 m uncertainty. The Ottawa check is a render-anchor numerical probe only. The current real contract permits EPSG:2953, a New Brunswick source CRS; it does not provide an Ontario source CRS. The tests explicitly reject EPSG:32189 rather than claim real Ottawa source support.
Published control records and source limitations are preserved separately from the application’s shipped data.
The initial browser run measured maximum corner offset 0.017452 m and centroid offset 0.00000234 m. Per-camera results and the test-only overlay screenshot are recorded. These are internal geometry/coordinate checks, not a claim that the mocked box matches a real surveyed building.
The unchanged FS3 spherical metre scale still fails the 100 m east / 0.05 m acceptance regression at Moncton. That test is explicitly marked as an expected failure and must become a normal passing test after a reviewed shared correction/rebind. Its expected failure does not constitute production georeferencing acceptance. The earlier read-only diagnosis documents the scale/extent error.
Final proof still needs approved ModelVersion / CoverageItem / preview artifact pairs actually served by M6, plus producer-owned coordinate/heading/ground reference evidence. Verify whether each coverage polygon is a host footprint or a broader coverage region before treating it as a shape oracle. Keep the contract’s NAD83 basemap notice separate from numerical tolerances and survey uncertainty. No shared source was edited.
