Tangent angular approximation fails exact-plane coverage

Frozen f349514f, import-order-only rerun febdaa2b. See protocol, plan, 27 exact-plane scenarios, 27 surface controls and schema.

Reject this tangent confidence approximation. Nine of27 exact-plane scenarios fail the predeclared >=94% empirical coverage gate. Minimum coverage is78.47%, median98.005%, maximum98.59%; the median hides the failed frontal-plane stratum. All nine failures have zero disparity slopes, across center disparities6/12/24 and all three tested image positions. Tightening the previous bound by a local linearization is not calibrated throughout this parameter range.

Exact frontal-plane center disparity	Principal-point coverage	Off-center coverage range
6px	78.47%	85.67–85.94%
12px	83.18%	85.36–85.71%
24px	88.58%	89.61–89.79%

The enclosing-ball baseline has minimum97.90% coverage on the same27 scenarios. The tangent method uses the covariance of a normalized noisy normal and a95% 2D ellipse/circle approximation; finite-noise nonlinearity and plug-in length/ direction mean that this approximation need not have nominal coverage. These results are not a failure of the known linear coefficient covariance itself. Each scenario uses20000 independent coefficient draws from the frozen .5px model. Image-patch overlap is not involved in that exact-plane validation.

Curved/boundary controls remain diagnostic

The surface controls were executed as explicitly planned even though the plane coverage gate failed. All27 depth-array hashes match the prior experiment. Clean data retains zero points in all nine frames with radius<=15degrees. For noisy seeds730/731, only near_foreground frame2 has nonempty selection: 10 and2 points respectively, both with100% joint surface correctness. Every other noisy frame retains zero. These tiny conditional successes do not pass the >=10% retention and spatial support screens. Null empty-population precision is preserved.

Across all-valid noisy patches, observed coverage ranges78.80–81.95%; within the previous depth/planarity gate it ranges89.16–94.61%. Curvature/boundary model error adds another limitation; exact-plane calibration must be fixed before trusting these intervals. There is no registration run or revised calibration recovery.

Execution and verification

An initial completed run used a test file with unsorted imports. The test import order alone was corrected, its source hash updated, and the full deterministic validation rerun. Original plan and outputs remain under initial-* in the same asset directory. Final surface case records exactly match the initial case records; no numerical algorithm, model, threshold, noise, seed or denominator was changed.

Four focused tests, repository Ruff and strict documentation build pass. Tests check the analytic tangent variance at a principal-point frontal plane, missing patch handling, the previous exact-plane confidence sanity check and far-plane behavior. Validation checks schema, frozen hashes, case counts, paired depth identity and deterministic rerun equality. Core/defaults and MI baseline remain.

Close the5x5 local confidence shortcuts. Neither a broad enclosing ball nor the uncalibrated tangent approximation provides a usable measurement-only gate here. A next measurement-level direction is coherent surface estimation with larger spatial support, explicitly balancing reduced disparity-slope variance against curvature and discontinuity bias. Freeze those model assumptions and independent coverage tests before any real-data filtering or ICP rerun; do not select window sizes or angular thresholds from the18 failed real frames. This work establishes no production readiness or SOTA claim.