[cailmdaley]

Measured pair-count shape noise now covers every tomographic bin pair as well as the whole sample.
On the blinded v2 tomographic data, pair-count noise doesn't rescue the joint ξ−B null test: PTE 0.0045, against 0.0046 with analytic noise. Analytic/measured noise medians are 0.973–1.006 at 12–83′.
This remains a draft pending validation on GLASS mocks with a known answer (#443).

OneCovariance's uniform-area estimate misses mask holes and spatial weight variation.
The workflow measures W = Σwᵢwⱼ and Q = Σ(wᵢwⱼ)² from positions and weights, then supplies effective counts through its tabulated-pair input.
For tomography it also reads the bin-ID field, excludes nonpositive IDs, and counts every auto and cross pair separately on each covariance grid.

• covariance.shape_noise: pair_counts is the cosmo_val default for both whole-sample and tomographic covariances; analytic restores the uniform-area estimate.
• Tomographic counts use shear.tomo_bin_col, with files in upper-triangle bin order; the covariance retains the binned sample's cov_th.tomo area and mask.
• The count cache contains no shapes, signals or blind records.
  The mixed, sample-variance, non-Gaussian and SSC terms are unchanged.

The normalization differs for autos and crosses.
OneCovariance takes per-component dispersions sᵢ.
Auto-bin tables contain 2W²/Q; disjoint cross-bin tables contain W²/Q, because only one Kronecker contraction survives.
Both give the independent-shape variance 2sᵢ²sⱼ²Q/W².
Doubling the cross count would halve its noise incorrectly.
The derivation and cache conventions are in docs/pair-count-noise.md.

The tests include explicit per-bin-pair weighted answers, shape-free HDF5 reads, DAG dependencies and bin-pair file order, analytic opt-out, and footprint inputs.
An opt-in probe executes the real OneCovariance table reader and real-space noise kernel with unequal bin dispersions; its Fourier cosmological terms are stubbed to zero.
It reproduces all auto/cross answers and vanishing cross-pair and ξ+/ξ− pure-noise covariance against candide OneCovariance 70265be3.
It does not validate the full cosmological integration.

Existing non-tomographic full-v2 evidence

The analytic/measured shape-noise variance ratio is 0.809–0.856 (median 0.839) over 12–83′ on the full v2 area.
The broad deficit persists at full area; the pilot's extra large-scale decline reflects its boundaries.

With fixed full-v2 B vectors at 12–83′ and only the noise diagonal substituted:

Statistic	dof	PTE, analytic	PTE, measured pairs
COSEBIs B, 20 modes	20	0.438	0.643
Pure-B ξ₊	7	0.347	0.480
Pure-B ξ₋	7	0.072	0.129
Pure-B combined	14	0.157	0.294

These are controlled noise-diagonal replacements, not fresh full-area OneCovariance runs.
They do not validate mask-dependent mixed or sample-variance terms, and they do not establish the tomographic correction.

Closes #431.

— Codex on behalf of Cail