test: Audit numerical invariants across every mesh interpolator

[Jammy2211]

Summary

Add independent numerical regression coverage for all mesh interpolator classes, including rank and kernel adaptive variants. The tests check applicable linear precision, both-axis cell/edge continuity, supported-region refinement, CDF discretization, kernel-neighbor oracles, guard support and an isolated uniform control. Deliberate mapping/weight, locator, inverse-table and frozen-mesh faults demonstrate discriminatory power.

The audit found two production defects, filed separately: #609 (partial KNN point block) and #610 (Sibson internal-edge coordinates and near-edge partition). Their regressions execute as strict expected failures. This PR does not repair production code or claim every numerical property passes.

Closes #603 once the linked workspace source-recovery PR also lands.

API Changes

None. Production modules and defaults are unchanged.

Test Plan

• Full PyAutoArray suite before the final audit-only extensions: 1,940 passed, 1 strict xfailed, 0 unexpected failures, exit 0.
• Final focused interpolation suite: 90 passed, 4 strict xfailed, 0 unexpected failures, 46.74 seconds; covers every final audit test addition.
• Workspace source-recovery smoke: 1 passed, 0 failed, 8.27 seconds under the actual smoke profile/runner. Compares solved coefficients with independent analytic physical truth for all six classes and rejects deliberately corrupted mappings; unchanged uniform control is bit-identical.
• No new JAX tests are added to the NumPy library suite; optional KNN cases skip when JAX is unavailable because production KNN requires it.

Numerical findings

At mesh sizes 16/32/64, maximum CDF roundtrip errors in index units with the default 64 knots were 0.002764 / 0.018877 / 0.101891. With 256 knots they were 0.000211 / 0.001042 / 0.006067; with 1,024 knots, 0.0000125 / 0.0000771 / 0.000311. These results support a separately validated scaling policy, not a default change in this audit.

Adaptive positive support uses flattened rows 2..n-1 and columns 1..n-2; zero-weight stencil entries can reference additional guards. The conventional mesh zeroed-pixel perimeter has different semantics, retained as a documented boundary-condition question rather than conflated with transformed-cell area repair #605.

Readiness

Heart has no RED reasons. Human explicitly acknowledged manifest drift: workspace checkouts (manifest ↔ disk) — 1 mismatch(es) vs PyAutoMind/repos.yaml and stale release validation incomplete: no rehearsal for current source, authorizing development PR shipping only. No merge or release is authorized.

[Jammy2211]

Linked workspace source-recovery checks and durable audit report: PyAutoLabs/autolens_workspace_test#342. Preserve library-first merge order; issue #603 covers both PRs.