Saved endpoint-derivative curvature diagnostic

Use the twelve completed fixed nonlinear probes and their saved zero-expansion baselines. Compare the reference linear expression with the endpoint derivative average at all 293 epochs, retaining every physical derivative column. At the six epochs with existing shared-source maps, compare all three expressions. Do not interpolate maps or silently reduce the full-history derivative check.

The endpoint average approximates the derivative line integral. Check exact affine/quadratic examples, a cubic counterexample and missing-column failures before execution. Do not fit coefficients to the outcomes or adjust any point. Decompose the directional curvature expression into the original physical parameter groups; these are algebraic contributions, not unique causal shares.

Finite maximum: two identities, twelve point comparisons/decompositions, 10,548 coordinate comparisons per derivative expression and 216 shared-source coordinate crosschecks. Reserve 569,592 physical-column products. No native calls, source oracles, LP proposals, aircraft reads or model-fit starts. Use one numerical thread with other allocations reserved, a 120-second cooperative ceiling, no retries, process signals or allowance expansion.

Freeze source, environment and hash references to the existing sealed arrays before launch. Verify each referenced file before reading it. Crosscheck the reference expression against the prior shared-map audit within a fixed 1-micrometre numerical tolerance; this is not a sensor-accuracy allowance. Publish only from sealed results.

This is not a curvature bound, nonlinear uncertainty coverage, an observation error model or a scientific model comparison. It identifies where a separately proved bounding method needs improvement.