Affine and Differentials graduation audits

These audits record evidence as of 2026-07-16. They are deliberately stricter than an implementation checklist: a module can be capable and well tested without yet carrying a stable compatibility promise. Neither module is promoted by this review.

Affine audit

AffinePlan graduation gates

Gate	State	Evidence	Remaining work
Geometry and numerical contract	Pass	2-D/3-D pull-back matrices, output geometry, degrees 0–7, mirror/zero modes, analytical cases, and SciPy parity for matched degrees 0–5.	Keep higher-degree claims SplineOps-specific.
Axes, dtype, mutation, and buffers	Pass	Explicit batch/channel axes, float32/float64 execution, source preservation, exact and strided output buffers, and randomized scalar dispatch parity.	None for the documented surface.
Bounded resources	Pass	Cached plans enforce max_retained_bytes; one-shot tiled execution and retained-byte inspection are tested and benchmarked.	Continue platform monitoring.
Coefficient reuse and persistence	Pass	Immutable compatibility fields, atomic schema-2 archives, permanent schema-1/schema-2 byte fixtures, checked migration, corruption rejection, and endian/metadata validation.	Preserve every historical fixture when a future schema is added.
Concurrency and restart behavior	Evidence collecting	Thread reuse and spawned-process loads pass unit tests and the repeated public-API soak; the cross-platform soak workflow publishes JSON.	Accumulate real workflow runs without contract changes.
Performance claims	Pass, narrowly scoped	Complete-workflow gains exist for coefficient sharing, but matched one-shot SciPy affine execution remains faster. Profile evidence, not a universal speed claim, governs implementation changes.	Re-profile before any further optimization.
Compatibility duration and migration	Open	The public contract and failure modes are documented.	Complete an API-soak period and prepare migration guidance if names or buffer rules change.

Decision: AffinePlan is technically strong but remains experimental. The open gate is compatibility evidence over time, not a missing attempt to win every one-shot timing comparison.

Differentials audit

DifferentialPlan graduation gates

Gate	State	Evidence	Remaining work
Derivative convention	Pass for current scope	Increasing-coordinate gradients, packed upper-triangular Hessians, Laplacians, physical spacing, analytical polynomial/trigonometric invariants, and whole-sample mirror boundary fixtures.	Do not imply boundaries other than the documented cubic mirror model.
Axes, dtype, mutation, and buffers	Pass	2-D/3-D explicit spatial axes, integer promotion, float32/float64 preservation, selective families, prevalidated exact/strided structured buffers, non-aliasing, and randomized scalar dispatch parity.	None for the planned API’s documented surface.
Resource behavior	Pass	One lazy workspace shares coefficient and derivative intermediates; Laplacian-only calls skip mixed Hessians and the plan retains no arrays.	Continue batch-size monitoring.
Concurrency and repeated buffers	Evidence collecting	The repeated volume-feature soak preserves source data and destination identities with exact scalar parity.	Accumulate cross-platform soak artifacts without contract changes.
Legacy relationship	Open	DifferentialPlan provides the general planned numerical API while Differentials preserves scalar convenience and 2-D angular maps.	Publish migration guidance before changing or deprecating either name.
Compatibility duration	Open	Errors and output selection are explicit and covered.	Complete the API-soak period; keep angular behavior outside a general N-D claim unless it receives its own contract.

Decision: DifferentialPlan also remains experimental. It is ready for serious evaluation, but stabilizing it today would turn a short, internally reviewed history into a compatibility promise prematurely.

Next review trigger

Review these tables after a meaningful period of unchanged downstream use and cross-platform soak artifacts—not merely after another green unit-test run. Promotion can then happen module by module. TensorSpline, affine, differentials, resize, and the research modules remain independent APIs throughout that process.