refactor: chunk the kernel-CDF forward transform (512-query blocks)

[Jammy2211]

Summary

Removes the kernel-CDF meshes' O(M×N) memory wall (#376): the exact forward F(q) = Σᵢ wᵢ·Φ((q−xᵢ)/h) previously broadcast an (M, N, 2) array — ~60 GB at production imaging scale (M≈246k over-sampled queries × N≈15.4k traced points, observed OOM). It now evaluates in fixed KERNEL_FORWARD_BLOCK = 512 query blocks: lax.map over padded blocks under jax (static shapes, AD flows through the scan), a plain block loop under numpy. Peak block memory ≈ 126 MB at that scale; measured end-to-end: 1.06 GB peak RSS where the unblocked code failed to allocate. CPU wall-time at that scale is ~10 min/eval (2×10⁹ erf evaluations — the arithmetic, not the blocking); GPU remains the production target.

API Changes

None — internal changes only (one additive module constant, KERNEL_FORWARD_BLOCK).

Test Plan

• Refactor invariant — float-identical values: block-invariance unit test (batched == row-by-row across a non-multiple-of-block query count); full suite 894 passed
• Both jax_grad certification scripts pass on this branch with byte-identical FoM/parity values to the unblocked run (imaging 7 variants, interferometer 5)
• Function-level JAX sanity: eager == JIT; AD == FD through lax.map
• Previously-OOM scale probe (M=246,080 × N=15,380): completes at 1.06 GB peak RSS

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: safe (refactor cap; human-directed follow-up "do these two")
• Plan: on issue refactor: chunk the kernel-CDF forward transform (remove the O(M×N) memory wall) #376
• Gate: tests 894 + certification byte-identical · smoke n/a (jax_grad scripts are the downstream surface, both re-run) · review CLEAN · Heart YELLOW acked in-session (same 6 stale reasons)
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

🤖 Generated with Claude Code