Asynchronous CUDA normal preparation

Scope	Synchronous preparation	Asynchronous preparation	Observed ratio
Complete worker	57.791 s	51.622 s	1.120x

Repeat	Variant	Complete seconds
0	prior	57.812417
0	current	50.961965
1	current	52.282880
1	prior	57.770263

Two alternating fresh-process pairs replay six seconds of the original public GLIM Livox bag, four map rounds, one BLAS thread, CUDA odometry and CUDA fitter planes. Every worker has its own initially empty CuPy compiler cache. Whole time includes initialization/compilation, all initial/gyro/feedback/final odometry, association, fits/covariance, noise updates, holdout, controls, hashing, saves and frozen-import verification. No warmup or slow worker is discarded.

Baseline 105e1031206ba6c551d08a1b07742b0307e5f523; candidate 1f1b7c46a49cdbadbcfa42761dd9676598968d59; input bag SHA-256 f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643. The unchanged dense policy uses 3 cm voxels, three local scans, a 1 nm coordinate grid and returned-motion grids of 1 micrometre / 0.1 microradian. CPU remains the default and all grids remain disabled by default.

One bounded dense normal batch is submitted before CPU source deskew/preprocess. The target retains GPU temporaries, pinned download buffer and completion event until the unchanged native repairs finish. There is no extra CPU thread. Larger targets and sparse source hints retain the existing demand-driven preparation. Tests cover cross-stream consumption, snapshots, sparse hints, the chunk boundary, abandoned targets, failed native repairs and retries, and empty sources.

The exact gate permits only run metadata and the declared schedule suffix / gpu_async_normal_batches counter to differ. Every calibration/evidence field, old work counter and complete fit field including trajectory and covariance matches exactly. The original physical/evidence checks also pass. This checks preservation of previous outputs; the clip has no independent extrinsic truth. The final policy verdict remains INCONCLUSIVE. All 130 public schema/ledger exports are byte-identical between the two frozen runtime trees.

Host load is unreserved; two pairs do not establish a universal speed ratio. Ratios are not multiplied across scopes or earlier cohorts. The retained normal-preparation timer records host time in submit/finish calls, including native repairs, excludes CPU work between calls and does not measure overlapped GPU work. Its ratio is not a normal-kernel arithmetic speedup.

All four full results/fits/logs, deterministic source capsules/manifests, launcher/schema, publisher, schema checker/receipt and source-bound correctness run are indexed by SHA-256. The audit runs without CUDA. Raw bag data and compiler caches are not redistributed.