Maintained CPU grid rotation sharing

Complete evidence, 3,490,476 bytes, contains 871 indexed files: both 401-module source trees, one frozen timing toolchain, input receipt, schemas, all six complete results/configurations/measurements, all 180 per-update observations, original logs, and packaging provenance. Raw images and scans are read in place and are not included.

Archive SHA-256: 8e604eba37e5b12e2ba0e8dc1d5ec292f75d6cdf51cc7c6b2acffa6d44ab28bb. The inventory binds the archive, internal index and original summary, which follows this schema.

The baseline is the maintained Numba backend at b7d6203fa3024a78b6693e7536329b75d2b1dc6e; the candidate is maintained source c2e694e106ebdfec290538b7fd018f45c17ead7f. Both include the typed PNG array reader. All 48 prepared frame snapshots and all six complete quality payloads match. Each worker executes 30 rolling updates, 21,958 objective evaluations and 395,476,000 logical point rows, including holdout, controls and curvature.

Median of three fresh workers	Previous maintained CPU	Shared rotations	Ratio
Complete solve API, including first projection JIT	22.018838 s	5.964614 s	3.691578x
Median online update, all 30 retained	0.681927 s	0.126579 s	5.387368x
First update, including projection JIT	2.420839 s	2.167978 s	1.116635x
Complete worker, including startup and preparation	32.604431 s	16.219636 s	2.010183x

All workers use the same 48 already-spent KITTI image/cloud pairs, three ordered preparation workers, one BLAS thread and no discarded warmup. The host is unreserved. Input verification warms the filesystem cache. The predeclared 2x gate concerns the complete solve API; startup is a separate timing scope. This cohort is not pooled with earlier prototypes and establishes no new calibration accuracy claim or portable latency guarantee.

python tools/audit_online_grid_reuse_archive.py \
  docs/assets/online_grid_reuse/complete-evidence.tar.gz --verify-git

The audit checks the indexed bytes and both committed source trees without executing archived Python or importing optional acceleration packages. Omit --verify-git for an offline archive-only audit. See the benchmark description for timing boundaries, bounded buffers and reproduction.