Indexed online-edge CPU execution evidence

The complete archive retains every worker from the final six-process cohort and all complete results/configs, measurement records, source/tool snapshots, generated schemas, input receipt and stdout/stderr. Its index binds 449 files by SHA256 and size, including the packaging source and project license. Raw images/scans and compiler caches are excluded.

Implementation	Whole-worker median
Production NumPy	51.174 s
Compiled benchmark control	23.452 s
Maintained Numba CPU backend	23.724 s

The maintained backend measured 2.157x against production NumPy and 0.989x against the compiled benchmark control. It makes the compiled control's performance available through supported APIs. All 48 prepared frames and all six complete calibration result payloads match exactly, excluding only separately validated execution metadata. Each implementation has two cold observations on an unreserved host. Default CPU serialization is unchanged.

The measured source commit is bb90e12f2fa2b811a10ad92dfe222aaffef7831f. All 401 package modules and eight benchmark tool/schema files are frozen. The comparison is also available separately. The archive checksums cover both cohorts.

The initial image-JIT archive preserves all six earlier workers from ccf3f872663c724dc9bc1d00a4279f35853b6bab. Its whole-worker medians were 55.588 s NumPy, 26.107 s compiled control and 26.036 s Numba, with complete parity. It is a separate development cohort, not pooled with the final measurement. Its index also binds 449 files.

python tools/audit_online_edge_cpu_archive.py \
  docs/assets/online_edge_cpu/complete-evidence.tar.gz \
  --bind-commit bb90e12f2fa2b811a10ad92dfe222aaffef7831f \
  --output outputs/online-edge-cpu-archive-audit.json

The auditor uses CPU-only JSON/schema/hash checks without importing Numba, CuPy or Calibrex, or executing archived Python. It applies the shared archive limits, safe-path checks, complete index validation and local-only schema references. Optional Git binding checks normalized UTF-8 source/tool content.

The execution guide describes reproduction from existing inputs. Both cohorts reuse the same 48 already-spent real KITTI image/cloud pairs. Every input identity is pinned in the acquisition receipt; the dataset calibration initializes an execution control and supplies no new extrinsic accuracy or SOTA evidence. One frozen tree serves all workers, with less than 9 MiB of uncompressed evidence per cohort and a 64 MiB run limit.