Selective runtime loading

The CLI imports each command's dependencies when that command runs. Solver package exports resolve to their original objects on first access. Artifact validation loads the selected model and schema generator, while listing kinds uses lightweight declarations. This removes unrelated initialization from short jobs without changing their parsing, validation, optimization or output.

Complete fresh CLI command	Previous median	Current median	Ratio
Probabilistic Camera–LiDAR refinement and save	7.758 s	3.003 s	2.584x
Validate a 9,216-correspondence artifact	6.975 s	1.821 s	3.830x
Help	6.460 s	1.820 s	3.549x

Each case retains three alternating fresh-process pairs: prior/current, current/prior, prior/current. All 18 attempts succeeded on an unreserved host with one BLAS thread. No warmup or slower attempt is removed. Timing includes interpreter startup, normal CLI parsing and execution, and a small import-path receipt. Freezing and checking sources, external result validation and schema export comparison occur outside the timed process. The previous runtime already includes batched pose evaluation; its different worker timings must not be pooled or multiplied with this comparison.

The workload reuses that benchmark's nine public KITTI raw scans and 9,216 generated image correspondences, with mixed camera models and fixed noise. These correspondences test execution and parity; they are not independent camera measurements or an accuracy evaluation. Every refinement field and all 325 trace entries agree across all six runs, including acceptance and holdout. Only creation time, source-commit provenance and the CLI's declared output path are normalized. Validation and help outputs also agree completely.

All 130 schema and ledger exports are byte-identical between frozen runtimes. Accepted versions still come from model Literal annotations: namespace hints only prioritize candidate models, and a complete scan handles older or unrelated namespaces. Aliases, migrations, retired-version errors, digest verification and the public SCHEMA_REGISTRY tuple remain available. Concurrent first access returns stable cached entries, and imports happen outside the registry cache lock. Long-running jobs that already loaded these dependencies should expect less benefit.

The final runtime passes 155 targeted checks covering CLI dispatch, public solver identities, concurrent schema loading, migrations, digest checks, committed artifact readability and complete refinement parity. Four additional archive checks pass after binding the final correctness receipt.

The indexed archive contains every attempt, full refinement output, CLI stdout/stderr, actual import paths, source capsules and manifests, the launcher, input hashes and schema parity receipt. Raw dataset coordinates, generated correspondence files and model weights are not redistributed. The CLI's git_commit field records the enclosing checkout; the separate frozen-source manifests and import receipts identify the runtime actually executed for each variant.

Reproduce

Use the fixture builder and nine scan identities documented in the pose batching benchmark. At measured candidate c0402f431902c18cb6f5c61b8aaa6d505ca2823a, run:

python -m tools.benchmark_lazy_runtime_loading \
  --baseline-commit e7b89520f5391005c62016fd2549d39de0693c09 \
  --correspondence outputs/projected-control/correspondence.json \
  --problem outputs/projected-control/problem.json \
  --repetitions 3 --output-dir outputs/cli-comparison
python -m tools.audit_lazy_runtime_benchmark docs/assets/lazy_runtime_loading

Use a new output directory. The CPU-only audit verifies the inventory and source capsules, limits runtime changes to the three loading files, checks every output and input binding, and recalculates medians from all attempts.