Qualification scripts

architecture/scripts.toml inventories every source file under scripts/. Each entry declares its owner, role, qualified gate/journey, direct workflow/job callers, retained job evidence and whether CI reaches it. Helpers inherit the evidence of their calling jobs; resource entries describe inputs rather than claiming to emit reports. Evidence paths are caller-selected templates from artifact steps, alongside stdout/stderr. They are not a complete dynamic call graph or a claim that each helper writes every file retained by its job.

Implementations are grouped by responsibility:

Package	Responsibility
scripts.qualification.ci	Repository policy, dependencies, architecture, coverage, security and CI measurements
scripts.qualification.docs	Site identity, publication validation and executable reader examples
scripts.qualification.packaging	Coordinated candidates, installed distributions, Homebrew and release evidence
scripts.qualification.extensions	Plugin preparation, catalogues, profiles, synchronization and cleanup
scripts.qualification.interfaces	MCP clients, schemas, REST collections and onboarding
scripts.qualification.experiments	Scientific contracts, SDK capture, notebooks, history and installed journeys
scripts.qualification.delivery	Operator authority, artifact/image delivery, Attest receipts and admission
scripts.qualification.repository	Git acquisition, SSH/HTTPS and source provenance
scripts.qualification.fixtures	Disposable protocol, signing, registry and isolation fixtures
scripts.mcp_distribution	Existing named helpers for official MCP distribution and index verification
scripts/client_collection_tools	Exact native vendor manifests and compatibility tooling

The root Python commands remain supported, for example python scripts/check_package_architecture.py. They contain only a checked dispatch shape and forward to one implementation. Internal Python imports use the canonical package, and development workflows normally use module entry points from the checkout root:

uv run --locked python -m scripts.qualification.ci.check_script_architecture
uv run --locked python -m scripts.qualification.ci.check_package_architecture

Some installed-interpreter calls deliberately retain the stable CLI path. Standalone probes run the actual file without adding checkout directories to sys.path, preserving python -I and copied/source-free driver behavior. Their inventory entries declare source_free; they cannot import repository qualification helpers. Keep copied proof files self-contained and copy the implementation rather than its dispatcher. Historical release evidence and external Action/GitLab commands retain their existing root CLI names.

For a new proof, choose the package that owns the journey and a descriptive *_proof.py or smoke_*.py name. Role values are smoke, proof, check, prepare, release, fixture and tool. Add the inventory entry with its gate and callsites in the same change. Namespace initializers only compose; shared helpers belong in a named responsibility package, never utils, helpers, common or services. Do not add new root implementations or copy proof logic into a compatibility command.

The ownership gate rejects missing/stale files, absent workflow script/module targets, stale callsites/evidence, misplaced implementations, executable namespace initializers, business logic in dispatchers, compatibility imports inside implementations and checkout imports in source-free probes. Generated Python caches and the vendor tool's installed node_modules are excluded; their tracked manifests and compatibility check remain inventoried. The gate runs in pre-commit and the CI quality job. Pyright's existing recursive scripts include covers canonical packages and compatibility commands together.

The initial reference audit found one script without a workflow, documentation or other-script caller: completion_model.py. It is retained as a documented developer tool because it regenerates the captured modern CLI grammar used by completion conformance tests. After reviewing a deliberate CLI grammar change, regenerate the snapshot with:

uv run --locked python -m scripts.qualification.ci.completion_model

Review the resulting src/apizr/cli/completion_spec.py diff with the CLI change. This command is a development operation, not part of normal qualification.

Implement related tickets locally, use focused checks while the batch evolves, then push and qualify the whole batch once. Moving a proof or adding a ticket does not require starting another full platform matrix in the middle of that batch. The separate pytest performance comparison is manual only.

Academy source/shot-list preparation uses python3 -m scripts.qualification.docs.academy_lessons --output /absolute/new/directory. See the existing recording process; this command does not render or publish videos. The loopback-only python3 -m scripts.qualification.interfaces.telemetry_conformance --endpoint http://127.0.0.1:8080 --output /absolute/report.json checks a disposable receiver against the public synthetic corpus. It never qualifies production hosting, storage or retention.

The engine performance baseline is also manual only. Its scripts.qualification.experiments.benchmark_engine command records import, inspection, SDK, file digest, publication, history and editor timings, with function-call profiles. See the measurement method and budgets.