Reproduce the artifact-contract diff CLI from a clean clone

[stacknil]

Review target

Reproduce the standalone artifact-contract diff CLI from a clean clone using tiny sanitized artifact trees. This is a focused follow-up to the implementation in #85 and #90-#92; it is narrower than the earlier five-demo operator-path reproduction in #77.

Relevant paths on the current default branch:

• scripts/artifact_contract_diff.py
• docs/reviewer-artifact-diff.md
• schemas/artifact_contract_diff.schema.json
• tests/test_artifact_contract_diff_core.py
• tests/test_artifact_contract_diff_structured.py
• tests/test_artifact_contract_diff_json_report.py

Reproduction

1. Record the default-branch commit, OS, Python version, and clean-clone install command.
2. Create two disposable, sanitized directory pairs: one byte-equivalent text/JSON case and one case with a small reviewer-visible difference.
3. Run the documented human-readable comparison for both pairs.
4. Repeat with --json-out located outside both input roots, then run the three focused test files above.

The documented contract is exit 0 for unchanged artifacts and exit 1 for detected differences. Repeating the same JSON-report command on the same trees should be byte-identical. Output must not disclose artifact bodies, timestamps, or absolute checkout paths.

Evidence to post

• Exact commands and exit codes.
• A concise description of relative paths/statuses reported for both cases.
• Whether repeated JSON output is byte-identical and validates against artifact-contract-diff/v1.
• Focused test results.
• One classification: reproduced as documented, documentation ambiguity, environment-specific limitation, or probable defect.

Change threshold

Post reproduction evidence before opening a PR. Only propose a change for a concrete mismatch, and preserve the smallest affected CLI, contract, test, or instruction.

Boundaries

• Keep fixture trees tiny, synthetic, and disposable.
• Do not include real telemetry, secrets, usernames, absolute machine paths, or artifact bodies in the issue.
• Do not accept regenerated outputs, change release compatibility labels, or broaden the comparator's fail-closed path semantics.