feat(artifacts): list and download pipeline-run artifacts

[arseniy-pplx]

Summary

• extend tangle sdk artifacts get with --list, --out-dir, --only and --include-children
• accept a run id or root execution id for listing and fetching artifact output
• JSON-encode inline values as <name>.json; stream signed-URL content to a file with owner-only permissions on POSIX (0o600; Windows ACLs are not adjusted) under a sanitized artifact name
• refuse symlinks, collisions and path traversal; check the initial signed-URL host and every redirect hop against loopback, private, link-local and CGNAT address literals by default
• validate --query top-level and nested shapes (tasks/executions/components/artifact_ids) as clean CLI errors; null per-entry output filters mean all outputs (same as []), and an explicit --query '{}' is a valid empty query rather than a missing one
• add the public TangleApiClient.request_raw seam so artifact fetching does not depend on private client methods

Context

The current CLI reports artifact metadata but cannot retrieve artifact contents. This completes the pipeline-run output workflow while keeping large files streaming and applying conservative filesystem, redirect and error-redaction safeguards. Self-hosted deployments with private-network object storage can opt out with TANGLE_ALLOW_INTERNAL_ARTIFACT_HOSTS=1; the host-literal guard is a conservative default, not a DNS-rebinding defense.

Testing

• uv run pytest tests/test_artifacts.py tests/test_artifacts_cli.py tests/test_client.py
• uvx ruff check on the touched source files
• uv lock --check
• git diff --check

[Volv-G]

(AI-assisted)

Could root resolution stay shallow here?

get_run_details() calls get_execution_details(), which recursively _enrich_execution_tree() across every descendant. Consequently, even --list / --out-dir without --include-children makes one API call per descendant and can fail because a nested execution the command never displays cannot be fetched.

Could we add a shallow helper that maps run ID → root execution ID and calls executions_details once, fetching direct children only when include_children is set? A real-client call-count test would pin that boundary.

[Volv-G]

(AI-assisted)

Can we avoid reading/logging the body for streamed requests before this ships?

This stream=True request still traverses _request_with_same_origin_redirects, whose TANGLE_VERBOSE branch evaluates response.text before returning. That buffers the entire artifact—defeating the large-file streaming guarantee—and writes artifact contents verbatim to diagnostics.

PR #39 has the centralized fix: log <streaming body omitted> when request_kwargs["stream"] is set. Since #42 is an independent branch, could this PR either include that guard or explicitly depend on/rebase onto #39?