Python(feat): accept list values for multi-value metadata writes

[dengjonathan]

What

Stacked on #699 (read side). Completes the client story for multi-value metadata: list values can now be written through the public API.

client.runs.update(run, RunUpdate(metadata={"part_number": ["ABC", "XYZ"], "env": "prod"}))
run.metadata.getall("part_number")   # ["ABC", "XYZ"]

• metadata_dict_to_proto accepts list[str] values, emitting one MetadataValue per element in list order (the canonical order). Lists must be non-empty and string-only, mirroring the backend rule that only string keys are multi-value capable. The parameter widens to Mapping so existing scalar-dict callers typecheck unchanged.
• RunCreate / RunUpdate / AssetUpdate widen metadata to dict[str, str | float | bool | list[str]] and gain a before-validator that expands Metadata mappings via the new expand_metadata_for_write helper. Without it, pydantic's dict rebuild (and the model_dump feeding proto conversion) flattens the mapping to its first-value view — so update(metadata=run.metadata) would silently drop values. With it, the round-trip is lossless.
• The streaming run form (_to_rust_form) rejects list values with a clear error until the Rust bindings support multi-value.

Scope

Run and Asset writes only — the entities in the ENG-13281 rollout. Report, channel, test-report, and calculated-channel models keep scalar annotations; widening them later is a per-model annotation + validator once backend support is confirmed.

Testing

• Unit: 8 new tests across test_metadata.py (list conversion order, round-trip, Metadata expansion, empty-list and non-string rejection), test_run.py (create/update list writes, Metadata round-trip, rust-form guard), test_asset.py (list writes, Metadata round-trip). Full non-integration suite: 1139 passed.
• Integration: TestMultiValueMetadata gains public-write coverage — test_update_run_with_list_metadata and test_metadata_round_trip_preserves_all_values — verified green against a local stack with the multi-value-metadata flag enabled (skip cleanly when it isn't). Full test_runs.py integration file: 33 passed.
• ruff check / ruff format --check / mypy / pyright clean; regenerated sync stubs unchanged.

Linear: ENG-13281 (Phase 3)

🤖 Generated with Claude Code

[github-actions]

Python docs preview: https://sift-stack.github.io/sift/python/pr-709/

Deployed from 0d54905. The link may take up to a minute to become live as GitHub Pages propagates.