feat: Surface the health payload's backend and decoding_strategy

[alex-dr]

Summary

HealthMetadata gains optional decoding_strategy and backend fields, so the two informational identifiers the service advertises on /healthz reach SDK callers and jointfm health instead of being dropped during parsing.

Pairs with the service-side change in datarobot/joint (feat/healthz-decoding-strategy), which adds decoding_strategy to the health payload alongside the backend field that landed earlier.

Behavior

Both fields are typed str | None and parsed with _optional_string, which tolerates a missing key and an explicit null alike:

• The service reports decoding_strategy as null for architecture-backend containers and for checkpoint configs predating the field.
• Images built before backend landed omit that key entirely.

So neither field can be required without breaking this client against older deployments. Both flow into jointfm health automatically, since the CLI serializes the record with asdict.

Neither field is validated. decoding_strategy is singular rather than a supported_* list because the strategy is fixed at training time and no request field can select another, so there is nothing for the SDK to check. Mirroring the service's closed vocabulary (parallel_dense, parallel_scalable, autoregressive) as a Literal would only add a failure mode — an older client hard-erroring on a deployment that adds a fourth strategy.

Callers still want the field to interpret responses: only autoregressive couples horizons, so under either parallel strategy samples are per-horizon marginal draws rather than coherent sample paths. That caveat is now documented in docs/api-reference.md.

Tests

• The checked-in health_metadata fixture advertises both fields, asserted in test_fixture_compatibility.py so the fixture stays honest against the service contract.
• Two parametrized tests in test_contract.py cover missing, explicit-null, and non-string payloads for each field.
• test_cli.py asserts both reach the jointfm health JSON output.
• The inline health payloads in test_transport.py and test_configuration.py deliberately keep omitting both fields, so the suite exercises older payload shapes next to current ones.

134 passed, 1 skipped. ruff check, ruff format --check, and ty check all clean.

🤖 Generated with Claude Code

---

Note

Low Risk
Additive, backward-compatible health parsing and documentation only; no changes to forecast requests, validation policy, or security-sensitive paths.

Overview
Health metadata now carries two optional, informational fields from the service: backend and decoding_strategy. HealthMetadata.from_payload maps them with _optional_string, so older deployments that omit the keys or send explicit null still parse without changing compatibility checks.

jointfm health and API docs pick up the same fields automatically (CLI serialization via asdict). The reference notes that decoding_strategy helps callers interpret samples—parallel strategies yield per-horizon marginals, while only autoregressive produces coherent multi-step paths.

Fixtures and contract/CLI tests assert the new fields; parametrized tests cover missing, null, and invalid non-string payloads.

^{Reviewed by Cursor Bugbot for commit a3c3e9d. Bugbot is set up for automated code reviews on this repo. Configure here.}