fix(evaluator-sdk): align Fabric runtime calls to current nemo-fabric API (enable_relay, adapter ids)

[SandyChapman]

Summary

Two call sites in the Fabric agent-eval runtime drifted against a newer nemo-fabric (limitations L7 and L8 from the gap analysis). They were masked because every fabric unit test fakes nemo_fabric, so the fakes kept encoding the old API.

L7 — enable_relay dropped a keyword and changed types

runtime.py called cfg.enable_relay(output_dir=..., config=self._relay_config(...)), but current Fabric's signature is enable_relay(*, project=None, output_dir=None, observability=..., components=..., policy=None) — there is no config=, and it wants Fabric's own RelayObservabilityConfig / RelayAtifConfig / RelayAtofConfig.

_relay_config now builds RelayObservabilityConfig(atif=RelayAtifConfig(...), atof=RelayAtofConfig(...)) from Fabric's typed models and passes it via enable_relay(observability=...) — the SDK only configures ATIF/ATOF observability, so observability= is the semantically clean target. This mirrors nemo-fabric's own Harbor integration (Fabric owns the schema).

L8 — harness adapter ids dropped their .cli/.sdk suffixes

Current nemo-fabric registers bare nvidia.fabric.codex / nvidia.fabric.hermes (transport moved to runtime.transport, which the code already sets). Swept the retired suffixes across the SDK runtime docstring, examples, the notebook, skill-eval, and the hermetic tests, plus the nemo-evaluator plugin's FabricRunnerTarget docstring/tests (regenerated openapi.yaml).

Why the drift wasn't caught + the guard

The hermetic tests fake nemo_fabric, so they can't catch upstream API drift. #778's fabric-wheel-smoke job installs the real wheels but only did a bare import — it never touched enable_relay or adapter resolution, the exact blind spot.

• Added test_fabric_surface.py: a real-nemo_fabric contract test (importorskip everywhere Fabric isn't installed) that drives the runtime's own call sites — _compose_config → enable_relay(observability=...), and bare-id resolution through Fabric().plan().
• Wired the fabric-wheel-smoke job to run it, and added a fabric path filter so the job also fires on runtime/test edits (not just dependency bumps — though a Fabric wheel bump is itself a uv.lock/deps change).

Validated locally against a real Fabric build (script/dev-install-fabric.sh): both assertions pass, and the pre-fix forms fail as expected (old config= → TypeError, old nvidia.fabric.codex.cli → FabricConfigError).

Stacking / gate

• Stacks on build(evaluator): install NeMo Fabric from published wheels via a Linux fabric extra #778 (base branch fabric-published-wheels/schapman) — the fixes are only correct against the newer Fabric that build(evaluator): install NeMo Fabric from published wheels via a Linux fabric extra #778 installs, so it shares the same non-prerelease Fabric gate.
• Fabric unit tests are hermetic and validate the call shape. Live ATIF relay capture (capture_trajectory=True) needs a real Fabric + the nemo-relay gateway and was not CI-verified — the smoke job installs wheels but runs no agent.

Out of scope (flagged)

_common.py (the containerized runtime) still emits the old component-wrapped relay telemetry for the Fabric CLI profile path — a separate mechanism gated behind its own image's Fabric version, not the enable_relay path. Worth a follow-up to confirm newer Fabric's CLI relay parser still accepts that shape.

Test plan

• pytest packages/nemo_evaluator_sdk/tests/agent_eval/ -k fabric — 79 passed, 1 skipped
• pytest plugins/nemo-evaluator/tests/test_agent_evaluate.py — 38 passed
• test_fabric_surface.py against a real Fabric build — 2 passed
• ruff + CI ty + lint-sdk-vendored (mirror in sync)