OpenAI Conformance Harness: runnable client with mock and real backend modes

[Zigfreidish]

Child of #1384.

Promote the in-process Conformance Matrix into a runnable Conformance Harness: a client that posts the same conformance rows to a live local endpoint or a configured remote Server Profile, with a deterministic mock-backend CI mode and a real-backend smoke mode, emitting the matrix report as a machine-readable operator evidence artifact.

Reuses the OpenAIConformanceReport schema; does not replace the in-process matrix as the authoritative proof (see docs/adr/0002-openai-conformance-at-control-plane-boundary.md).

Done when

• Harness runs against a live local endpoint and a configured remote Server Profile.
• Mock-backend mode is deterministic and CI-runnable without model weights.
• Real-backend smoke mode produces an evidence artifact.
• Failure output names the incompatible field/phase rather than a generic error.

[Akane-CN]

Design principle: live conformance needs to cover client bootstrap probes and route-specific model-name semantics, not only request JSON schemas.

Melix-specific goal: extend the runnable compatibility harness so a local endpoint is exercised the same way real SDK clients do: unauthenticated GET/HEAD / liveness probe, chat/messages route dispatch, strict OpenAI-style unknown-model rejection, and route-local passthrough for provider-native model names when the loaded runtime is intentionally acting as that compatibility target.

Measurable target: add harness rows that prove (1) root probe returns 2xx without exposing privileged state, (2) OpenAI chat still rejects unknown model IDs with a typed 404, (3) native-message compatibility can accept provider-shaped model names while returning the loaded model identity in receipts, and (4) streamed tool-call deltas never leak parser sentinels or emit empty-name clobbers under byte-by-byte and multi-tool chunking.

Next executable slice: add mock-backend fixtures for fused/split tool-call boundaries, literal sentinel strings inside JSON arguments, and a probe-before-request sequence; wire the failure output to name the incompatible phase (probe, model_validation, tool_delta_assembly).