decisiongrounding: OpenAI-compatible (LiteLLM) answering backend — first real-gateway validation

[tcballard]

What

The decisiongrounding benchmark now supports OpenAI-compatible gateways (the common enterprise LiteLLM config) via a new answering backend:

• --answering litellm:<model-alias> (OpenAICompatAnsweringModel), speaking /chat/completions + response_format: json_schema over stdlib urllib — no new dependency.
• Held-constant contract preserved (DG-ADR-0001): the litellm and claude backends share the same scaffold, user-prompt builder, and ProposedChange schema; only the wire format differs. Covered by tests, including a localhost mock-gateway round-trip.
• scripts/litellm_probe.py --mode openai probes the exact adapter request before any funded spend.
• ANSWERING=litellm:<alias> ./scripts/run_real.sh wired; --batch/BATCH=1 refuse the combination (no Batch API on the OpenAI surface).
• Recorded in decisiongrounding/rac/decisions/ADR-0005-openai-compat-answering-adapter.md (Accepted).

Remaining work this issue tracks

• Validate against a real LiteLLM gateway (both probe modes; confirm response_format json_schema translation for the chosen alias and that usage is forwarded).
• Confirm the gateway alias is pinned to a fixed model before any recorded run (version honesty rule in ADR-0005).
• First gateway-routed real run (Step 2 cheap validation from docs/AGENTIC_BENCHMARK_RUN_HANDOFF.md) and note any gateway-specific behaviour back into the README proxy section.

Why

Enterprise adopters route all model traffic through LiteLLM; supporting only the Anthropic-native passthrough made the benchmark unrunnable exactly where the rac-grounding thesis gets evaluated for adoption.

Roadmap: decisiongrounding/rac/roadmaps/publish-swe-decisionbench.md (multi-model generalization initiative — per ADR-093, execution tracked here).