Bridge decision-capable registered models to DecisionTarget and surface the answer in run detail

[avalyset]

Covers items 2 and 4 from the gap analysis in #17 (comment 6081082149); items 1 and 3 are SushantGautam's. Builds on c468a56 and f4e91d2.

What

A registered model that /connections/ probed as a decision model (RegisteredModel.is_decision) can now be used as an audit target, and the structured answer it returns — choice, per-option probabilities, calibrated confidence — is shown in run detail.

Item 2 — DecisionTarget bridge (infra/engine.py, infra/trace_target.py)

• The bridge. build_model_auditor() builds a DecisionTarget and calls auditor.set_target(...) when the target snapshot carries capabilities.decision. The snapshot already copies RegisteredModel.capabilities (audits.services._endpoint_snapshot), so target setup never re-probes: a run uses what was true when it was frozen.
• No provider allow-list. The gate is that flag alone. Since f4e91d2 probes every non-OpenRouter server at the same {base_url}/v1/systemone surface, target setup needs no per-provider endpoint logic either — a provider you add later works here unchanged. The .ollama() factory builds that URL and applies the 64 KiB body cap; a connection's own /v1 suffix comes off first, the same normalisation probe_systemone_decision does. OpenRouter is the one exception: a fixed remote URL whose key is resolved from the connection's secret_reference, like every other role, so a raw key never enters a snapshot.
• max_turns forced to 1. Done in auditor_kwargs rather than at either construction site, so both run paths agree — the repetition runner derives its model entry and max_retries_per_rep from the same kwargs.
• install_trace_context_target and DecisionTarget — checked. It does not cover one, and must not be applied to a decision run: it wraps auditor.target_client in a TraceContextModelTarget, a ModelTarget over an OpenAI-compatible chat client, which would replace the DecisionTarget outright. Nothing is lost by skipping it — DecisionTarget.send merges TargetContext.trace_headers into its own request headers, so per-turn traceparent still propagates. test_a_decision_target_forwards_the_per_turn_traceparent pins that, so it fails if the engine ever stops doing it.
• The unused chat client is not built. SimpleAudit's own Auditor facade does the same (_skip_target_client), and building it would demand the provider's any_llm extra and an API key for a client that is never called: a local Ollama decision run otherwise fails on any-llm-sdk[ollama] not being installed.
• Both run paths. The single-repetition path, and the repetition runner, which builds its own auditors and so needs the choice made on the class (decision_auditor_class).

Item 4 — picker and run detail (infra/ui.py, templates/)

• New Experiment picker: decision-capable models carry the same decision badge as on /connections/, with the constraint spelled out per role — a single-turn run needing a scenario with a decision block for the target, "answers options, not prose" for judge and auditor.
• Run detail: _rep_view carries the answer per reply (reply["decision"], which ModelAuditor copies from TargetResponse.decision) and per repetition, a later turn winning. The result panel renders a "Decision answer" card: the chosen option, confidence N% when the model reported one, and a probability bar per option, highest first. A chat reply has no such block and renders unchanged.

Not in this PR

• Items 1 and 3 (choice_match judge surfacing; the scenario decision block). Nothing here touches ScenarioRevision, scenario_dict() or the scenario form — the end-to-end test builds its decision scenario as a dict in the test, so this is mergeable before or after that work.
• Dependency on item 1: restricting the picker by which scenarios a set contains needs the decision field from item 3, so the badge is advisory for now. And a decision run is judge-less in practice — _validate_secrets demanding a judge secret_reference is item 1's call, which is why the tests here construct the auditor with the code-only choice_match judge directly rather than through a run's judge snapshot.

Tested

Tested end-to-end against clef-flash (9.1B, Q8_0) on local Ollama 0.35.1; the recorded response is the CI fixture (infra/fixtures/decision_clef_flash.json). CI runs Python 3.12.

The fixture is replayed through an httpx.MockTransport, so everything but the HTTP hop is the real chain: frozen snapshot → DecisionTarget → engine run with the code-only choice_match judge → _rep_view → rendered panel. test_live_clef_flash_answers_and_renders runs the same path against the real server and skips when no local Ollama serves the model.

54 new tests. Ruff clean. Five pre-existing failures (chat/tests/test_sync.py signal counts, infra/tests/test_db_resilience.py SQLite locking) and four errors (infra/tests/test_minimal_config.py embedded Hatchet lifecycle) reproduce identically on a clean f4e91d2 checkout, so they are not from this branch.

Refs #17

AI assistance. Implemented with Claude Code (Opus 5); planning and review with Claude Fable 5.1. The author ran the test suite, the live clef-flash run and the diff review locally before submission.