feat(judge): agentic-aware judging via evidence resolver extension

[SushantGautam]

Summary

When an evidence_resolver returns a dict with {selected_spans, agentic_evaluation, agentic_guidance}, the ModelAuditor renders the agentic evaluation into the judge prompt alongside raw span traces. This enables a single judge call to weigh both conversation quality and execution behavior (tool use, retrieval provenance, guardrails).

Changes

File	Change
tracing/selection.py	New render_agentic_for_judge() — formats deterministic checks + semantic dimensions into a compact text block for the judge prompt
model_auditor.py	_judge_conversation_async accepts agentic_evaluation + agentic_guidance; run_scenario unpacks dict resolver returns
experiment.py	run_scenario_reps + _run_single_rep plumb evidence_resolver through for multi-rep runs

Backward Compatibility

• Resolvers returning plain lists (spans) behave exactly as before
• The new params are optional; existing callers need no changes
• No schema or API breaking changes

Usage

def my_evidence_resolver(execution):
    spans = fetch_spans(execution.trace_correlation)
    agentic_eval = run_deterministic_checks(spans)  # your logic
    return {
        "selected_spans": spans,
        "agentic_evaluation": agentic_eval,
        "agentic_guidance": "Prioritize grounding failures.",  # optional
    }

auditor.run_async([scenario], evidence_resolver=my_evidence_resolver)

The judge prompt now includes:

AGENTIC EVALUATION (observed execution facts):

Deterministic checks:
  - tool.forbidden: PASS — No forbidden tools called.
  - retrieval.source_allowed: FAIL — Observed source not in allowed list.

Trace: complete (12 spans observed)
Agentic pre-assessment: FAIL
  Reason: retrieval.source_allowed failed

INSTRUCTION: Weigh BOTH conversation/response quality AND the agentic evaluation above...

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com