[043][Phase 5][US3] Prompt override surface + version discipline

[jwesleye]

Scope

US3 MVP — declarative prompt configuration: every judge-backed evaluator accepts custom template, few-shot examples, system prompt, output schema, use-reasoning flag, and feedback key without code change. Prompt version is always recorded in results.

Priority: P1 (MVP)

Tasks

• T104 [P] [US3] Tests in eval/tests/us3_custom_prompt_test.rs — .with_prompt(custom) replaces built-in, few-shot injection order, version bump is explicit opt-in, missing variable is construction-time error
• T105 [US3] Extend every judge-backed evaluator builder with .with_prompt(), .with_few_shot(), .with_system_prompt(), .with_output_schema(), .with_use_reasoning(), .with_feedback_key() — all route through JudgeEvaluatorConfig
• T106 [US3] Verify prompt_version recorded in every EvalMetricResult::details (re-verification of T056 in context of all evaluators)
• T107 [US3] Versioning smoke test — correctness_v0 (built-in) and correctness_v1 (custom) both resolvable; results distinguish them per-metric

Acceptance

• A consumer can override the built-in CorrectnessEvaluator prompt with a custom template + few-shot examples via builder methods only; no evaluator code is modified.
• The custom template's version string appears in the resulting EvalMetricResult::details.
• Bumping a prompt version (_v0 → _v1) is explicit; old version remains accessible.
• Missing or misspelled template variables surface at evaluator construction time.

References

• Spec FR-010, FR-011, Success criterion SC-005

Depends on

#750 (evaluators must exist to override).

[unseriousAI]

Landed a partial US3 advance directly on integration in 6871aa4.

This run wires the evaluator-side prompt override surface from T105:

• added .with_prompt(...), .with_few_shot(...), .with_system_prompt(...), .with_output_schema(...), .with_use_reasoning(...), and .with_feedback_key(...) across the judge-backed evaluator surface (quality, safety, RAG, agent, code LLM judge, multimodal)
• centralized prompt render context construction in build_prompt_context() so evaluator-level overrides now actually reach template rendering instead of being dropped by per-family helpers
• changed PromptContext::with_custom(...) to merge keys so evaluators can layer custom metadata without overwriting shared config metadata
• added eval/tests/us3_custom_prompt_test.rs covering custom prompt replacement, explicit version opt-in (correctness_v0 vs correctness_v1), construction-time missing-variable rejection, and builder availability across representative judge-backed evaluators

Validation:

• cargo test -p swink-agent-eval --features "judge-core,evaluator-quality,evaluator-safety,evaluator-rag,evaluator-agent,evaluator-code,multimodal" --test us3_custom_prompt_test
• cargo test -p swink-agent-eval --features "judge-core,evaluator-quality,evaluator-safety,evaluator-rag,evaluator-agent,evaluator-code,multimodal" --test evaluators_quality_test --test evaluators_safety_test --test evaluators_rag_test --test evaluators_agent_test --test evaluators_code_test --test evaluators_multimodal_test
• cargo clippy -p swink-agent-eval --features "judge-core,evaluator-quality,evaluator-safety,evaluator-rag,evaluator-agent,evaluator-code,multimodal" --all-targets -- -D warnings

Issue stays open because T106/T107 still need the broader per-evaluator prompt-version re-verification and versioning smoke coverage.

[unseriousAI]

Completed US3 on integration in 45d9672.

This run finishes the remaining scope from T106/T107:

• expanded eval/tests/us3_custom_prompt_test.rs with a full sweep asserting every judge-backed evaluator records its built-in prompt_version
• added a registry-level smoke test proving builtin correctness_v0 and custom correctness_v1 prompt variants can coexist and be selected independently

Validation:

• cargo test -p swink-agent-eval --features "judge-core,evaluator-quality,evaluator-safety,evaluator-rag,evaluator-agent,evaluator-code,multimodal" --test us3_custom_prompt_test
• cargo clippy -p swink-agent-eval --features "judge-core,evaluator-quality,evaluator-safety,evaluator-rag,evaluator-agent,evaluator-code,multimodal" --all-targets -- -D warnings

Closes #754.