[043][Phase 3][US1e] Registry wiring + US1 end-to-end integration

[jwesleye]

Scope

US1e — registry wiring + end-to-end integration. Final wiring for the MVP's US1 layer: unique-name validation, crate-root re-exports, and the end-to-end integration test that lifts the US1 acceptance gate.

Priority: P1 (MVP)

Tasks

• T087 [US1] Extend EvaluatorRegistry::add to reject duplicate evaluator names with EvalError::DuplicateEvaluator (spec edge case)
• T088 [US1] Update eval/src/lib.rs public re-exports covering every US1 surface under correct feature gates
• T089 [US1] Integration test eval/tests/us1_end_to_end_test.rs: registry with one evaluator per family, wiremock judge with canned verdicts; assert per-evaluator score + non-empty reason + prompt_version recorded; non-applicable evaluators produce no entry

Acceptance

• US1 acceptance scenario from spec.md passes end-to-end with MockJudge / wiremock.
• cargo test -p swink-agent-eval --features all-evaluators is green.
• Duplicate evaluator registration returns a structured error, not a panic (spec edge case).

References

• Spec US1 acceptance scenarios 1–6
• SC-001 (10-line happy path)

Depends on

#748, #749, #750, #751.

[unseriousAI]

Advanced #752 with the duplicate-name edge case from T087.

What landed on integration in a7ba0fe:

• added EvalError::DuplicateEvaluator { name }
• added EvaluatorRegistry::add(...) -> Result<(), EvalError> that rejects duplicate Evaluator::name() values
• kept register(...) as the infallible convenience wrapper, now backed by add(...)
• added unit/integration coverage for duplicate registration failures

Validation run:

• cargo fmt --all --check
• cargo test -p swink-agent-eval --test evaluator

#752 stays open because T088/T089 are still outstanding.

[unseriousAI]

Advanced #752 with the missing US1 acceptance coverage from T089.

What landed on integration in 44cd34b:

• added eval/tests/us1_end_to_end_test.rs using only crate-root swink_agent_eval::* imports so registry wiring and eval/src/lib.rs re-exports regress together
• covered one evaluator per US1 family slice exercised here (correctness, harmfulness, rag_groundedness, task_completion, json_schema, exact_match, code_llm_judge, image_safety)
• asserted judge-backed metrics record prompt_version + non-empty reasons, and that non-applicable evaluators are omitted from registry output
• documented the crate-root acceptance-test convention in eval/AGENTS.md

Validation run:

• cargo test -p swink-agent-eval --features all-evaluators --test us1_end_to_end_test
• cargo clippy -p swink-agent-eval --features all-evaluators --test us1_end_to_end_test -- -D warnings

#752 stays open for now because the broader cargo test -p swink-agent-eval --features all-evaluators suite did not finish within this automation run's timeout window, so I did not claim the full acceptance matrix.