Gate Agent answers on native citation checks

[acgxv]

Change

Block final Agent answers when any native citation binding fails, including the last tool turn. Tool submissions and direct JSON readers now share one finalizer over the existing evidence-check API. Retain rejected drafts and verdicts, allow correction only within the existing workflow budget, and return the checked object without another model rewrite.

Include the completed ELF/OpenViking/RAGFlow comparison and reuse assessment requested with this change. Most added lines are frozen test evidence, not product code. Preserve the original ELF result of 13/13 semantic answers and 12/13 with valid quotes; do not replace it with the correction result.

Alignment

This implements the source integrity and traceability requirements in docs/spec/agent_memory_knowledge_system_v1.md. The observed Cobalt quotation inserted an ellipsis; the native checker rejected it, but the host did not consistently enforce that result. No source, model, embedding dimension, store, or retrieval behavior changes. External MCP clients still need their own final-output hook; publishing instructions does not install one.

Validation

• cargo make test-benchmark-runner-contract: 222 tests passed, including last-turn rejection, incomplete verdicts, unavailable checks, and unchanged accepted answer snapshots.
• Live replay of the same 13 saved drafts against restored native originals: 12 passed directly; one was blocked and corrected with one same-model call. All 14 original objects remained unchanged. No new extraction, retrieval, or embedding calls.
• Correction cost: USD 0.000842050. Campaign settled total: USD 2.016143006; conservative exposure: USD 2.088323706 of USD 10. Test containers and volumes removed.
• cargo make check: passed (479 Rust tests, 222 runner-contract tests, 4 tooling tests; 111 environment-dependent Rust tests skipped locally). Documentation, formatting, Clippy, workflow lint, artifact manifests, and tested implementation hashes verified. PR integration, E2E, trace, and CodeQL checks must pass before merge.

The new replay verifies final-output enforcement. It is not a fresh competitive benchmark, semantic-verification guarantee, or customer acceptance.