[AIGNLAI]

Summary

• Remove BaseRouter.cal_rci and all RCI calls, logs, and diagnostic-only allocation tracking from shared, Carrot, EA-RAM, and reward-model evaluation paths.
• Remove RCI documentation and benchmark protocol references; retain nAUC, Peak Score, QNC, allocation rules, and the existing Pareto JSON schema.
• Update smoke checks to validate finite Peak Score instead of RCI; preserve EA-RAM null-action coverage with a real-auction regression test.

Validation

• Rebased onto latest main after PR Add xRouteBench, MMR-Bench V2 and R2-Bench; fix RouteLLM pair selection #16 merged; this PR contains only RCI-removal changes.
• Full unittest discovery: 67 tests passed, including the 30-method x 3-benchmark offline matrix and 24 compatible multimodal method checks.
• Persistent offline smoke: 90/90 passed.
• Bounded real-data smoke: 90/90 passed across XRouteBench, MMRBenchV2, and R2Bench.
• All 180 persistent smoke results have identical Pareto curves and nAUC to the corresponding pre-removal results; new logs contain no RCI output.
• Syntax compilation and git diff --check passed.

Smoke runs exercise real router training, prediction, allocation, evaluation, and JSON output. They use small samples, deterministic hash embeddings, tiny local transformer checkpoints, and a stub at the paid NIRT annotation boundary; RouteFM uses existing released local weights. These are runtime/regression checks, not full benchmark quality claims.

Compatibility

The deprecated cal_rci API is intentionally removed. Existing routing curves and result JSON formats are unchanged.