Evaluation: Include judge cost tooltip

Is your feature request related to a problem?
The cost tooltip on /evaluations only lists Response generation and omits the judge cost, leading to discrepancies in the cost breakdown compared to the total displayed. This causes confusion for users trying to understand the complete cost.

Describe the solution you'd like

• Update EvalCost to include judge?: EvalCostEntry.
• Render a Judge scoring entry in the tooltip of EvalRunCard when job.cost.judge is present.
• Ensure all cost entries (response, judge, embedding) are shown in the tooltip to match total_cost_usd.

Additional Context

Backend response

        {
            "id": 887,
            "run_name": "assistant_v2_v1_0_ai_cohort_2_evals_demo_goldenqna_1788945623629",
            "dataset_name": "ai_cohort_2_evals_demo_goldenqna",
            "config_id": "dc576d3c-5e86-4eef-9b95-6d1e2194cce4",
            "config_version": 1,
            "dataset_id": 709,
            "batch_job_id": 1805,
            "embedding_batch_job_id": null,
            "status": "completed",
            "run_mode": "fast",
            "object_store_url": null,
            "score_trace_url": "s3://ai-platform-documents-staging/3ce7b9fe-2900-4f33-9a68-8162568a41be/evaluations/score/887/traces_887.json",
            "total_items": 9,
            "score": {
                "overall": {
                    "verdict": "Needs Refinement",
                    "breakdown": [
                        {
                            "key": "ground_truth",
                            "name": "Adherence to Ground Truth",
                            "delta": -0.45,
                            "score": 3.44,
                            "weight": 0.71,
                            "verdict": "Needs Refinement"
                        },
                        {
                            "key": "prompt",
                            "name": "Adherence to Prompt",
                            "delta": 1.11,
                            "score": 5,
                            "weight": 0.29,
                            "verdict": "Good"
                        }
                    ],
                    "ai_summary": "**Overall read:** The run is in generally good shape — most questions score 4–5 on ground truth and a clean 5 on prompt adherence, with no KB in play. The model answers are consistently substantive and well-structured; the main tension is between the model giving richer, modern-science answers and golden answers that expect specific, textbook-narrow responses.\n\n**Top 3 to check:**\n\n**Question 9** — Ground-truth score of 0: the golden answer expects a very specific socio-demographic list (sex, skin colour, caste, mother tongue, etc.) but the model answered from a biological/population-genetics frame; this looks like a golden-dataset framing issue more than a model failure, but needs a human call on which answer the use case actually wants.\n\n**Question 7** — Borderline ground-truth score (2): the model explicitly refuses to classify by skin colour and race, directly conflicting with the golden answer that includes skin colour; this is a values/alignment tension between the model's safety behaviour and the expected answer — worth deciding whether the golden answer or the model's stance is appropriate for this context.\n\n**Question 4** — Minor: the macrophage-as-viral-factory stage (a key step in the reference answer) is omitted; solid overall but worth a quick check if curriculum accuracy to the specific textbook is required, pointing at the model.\n\nThese are go-verify pointers — open each item, read the actual answer against the use case requirements, and decide based on what the deployment needs.",
                    "overall_score": 3.89
                },
                "summary_scores": [
                    {
                        "avg": 3.44,
                        "std": 1.42,
                        "name": "Adherence to Ground Truth",
                        "data_type": "NUMERIC",
                        "total_pairs": 9
                    },
                    {
                        "avg": 5,
                        "std": 0,
                        "name": "Adherence to Prompt",
                        "data_type": "NUMERIC",
                        "total_pairs": 9
                    }
                ]
            },
            "unscoreable": null,
            "is_score_updated": true,
            "is_judge_run": true,
            "cost": {
                "judge": {
                    "model": "gpt-5.6-luna",
                    "cost_usd": 0.002832,
                    "input_tokens": 16061,
                    "total_tokens": 18104,
                    "output_tokens": 2043
                },
                "response": {
                    "model": "gpt-5.6-luna",
                    "cost_usd": 0.001942,
                    "input_tokens": 276,
                    "total_tokens": 3466,
                    "output_tokens": 3190
                },
                "total_cost_usd": 0.004774
            },
            "error_message": null,
            "organization_id": 1,
            "project_id": 1,
            "inserted_at": "2026-09-09T09:20:25.182635",
            "updated_at": "2026-09-09T09:24:23.885105"
        },

Commencez par la page /evaluations et suivez le type EvalCost ainsi que le rendu de l’infobulle de EvalRunCard. Ajoutez l’entrée du coût du juge à côté des coûts de réponse et d’embedding lorsque job.cost.judge existe, puis vérifiez que l’infobulle affiche tous les coûts disponibles et correspond à total_cost_usd.