feat(eval): add observability — SDK sends results to server, UI shows them

[vishesh-orkes]

Eval Observability — Phase 1 (Issue #215)

When running CorrectnessEval from the Python SDK, there was no way to see results beyond the terminal. This PR closes that gap: eval results are automatically persisted and visible in a new Experiments section of the UI, and eval runs are kept separate from production agent traffic.

---

What's new

Python SDK

• CorrectnessEval.run() automatically sends the full suite result to the server after each run. Failures don't surface to the caller — eval logic is never blocked by network issues.
• Eval results now carry richer metadata: run name, strategy, ran_by, per-case prompt and output, and optional score / reasoning for LLM-judge assertions.
• runtime.push_dataset(name, cases) stores a named, reusable test dataset on the server.
• Cases can be filtered by tags when calling run().

Server

• New eval_runs, eval_cases, eval_checks, and eval_datasets tables (SQLite and Postgres).
• REST endpoints: POST /api/eval/runs, GET /api/eval/runs, GET /api/eval/runs/{id}, POST /api/eval/datasets, GET /api/eval/datasets, GET /api/eval/datasets/{name}.
• Eval runs are excluded from Agent Executions search results by default. An includeEvalRuns flag re-includes them when needed.

UI — Experiments (new sidebar section)

• Eval Runs: table with pass rate bars, pass/fail case counts, search, and agent/result filters. Click any row to see the full run detail — cases, per-check results, agent output, and semantic score with LLM reasoning where applicable.
• Datasets: split-panel view listing all pushed datasets on the left; click to inspect individual cases on the right.
• Agent Executions: a new Type column distinguishes production runs from eval runs. A "Show eval runs" toggle (next to "Hide sub-agent executions") reveals them when needed.

Tests

• Python unit tests covering serialization, session tagging, result posting, tags filtering, and metadata fields — no LLM calls.
• Spring Boot integration tests for all eval storage operations.
• Playwright E2E tests for the Eval Runs and Datasets pages.

---

Out of scope (Phase 2)

• Triggering an eval run directly from the UI
• Creating or editing dataset cases in the UI
• Model comparison / benchmarks view

---