feat: eval pipeline run metadata and comparison

[rajnavakoti]

Description

Add run identity, timestamps, and comparison capability to the eval pipeline and viewer. Currently eval results are flat JSON arrays with no run metadata — users can't tell when an eval was run, what changed between runs, or track coverage improvement over time.

Acceptance Criteria

• Eval pipeline writes run metadata (run_id, timestamp, domain, source_mix, entity_count_at_time) to eval output
• Viewer loads multiple eval result files and presents them as named runs
• Run comparison mode shows CLEAN/INCOMPLETE/MISSING deltas between two runs
• Coverage trend visualization (sparkline or small chart) when 3+ runs exist
• Run selector dropdown in Evals page header

Out of Scope

• Multi-domain support (separate issue)
• Retrieval quality vs KB quality distinction (separate issue)
• Automated scheduled eval runs