v1.0.0

Added

• Structured agent trace model with Trace, Span, ArtifactRecord,
  StateDelta, TraceSink, WithTraceSink, and DefaultTraceSink
• Trace linkage through Case.TraceID, Case.Trace, Result.TraceID, and
  JSONL trace_id result fields
• Agent metrics: TaskCompletion, ToolArgumentAccuracy, PlanAdherence,
  and StepEfficiency
• RAG metrics: ContextRecall, AnswerCorrectness, and NoiseSensitivity
• JSON scenario definitions via LoadScenarios, DecodeScenarios, and
  BindScenarioDrivers
• Static reports through compare.ReportHTML, compare.ReportMarkdown,
  compare.ReportJSON, and goeval report
• Judge calibration and pairwise summaries through compare.Calibrate,
  compare.CalibrateFile, and goeval calibrate

Changed

• Scenario steps can declare required and forbidden artifact keys directly
• Result and trace redaction now share the same WithRedactors hooks
• Case.TraceID now seeds empty structured trace IDs, and trace sinks write a
  shared trace ID once per Runner
• Repeated scenario summary rows now keep all run trace IDs under
  scenario_summary.trace_ids
• Tool-call metrics and scenario tool contracts now use trace tool-call spans
  when available, with Turns fallback for legacy cases
• Scenario trace assembly now avoids synthesizing turn-derived tool spans when
  a step result already provides trace tool-call spans
• Trace sink deduplication now marks a trace ID as written only after a
  successful sink write, allowing transient failures to retry
• Judge calibration now aggregates duplicate judge or variant rows instead of
  replacing earlier rows
• goeval report --out now rejects unknown file extensions unless --format
  is supplied explicitly
• Markdown reports now escape table cell pipes, newlines, and HTML-sensitive
  characters