[AIGTWY-4880] Add native budget and model evidence for CUJ5

[andy-xu-db]

🥞 Stacked PR

• stack/andy/aigtwy-4880-account-auth [Files changed]
  • stack/andy/aigtwy-4880-budget-fixture [Files changed]
    • stack/andy/aigtwy-4880-budget-evidence [Files changed] ← this PR
      • stack/andy/aigtwy-4880-budget-defaults [Files changed]

---

Summary

Add independent evidence helpers for CUJ5's budget-driven smart-default journeys. They verify the selected native-agent process, completed task request model, displayed budget values, and recommendation.

Changes

• Compare the launch budget panel and ug usage with backend spend, limit, percentage, and meter values using their display rounding.
• Identify real Claude and Codex processes without treating launcher arguments as native processes.
• Correlate each completed file task with the request model recorded in its root native Claude or Codex session, requiring Codex's matching turn context to precede completion.

Configuration

The concrete CUJ5 target is declared by the workflow in PR #911. PR #913 forwards that metadata to pytest and keeps it out of native agent environments.

Verification

The complete stack passed 119 focused tests. Evidence cases cover terminal formatting, native process identification, root-session filtering, model spelling normalization, and rejection of missing, mismatched, or later Codex turn contexts. The seven-case live run verified completed-task model correlation for both native agents.

[andy-xu-db]

Consolidated into #911, which now contains the complete CUJ5 integration test change as one commit against main.