feat(viewer): render run traces from a payload with a JS renderer

[Polarisyjr]

Summary

This is Step 1 of the BenchFlow trace viewer work. More features will be added in follow-up PRs; at this stage, the main goal is to establish the rendering architecture and bring the viewer's overall style and trace presentation closer to PostTrainBench.

Refactors bench eval view around a shared payload-based renderer:

rollout artifacts → viewer.py → versioned payload → render.js → HTML

Instead of generating trace HTML directly in Python, Python now handles artifact normalization while JavaScript handles presentation. This also creates a shared renderer that can later be reused by a hosted/static viewer.

What changed

• Add a versioned viewer payload generated from rollout artifacts.
• Treat trajectory/acp_trajectory.jsonl as the canonical trace source, with legacy turn*.txt as a fallback.
• Move trace HTML rendering into render.js.
• Separate browser/DOM setup into boot.js, keeping the renderer independently testable.
• Rework the viewer layout and styling to roughly align with the PostTrainBench trace viewer.
• Vendor the PostTrainBench stylesheet as the base visual layer, with BenchFlow-specific overrides isolated in benchflow.css.
• Add BenchFlow-specific run information including:
  • reward and run status
  • token usage
  • environment / agent / verifier timing
  • skill mode and skill invocations
  • agent/verifier errors and partial trajectories
• Improve tool-call rendering:
  • preserve complete ACP content blocks
  • render diffs separately
  • handle binary output per block
  • truncate very large outputs while preserving the beginning and end
  • support ANSI formatting and a limited escaped Markdown subset
• Bundle viewer assets locally so generated pages remain portable and work without network access.

Compatibility

This does not introduce a new on-disk trace format.

The viewer payload is derived at render time and is not persisted. Existing ACP trajectory artifacts remain unchanged, and legacy Claude Code turn*.txt traces continue to work as a fallback.

When both formats exist, ACP is used as the source of truth rather than merging or rendering both.

Tests

Added two layers of viewer tests:

• tests/test_trajectory_viewer.py

  • payload structure
  • run status
  • token metadata
  • turn grouping
  • tool-output completeness and truncation
  • payload/script escaping

• tests/test_trajectory_viewer_render.py

  • runs render.js through Node
  • tool output and diff rendering
  • Markdown and ANSI rendering
  • binary placeholders
  • injection protection
  • reward 0 vs. unscored
  • schema-version mismatch

Also updates the trajectory integration tests for the new payload-based rendering path.

Validation

uv run pytest \
  tests/test_trajectory_viewer.py \
  tests/test_trajectory_viewer_render.py \
  tests/test_trajectory_integration.py -q

Manual E2E:

bench eval view <rollout-dir>

Next steps

This PR intentionally focuses on the basic single-run viewer and visual alignment with PostTrainBench.

Follow-up work will add more viewer features such as job-level navigation, verifier details, review/audit results, cost and per-call metrics, system metrics, and additional trace metadata.

[1digitaldesign]

Cool cool!