feat(evalboard): show conversation transcript for simulation tasks

[CarlesUIPath]

Summary

Simulation runs already write a clean user<->agent transcript to conversation.log (orchestrator._log_conversation), but the evalboard never surfaced it. Add a Conversation section to the task detail page that renders the dialog as chat bubbles, below the Success criteria section. Gated on parsed turns, so single-shot tasks (which never write conversation.log) are unaffected.

Changes

• lib/runs.ts: readConversationLog (mirror of readLogTail) + a conversation parser returning ConversationTurn[].
• _sections.tsx: ConversationSection renders turns as bubbles (user right/gray, agent left/blue).
• page.tsx: fetch + parse, render below Success criteria. It only renders the conversation if "conversation.log" exists in the runs/ directory.

Tests

Tests mirror repo conventions: unit tests for parseConversation (alongside parseMessages) and a render test for ConversationSection (alongside message-timeline). readConversationLog is left untested, matching its untested twin readLogTail.

Ui Example

When running a simulation task, a "conversation.log" text is generated. This is rendered in the run page after the Success criteria" as follows.

[github-actions]

Claude finished @CarlesUIPath's task in 1m 33s —— View job

---

Code Review in Progress

Todo List:

• Read .github/code_review.md for review standards
• Review CLAUDE.md for project conventions
• Get full diff with git diff origin/main...HEAD
• Read and analyze each changed file in full context
• Perform cross-file consistency checks
• Analyze what might be missing
• Format and post the review

[bai-uipath]

Approve. Clean, well-scoped feature — parser, presentation, and page wiring live in the right seams, and it's verified against the actual writer. All the design calls here are the ones I'd make.

• Parser matches the writer exactly. parseConversation targets the format _log_conversation emits (orchestrator.py:1843) — header shape, the — metadata separator (both U+2014), and the header\n{body}\n\n framing all line up. The always-present opener metadata (pinned initial_prompt / simulator-generated opener) parses correctly.
• The no-simulation default is right, and better than described. Gating on parsed-turns (conversation.length > 0) rather than file existence means single-shot tasks render nothing, and empty-file / no-header-garbage / read-error cases all collapse to [] and hide cleanly — no crash, no empty section.
• Dialogue-only rendering is the correct split. Tool-use and thinking are stripped on the Python side before the log is written, so bubbles show the clean final utterance; the per-turn caption surfaces N tool calls + cost, and the full trajectory already lives in the message-timeline / CostExplorerSection below. Not redundant with the timeline.
• Tests mirror repo conventions. Parser unit tests beside parseMessages, render test beside message-timeline; two-turn / empty / metadata / multi-line all covered. Appropriately scoped.