feat(cli): show tool call summaries with token estimates in watch recent steps

[zawakin]

Summary

Claude Code's adaptive thinking means thinking text is often absent from logs, making the reasoning-first Recent Steps display in agtrace watch useless (it also leaked chain-of-thought when present). This PR makes Recent Steps tool-centric and adds intuitive token visibility per step.

Changes

• Priority flip in step previews: tool executions > message > reasoning. Reasoning-only steps show a Thinking... placeholder instead of the thinking text.
• Rich tool summaries: tool name + salient argument (shortened file path for Read/Edit/Write, command for Bash, pattern for Grep/Glob, server.tool for MCP), execution duration ([1.2s]), and (+N more) for multi-tool steps. Existing TodoWrite/AskUserQuestion/Task special cases are preserved.
• Token visibility:
  • →~3.4k — estimated tokens fed back into context by tool results (~chars/4), making the previously blackbox tool-result cost visible.
  • Δ+5.2k — context window growth vs the previous step (input token delta), color-coded: gray < 2k, yellow ≥ 2k, red ≥ 10k, cyan for negative (compaction).
• Error highlighting: steps with failed tools render the description in red.
• Deduplicated the two step-line render paths in turn_history.rs into a shared step_preview_line helper.
• Re-exported MessageBlock/ReasoningBlock/ToolCallBlock/ToolResultBlock from agtrace-sdk (CLI depends on sdk only).

Tests

10 new unit tests covering priority order, thinking-text suppression, result token estimation, duration formatting, multi-tool counts, delta chaining across steps (including in-progress steps without usage), negative delta after compaction, and error flagging.

mise run verify passes.

🤖 Generated with Claude Code