[babaliauskas]

What this changes

A follow-up to #65 from looking at a real converter report.

• The target's trace box is red only when its tool calls differ from the source's. A card flagged for its wording alone (semantic similarity) showed both sides making the same convert_currency call, with the target's in a red box. That read as a tool failure when there wasn't one. "Differ" means the pair's tool diff is not empty: same rule as the card's "Tool diffs" list, so the box and the list agree. Applies to imported agent traces too (their trace_diff).
• A target's "(no tool calls)" is red only when the source did call tools. When neither side called a tool, it was red anyway.
• A within-threshold finding sits in a neutral notice. The plain notice style uses the warning tint, so a "Within threshold" card still looked like a warning.

Template and CSS only (report.html.j2, report.css); no data or report.json change.

Test plan

• New tests:

  • the target trace is red only when its calls differ (same call vs a different call);
  • "(no tool calls)" isn't flagged as missing when neither side called a tool, and is when the target dropped one;
  • a within-threshold finding uses notice-neutral and regressed cards don't.

  Each failed before the change.

• make ci: ruff, ruff format, mypy --strict, and pytest (3396 passed, 95.3% coverage)

• pre-commit run --all-files

• Re-rendered the real converter run and checked headless-Chromium screenshots: a regressed card with matching calls and an opened within-threshold card are both neutral where nothing differs

🤖 Generated with Claude Code