feat: visual eval trend dashboard (agent-strace dashboard --trend)

[Siddhant-K-code]

What

Extends agent-strace dashboard with --trend mode: reads per-session eval scores from eval.json, computes error/retry/cost time-series, and renders a self-contained HTML report with inline SVG charts.

Closes #68

New flags

# Terminal trend summary
agent-strace dashboard --trend --since 30d

# Self-contained HTML report
agent-strace dashboard --trend --since 30d --html trend-report.html

# Add a timeline annotation (appears as a vertical marker on all charts)
agent-strace dashboard annotate --date 2026-05-10 --note "Added retry policy to AGENTS.md"

What the HTML report shows

Eval quality section (one sparkline per judge, requires agent-strace eval scores):

• Pass rate trend per judge with color-coded lines
• Annotation markers showing when config or model changed

Behavioral metrics section (four sparklines, no eval required):

• Error rate per session
• Retry rate per session
• Estimated cost per session
• Session duration

Recent sessions table with eval scores inline.

Design constraints met

• Zero CDN dependencies — all CSS inline, all charts inline SVG
• No JavaScript libraries — polylines rendered as SVG paths
• Self-contained: single HTML file, openable in any browser
• Works without eval scores (falls back to behavioral metrics only)

Files changed

• src/agent_trace/dashboard.py — TrendPoint, TrendReport, EvalScorePoint, build_trend_report(), format_trend_terminal(), render_html_trend(), save_annotation(), load_annotations(), _svg_sparkline()
• src/agent_trace/cli.py — --trend, --since, --html flags; dashboard annotate subcommand
• tests/test_dashboard_trend.py — 25 new tests

Test results

Ran 591 tests — OK (25 new)