Release v0.2.0 — Signal + UX overhaul · grippado/claude-atlas

Signal

Severity classification (high/medium/low) on every issue, surfaced in the sidebar, graph edges, and stats.
Domain stopwords drop noise from trigger collisions. Generic tokens like agent, user, task, proactively no longer cause false positives — this was the main noise source in v0.1.x.
Short tokens (< 5 chars) excluded from collision detection. A 72-agent real-world scan went from ~100 issues to ~17 meaningful ones.

Search box filters nodes, issues, and orphans live.
Severity filters alongside kind filters.
Grouped issues by severity → kind with accordion; HIGH expanded by default.
Bidirectional click: clicking an issue in the sidebar highlights the pair on the graph and zoom-fits them. Clicking a node still populates the detail pane.
Hover tooltips on nodes show a quick description preview.
Isolated artifacts get their own tab and a dashed border on the graph — for CLAUDE.md files in repos without a .claude/, or skills with no overlap.
Suggested fixes (💡 text hints) on every issue. No automation — you still decide.

Edge dataclass gains a severity field. External consumers of ScanResult.to_dict() will see a new severity key.

Full changelog: v0.1.1...v0.2.0