feat(node-summary): AI summaries with disk cache and LLM-gated invalidation

[SobshDev]

Summary

• Clicking a code node streams an AI-generated overview into the existing summary sheet (uses the configured default Claude model).
• Persists results to .graphy/summaries.json keyed by node id; second click on an unedited node is instant with no network call.
• Fingerprint = SHA-256 of source + sorted caller/callee ids. Mismatch triggers a Haiku validator that judges whether the cached prose is still accurate — only regenerates when the validator says no, so whitespace and trivial refactors keep the cache.
• Header / facts / callers / callees render live from the graph the whole time. A RefreshCw button forces a regenerate. No-key state points at settings.

Test plan

• bun run dev, configure a Claude key, click a code node → header is instant; overview streams in; .graphy/summaries.json is created with an entry for that node id.
• Re-click the same node → overview appears instantly, no Anthropic request.
• Double-click → save a tiny whitespace-only edit → re-click for summary: brief "Checking cached summary…", then the same cached prose is reused.
• Edit the function body meaningfully → re-click: cached prose is invalidated, fresh summary streams in.
• Remove the Claude key → click a node: settings CTA shows, no network calls.
• Click the Refresh icon → regenerates even when the cache would have been valid.