fix(copilot): stop chat panel from wiping on tool invocation (ENT-517)

[dustywusty]

Summary

The deployed Copilot chat would silently wipe back to the welcome screen any time the model invoked a tool. Three independent bugs all produced the same symptom; this PR fixes the root cause + the worst secondary issues that surfaced once the chat stopped wiping.

Root cause — CopilotKitProvider thrashing

@copilotkit/react-core/v2's CopilotKitProvider declares headers = {}, properties = {}, agents__unsafe_dev_only = {}, selfManagedAgents = {} as default param values. JS evaluates those defaults on every render, producing fresh {} literals each time. Those refs invalidate the internal useMemos for mergedHeaders / mergedAgents. The deps of the giant setup useEffect (setRuntimeUrl + setRuntimeTransport + setHeaders + setProperties + setAgents__unsafe_dev_only) include those memos, so the effect re-fires on every render. Mid-tool-run, this reconnects the runtime and replaces the agent's threadId — the visible "chat reset to welcome" symptom.

Fix: pass module-level constant {} for those four props in CopilotProvider.tsx. Refs stay stable, the effect runs once at mount, no thread thrashing.

Secondary bugs surfaced after the chat stopped wiping

• Quick-action buttons (Search Contacts, Lead Triage) silently did nothing. CopilotKit clones agents per-thread (getOrCreateThreadClone); useAgent() outside <CopilotChat> returned the BASE agent while the chat rendered a CLONED agent for its own thread. Lifted threadId ownership up to CopilotOverlayPanel and Dashboard, pass it to useAgent({ threadId }), and forward it to <CopilotChat> so both ends share the same threaded clone.
• Tool throws caused CopilotKit's RunHandler to abort the run and spawn a fresh thread on the next message — same wipe symptom, different cause. Wrapped frontend-tool handlers in useAuditedFrontendTool to catch and return { error, ok: false } instead of re-throwing.
• res.json() exploded on HTTP errors in useGetTopLeads, useGetContactsByCompany, useSearchContacts (no res.ok check). Added guards.
• useCreateTask now validates contactId via dataProvider.getOne before creating, and sets sales_id from the logged-in identity (the dashboard's Upcoming Tasks widget filters by sales_id, so chat-created tasks were invisible).
• useSearchContacts now accepts firstName / lastName parameters (the server already supported them); previously the agent had no way to look up a contact by name.
• useGetTopLeads now pre-shapes its response to match LeadPriorityList's prop names so the agent doesn't have to translate snake_case → camelCase per row.
• Wrapped CopilotOverlayPanelGate in its own <ErrorBoundary> so a render error in the panel shows a visible fallback instead of silently unmounting.

Dev-quality changes

• Vite dev proxy now matches /api/* (was /api/copilotkit) and accepts a COPILOTKIT_PROXY_TARGET env var, so frontend devs can hit the deployed CopilotKit backend without running the local server.
• README updated with the new dev workflow.

Test plan

• npm run typecheck clean
• npm run lint clean (no new warnings on changed files)
• Click Lead Triage on the dashboard → user message appears in chat, getTopLeads runs, LeadPriorityList renders with names linking to contact pages, scores, and "Lifecycle · Last activity: …" subtitles
• "create a task to call Shelley Kerluke on 5/10/26 for a CopilotKit interview" → searchContacts finds her, createTask succeeds, task appears under Shelley's contact page AND on the dashboard's Upcoming Tasks widget
• Force a tool failure (e.g. unreachable backend) → chat shows the model's apology, threadId stays stable, panel does not wipe
• Verify on Render after deploy

[linear]

ENT-517