feat: local dev observability — cockpit error boundaries + Sentry Spotlight with sentry_mode tool

[Foxfire1st]

Summary

Give agents-remember a code-level error/tracing layer for local development:

1. Cockpit error boundaries — the dashboard currently has no React error boundary; one render crash white-screens the whole cockpit (precedent: the EnclosureCanvas node.landing.find crash). Add a reusable CockpitErrorBoundary grammar primitive at the root and around viewport panels, with a cockpit-styled fallback + retry.
2. DSN-free Sentry Spotlight across all three runtime surfaces — MCP server (FastMCP instrumentation), FastAPI dashboard server, and the React cockpit send errors/traces to a local Spotlight sidecar. No DSN, no cloud; nothing leaves the machine.
3. A sentry_mode MCP tool gates the whole thing: enable/disable persisted sentry mode, default off. No env-flag gating.

Series (master + 3 sub-tasks)

• 01_cockpit-error-boundaries — standalone, no Sentry dependency; exposes an onError seam.
• 02_python-sentry-mode-and-spotlight-wiring — sentry-sdk in [dev] extras only; sentry_mode tool + persisted state; init seam called from MCP server + dashboard CLI entrypoints; Python-side docs + smoke.
• 03_dashboard-spotlight-and-smoke-docs — @sentry/react + @spotlightjs/spotlight as devDependencies behind DEV-gated dynamic imports; hooks the boundary seam; prod bundle proven unchanged; cross-layer smoke + dev workflow docs.

Invariants

• No production telemetry, ever: dev extras / devDependencies only, sentry_mode default off, no DSN anywhere in the repo, prod dashboard bundle byte-identical.
• Error boundaries never swallow silently (console + fallback always).
• stdio MCP transport untouched (SDK speaks HTTP to the local sidecar only).
• Sub-tasks never bump the version; the master owns any release decision at series end.