Maintain Connector Summary Read Model

Why

GET /_ref/connectors is a core owner-console read path. Recent performance tranches made it fast by adding a short in-process cache and by making overview rows shallow, but that is not the final construction: the read path still has a large per-connection evidence fan-out underneath the cache, and record-ingest writes do not event-invalidate the summary projection.

The reference already has the right pattern in the retained-size read model: maintained, dual-backend, incrementally updated evidence with dirty reconciliation and honest rebuild state. Connector summaries should use the same construction instead of a parallel TTL cache.

What Changes

• Add a maintained connector-summary evidence read model for durable per-connection facts.
• Keep time-relative health, freshness, verdict, and next-action synthesis on read, using current now and runtime/controller liveness.
• Mark connector-summary evidence dirty from owner mutations, run lifecycle changes, and record ingest hooks.
• Replace the full-list fan-out cache with one indexed evidence read plus pure synthesis.
• Preserve scoped detail reads and diagnostics without falling back to shallow overview evidence.

Capabilities

Modified:

• reference-connector-instances

Impact

• Affects reference storage, migrations, SQLite/Postgres parity, and /_ref/connectors internals.
• Does not change the owner API response shape except for optional freshness/read-model metadata if needed for diagnostics.
• Expected benefit: event-fresh summary counts/evidence with stable low-latency reads and no time-stale verdict cache.