Maintain Connector Summary Read Model

1. Projection Model

• Refactor connector-summary projection into durable evidence extraction plus read-time synthesis.
• Add SQLite and Postgres connector-summary evidence tables and migrations. (connector_summary_evidence in server/db.js + server/postgres-storage.js; durable evidence only — no synthesized columns.)
• Add dirty-marking and reconcile helpers mirroring the retained-size read-model pattern. (markConnectorSummaryEvidenceDirty, markAllConnectorSummaryEvidenceDirty, reconcileDirtyConnectorSummaryEvidence in server/connector-summary-read-model.js.)
• Add a full rebuild/repair path for connector-summary evidence. (rebuildConnectorSummaryEvidence derives from connector_instances + maintained retained_size_stream; no connector re-run, no credentials.)
• Persist retained stream and byte evidence needed by overview rows. (stream_records_json, retained_bytes_json, and total_retained_bytes are derived from the maintained retained-size projection in both rebuild and dirty reconcile; no freshness, verdict, health, or UI copy is stored.)

2. Write Hooks

• Mark connector-summary evidence dirty from existing owner mutation seams. Every owner mutation route that calls invalidateConnectorSummariesCache now also awaits a scoped markConnectorSummaryEvidenceDirty({ connectorInstanceId }) right after it, with the known connection id (no markAll fallback is needed — every seam knows its connection): the bearer owner-agent routes revoke (prior slice), reactivate, run, schedule pause/resume + delete, delete, and PATCH /v1/owner/connections/:id rename (server/routes/owner-connection-{reactivate,run,schedule,delete}.ts, owner-connections.ts), plus all cookie-authed /_ref mutations sharing one refConnectorsContext — set-display-name, connector/connection run, schedule upsert/pause/resume/delete, revoke, delete, reactivate (server/routes/ref-connectors.ts). Each route declares an injected optional markConnectorSummaryEvidenceDirty? in its context interface; the marker is wired through server/index.js (7 context objects).
• Mark connector-summary evidence dirty from record ingest hooks that already update retained-size evidence. Scoped hooks colocated with the retained-size delta and scoped to the known connector_instance_id now cover ingestRecord (prior slice), deleteRecord (Postgres + SQLite arms, only on a changed delete), deleteAllRecords (Postgres + SQLite arms, only when rows were cleared), and deleteAllRecordsForConnector (per-instance in the post-commit loop) in server/records.js. No-op deletes (missing record / empty stream) do not dirty evidence, pinned by tests.
• Mark connector-summary evidence dirty from run lifecycle and gap/backlog changes. PARTIAL: the run-start seam is covered — the owner run-now routes (bearer owner-connection-run.ts and cookie /_ref/connectors|connections/:id/run) dirty the connection's evidence when a run begins (counted under the owner-mutation task above), because that is the only run-lifecycle seam that already had a known connection id at a summary-invalidation point. RESIDUAL: there is no existing controller-level run finish/fail or gap/backlog-drain seam that invalidates connector summaries or marks retained-size dirty today, so hooking those would mean introducing a NEW seam in the controller/run pipeline rather than mirroring an existing one. Left for a follow-up slice that adds those controller hooks deliberately (run finish/fail evidence and gap-drain completion), since record-ingest dirtying already covers the count/stream changes a run produces.

3. Read Paths

• Make unscoped /_ref/connectors read maintained evidence plus read-time synthesis.
• Preserve exact scoped connection/detail diagnostics with deep run evidence.
• Remove or gate the short TTL cache once the maintained read model owns the hot path.

4. Validation

• Add SQLite tests for dirty marking, lazy reconcile, and time-relative synthesis. Storage evidence tests exist in test/connector-summary-read-model.test.js (rebuild, dirty→stale, lazy reconcile of only-dirty rows, drop-on-delete, synthesis-free-columns guard). Write-hook seam tests now exist in test/connector-summary-dirty-hooks.test.js: record ingest dirties the matching connection and only that connection; no-op re-ingest does not dirty; deleteRecord and deleteAllRecords dirty the matching/cleared connection and only that one, while no-op deletes (missing record / empty stream) do not dirty; the owner revoke route AND a non-revoke owner mutation (the PATCH /v1/owner/connections/:id rename route) each dirty the targeted connection end-to-end. Time-relative synthesis remains deferred until task 1.1/3.x wires the read path.
• Add Postgres parity tests for the storage fixtures. (Same file, gated on PDPP_TEST_POSTGRES_URL; same rebuild/dirty/reconcile shape as SQLite.)
• Add a query-count or dependency-injection test proving the full-list path avoids per-connection evidence fan-out.
• Run reference and console type checks.
• Run browser perf harness on dashboard, records list, exact source detail, and runs.
• Deploy in a declared live-stack window and verify /_ref/connectors latency plus owner-journey acceptance.