[schemas] Connector sync state

[alanshurafa]

What this adds

schemas/connector-sync-state/ — one generic table that any capture connector can use to remember where it left off and whether its last run failed.

Every connector that pulls from an external source on a schedule (email importer, calendar poller, RSS reader, webhook receiver, backfill job) needs the same three things: a resume cursor so the next run doesn't re-read everything, lifecycle timestamps, and a last-error plus error count so a stalled or failing connector can be spotted. Writing that table once per connector gets repetitive and inconsistent. This gives them one shared table and three RPCs that wrap the begin / success / error transitions.

What's in it

• public.connector_sync_state — one row per (connector, surface, sync_key). Holds cursor_value, high_watermark, the four lifecycle timestamps, last_error, error_count, and free-form counters / metadata jsonb.
• connector_sync_begin(...) — upserts the row and marks the run started.
• connector_sync_success(...) — advances the cursor, stamps success, resets the error count, merges counters. A NULL cursor leaves the existing one in place.
• connector_sync_error(...) — stamps the error and increments the count, but leaves the cursor untouched so the next run retries from the last known-good position instead of skipping the failed window.

Notes for reviewers

• No thought ids. This table is keyed by connector identity, not by any thought. The surfaced primary key is a BIGINT id; the canonical public.thoughts.id (a UUID) never appears here, and the schema is standalone — it doesn't touch public.thoughts at all.
• Operational state, service-role only. RLS is on with no policy (deny by default), all three functions are SECURITY INVOKER, and table/function privileges are granted to service_role only. The table also explicitly REVOKEs from PUBLIC, anon, authenticated, so a Supabase project that blanket-grants new tables to its API roles still can't expose this one.
• Idempotent and additive — CREATE TABLE IF NOT EXISTS, CREATE INDEX IF NOT EXISTS, CREATE OR REPLACE FUNCTION; no DROP/TRUNCATE/unqualified DELETE. Safe to re-run.
• Pairs with brain-health-monitoring. The README shows the two queries (erroring connectors, stalled connectors) you'd wrap as a health view to alert on.

Testing

Applied to a throwaway Postgres 18 and exercised end to end: begin → success advances the cursor and keeps error_count at 0; begin → error increments the count and preserves the cursor; a later clean run resets the count; a NULL-cursor success preserves the prior cursor; distinct surfaces keep independent rows. Verified the status CHECK, RLS-on, SECURITY INVOKER, and that grants resolve to service_role only — including under a blanket ALTER DEFAULT PRIVILEGES ... GRANT ALL ON TABLES TO anon, authenticated, where the explicit revoke keeps anon/authenticated off the table. Re-applied the migration to confirm idempotency. Markdownlint clean against the repo config.