feat: CRM import pipeline for people and orgs

[khanal-samir]

What

A staged ETL pipeline to load people and organizations into the CRM from:

• CSV upload — no OAuth, universal fallback
• Inbound webhook (API-key auth) — the generic entry point for Zapier / Make / n8n / scripts
• Gmail & Google Calendar — mines sent-mail recipients and meeting attendees
• Calendly — booking invitees with custom question answers
• Google Sheets — re-syncable, same path as CSV
• PostHog — product users via the Persons API
• Outlook / Microsoft 365 — contacts, sent mail, and calendar through one Graph grant

Base is feat/sequence (not main): the import code reuses the sequence infra — token-crypto, encrypted-token storage — which only exists on that branch. Targeting feat/sequence keeps this PR to the import diff alone.

Design

Follows the existing sequence patterns; no new infrastructure.

• Extract is separated from load. A connector stages raw records into import_records and stops. The user reviews mapping + match preview, then a loader writes to the CRM. Fixing a mapping re-runs the load without re-fetching the source.
• Idempotent re-sync via external_identities, keyed on (workspace, provider, external_id). Sources with no stable id (CSV rows, anonymous pushes) fall back to a content fingerprint, so a re-upload updates instead of duplicating — this closes the nullable-email uniqueness gap for imported records.
• DB-polled worker mirroring sequence-worker (worker:imports). Records load in per-batch transactions — a bad row is flagged and skipped, not fatal to the run. 429s reschedule from the persisted cursor; auth failures flag the connection reconnect_required.
• Fixture mode behind INTEGRATION_LIVE_FETCH_ENABLED (mirrors SEQUENCE_LIVE_SEND_ENABLED): off, connectors resolve from fixtures, so the full pipeline runs end to end with no credentials.
• Status seeding by signal tier (payment → customer, conversation → qualified, interest → lead). Applies on create only — an import never downgrades an existing record.

Schema

Migration 0004: integration_connections, import_jobs, import_records, external_identities, integration_api_keys; people_source gains import. gmail_integrations is left untouched.

Testing

• pnpm test — 127 pass
• pnpm test:coverage — 100% on the DB-free boundary (statements/branches/functions/lines)
• pnpm check-types and pnpm lint — clean
• Verified end-to-end against a seeded DB (CSV import, re-sync idempotency, webhook push, automated-sender rejection). That live run is what surfaced the emailless-row duplication bug now fixed by the fingerprint fallback.

Covered: CSV parsing, mapping + auto-detection, normalization + match resolution, conflict policy, fingerprint idempotency, every connector (pagination / 429 / auth failure / dedupe), route wiring, and the worker.

Follow-ups (not in this PR)

• Web UI (settings/integrations + import wizard)
• Live OAuth connect flows per provider (env wiring pending)
• env.example documents the new vars; real values to be added separately

🤖 Generated with Claude Code