Purge collected people, authors and task transcripts with their project

[egeozin]

What and why

Deleting a project keeps the project and run rows for the billing ledger, so foreign-key cascades from projects and workflow_runs never fire. purge_project_data didn't list these tables, so they outlived the project:

• connection_collection_jobs, connection_collection_pages and connection_collection_account_leases: the people LinkedIn collection gathered (names, headlines, up to 3,000 characters of page text each);
• project_authors: author bindings;
• project_task_entries: task transcripts. chat_messages were already purged; these weren't.

A collection interrupted mid-lease also stopped deletion from completing. The job still pointed at its extension device (lease_device_id, no ON DELETE). Deleting integration_connections cascades to the devices, so that foreign key failed and the purge transaction rolled back.

The purge now deletes those rows, with collection jobs going before the connections. tests/test_project_purge.py adds two tests:

• A schema guard. After the real migrations, every table holding project data must either be cleaned by deletion (named in the purge, or cascading from a purged table) or be listed in KEPT with a reason. A new table that forgets the purge fails CI.
• A purge test. It seeds an interrupted collection, an author and a task entry, then purges twice. Everything is gone, another project's author stays, and the run row stays.

Verification

• The new tests fail on main: the guard names the missing tables, and the purge test hits the connection_collection_jobs_lease_device_id_fkey violation. Both pass with the fix.
• ruff format --check, ruff check and lint-imports are clean.
• Ran serially against a disposable local Postgres 14: test_project_purge, test_project_deletion, test_project_deletion_api, test_x_project_purge, test_connection_collection, test_run_lifecycle_regressions. All passed. Running them in parallel locally exhausted Postgres shared memory, which is an environment issue.

Notes

• paid_ads_campaigns and paid_ads_proposals are listed as kept. They hold no personal data, and whether Tin keeps its record of the Google Ads campaigns it created is a separate decision.
• Projects deleted before this change still hold their collected people. Cleaning them up is a separate, confirmed data step.

🤖 Generated with Claude Code