fix: release watched_repos on offboard so re-hire isn't blocked

[michaelzwang13]

Closes #31.

What was broken

Offboarding (`DELETE /agents/{id}`) calls `Orchestrator.stop_agent`, which only sets `status='stopped'` and clears the container + token — the DB row stays, so the `on delete cascade` on `watched_repos` never fires.

Combined with #28's cross-agent uniqueness (`unique(user_id, owner, repo)`), a fired employee's stale `watched_repos` row permanently occupied the `(user, owner, repo)` slot. Re-hiring under the same user and re-subscribing to the same repo returned a 409 against a row tied to a stopped agent.

Why #28 is still correct

Two different users can still watch the same repo — the constraint is per `user_id`, not global. So a senior engineer and another senior engineer (different accounts) can both auto-review a junior's PR. Only intra-user duplicates are blocked. The fix is purely about the stale-row-after-stop case.

Fix

`Orchestrator.stop_agent` now calls `WatchedRepoModel.delete_by_agent(agent_id)` before the AgentModel.update that flips the status to `stopped`. Three observations:

• `agent_memory`, `agent_action_log`, `reviewed_prs` are deliberately untouched. They're the audit trail of what the agent actually did. Offboard means "stop working," not "erase ever existed." Recommended option from the issue.
• No new migration. The cleanup is at the model layer; existing schema unchanged.
• Fire-and-forget. `delete_by_agent` returns nothing — we don't care about the row count.

Test

One new test in `TestStopAgent` (`test_stop_agent_releases_watched_repos`) patches `WatchedRepoModel.delete_by_agent` and asserts it was called with the right agent_id. Full suite: 142 → 143 passing.

Test plan

• `pytest` passes (143/143)
• Manual: hire CRE A, subscribe to repo X, offboard A, hire CRE B, subscribe B to repo X — should now 201 instead of 409

🤖 Generated with Claude Code

[michaelzwang13]

LGTM ✅

Review Summary:

This is a clean, well-documented fix for the offboarding issue (#31).

What I like:

• Correct root cause fix: Releasing watched_repos before the agent status update allows the cascade to work properly
• Audit trail preserved: Keeping agent_memory, agent_action_log, and reviewed_prs is the right call for compliance
• Good test coverage: The new test verifies delete_by_agent is called with the correct agent_id
• Clear documentation: Comments reference #31 and #28, explaining the "why"
• Minimal/surgical change: Only 51 lines added, no schema migration needed

The fix correctly balances cleanup (releasing repo subscriptions) with audit preservation (keeping action logs). Code looks ready to merge — would APPROVE if GitHub allowed self-approvals.