fix(db): add FK constraints to convergence_trajectories and worktrees

[polishfreak]

Fixes #61

Adds migration 008 that recreates convergence_trajectories with ON DELETE CASCADE and worktrees with ON DELETE SET NULL, both referencing tasks(id). Since SQLite doesn't support ALTER TABLE for FK constraints, the tables are recreated and data is preserved via INSERT...SELECT (orphaned rows are dropped during the copy).

Also cleans up existing orphaned agent_instances by NULL-ing out current_task_id references to nonexistent tasks.

SET NULL is used for worktrees because on-disk cleanup should be handled explicitly rather than auto-deleting the record silently.

submitted by Polish's bot :)

[odgrim]

Review: FK constraints migration — needs test fixes

Issue

The migration itself is well-written — the FK constraint additions, orphan cleanup, and SET NULL strategy for worktrees are all correct. However, the PR introduces 18 test failures:

• All 15 trajectory_repository tests fail with FOREIGN KEY constraint failed
• All 3 worktree_repository tests fail with the same

Root cause

The test setup for these repositories inserts rows with task_id values that don't exist in the tasks table. With the new FK constraints enforced, SQLite rejects these inserts.

Fix needed

Each test that inserts into convergence_trajectories or worktrees must first insert a corresponding task row into the tasks table. For example, in the trajectory tests, before calling repo.save(&trajectory).await, ensure a task with id = trajectory.task_id exists.

A helper function like insert_test_task(pool, task_id) would keep the fix clean.

Everything else looks good

• Migration structure (CREATE new → INSERT...SELECT with JOIN → DROP old → RENAME) is the correct SQLite pattern
• Indexes are properly recreated
• SET NULL for worktrees is the right choice (disk cleanup should be explicit)
• CASCADE for convergence_trajectories is correct (no external resources)
• agent_instances orphan cleanup is a nice bonus