v2.0.1 — Bug Fixes

Patch release fixing the two issues that broke the scheduler after deployment.

What changed since v2.0.0

Startup migration crash fixed. o9p8q7r6s5t4 no longer fails with
"column ... already exists" when the database was originally created via
create_all. The migration is now idempotent — it only adds created_at /
updated_at to keywords and pipeline_lock when the columns are missing.
Scheduler greenlet_spawn has not been called error fixed.
- The dispatcher success path refreshes the source ORM instance after the
  pipeline's commit before reading it.
- The failure path now passes pre-captured source_id / source_name
  scalars into _mark_unrecoverable instead of touching the expired ORM
  instance after session.rollback() — accessing any attribute (even the
  primary key) of an expired object raises MissingGreenlet under async
  SQLAlchemy 2.0.51, so sources that 401/403/404 (e.g. a dead RSS feed like
  the TechTarget redirect) were crashing the loop instead of being disabled.
- Breaking feeds are now properly disabled instead of looping errors every ~60s.

372 passed, 15 skipped.
New regression test: tests/test_dispatcher.py::test_dispatch_tick_disables_unrecoverable_source
guards that the unrecoverable path passes scalars (not the ORM instance).
ruff check and ruff format --check pass.

Full Changelog: v2.0.0...v2.0.1