fix(kanban): wrap ALTER TABLE ADD COLUMN with race-safe helper

[baocin]

The kanban DB migration in _migrate_add_optional_columns reads a single
PRAGMA table_info snapshot at entry, then issues multiple ALTER TABLE
ADD COLUMN statements guarded by that snapshot. When two gateway
processes restart concurrently (e.g. --replace takeover + cron job
tick), the PRAGMA can be stale by the time a given ALTER TABLE
executes, causing:

sqlite3.OperationalError: duplicate column name: <name>

This has been hit in production on the consecutive_failures column.
The same race will fire for any new column added to the migration in
the future.

Add _safe_add_column() which tries the ALTER TABLE and silently
ignores duplicate column name — another process already added it.

Changes:

• New helper: _safe_add_column(conn, table, col_def) with
  try/except for OperationalError, only re-raising non-duplicate errors.
• All 12 ALTER TABLE ADD COLUMN calls in _migrate_add_optional_columns
  now go through _safe_add_column.
• The task_events.run_id ALTER TABLE also uses the helper.
• Updated the NOTE docstring to reflect the new guard strategy.

[alt-glitch]

Duplicate of #21529 (same fix: wrap ALTER TABLE ADD COLUMN in try/except for duplicate column error). Also competes with #21021. Fixes #21503 and #21374.

[teknium1]

Closing as redundant — the same race-safe migration helper landed on main via PR #22627 (commit 7869838, by @wesleysimplicio) on May 9. The current _add_column_if_missing() does exactly what your _safe_add_column() does: tries the ALTER, catches OperationalError, swallows only "duplicate column name", re-raises everything else.

We also followed it up with PR #22994 which removed the redundant double-init pattern in the gateway watchers (connect() followed by init_db()), which was the upstream cause of the cross-process race firing in the first place. Together those two fixes should fully address the production incident you cited.

Thanks for the clean writeup @baocin — the diagnosis was spot-on.

References:

• PR fix(kanban): make _migrate_add_optional_columns idempotent on concurrent open #22627 — _add_column_if_missing() race-safe helper
• PR fix(kanban): drop redundant init_db() in gateway watchers (salvage #21874) #22994 — drop redundant init_db() in gateway watchers (sqlite3.OperationalError: database is locked at first kanban dispatcher tick after every gateway startup #21378)