Problem

The kanban dispatcher's _tick_once_for_board() in gateway/run.py opens a SQLite connection via _kb.connect() on every 60-second dispatch tick, but never closes it. Over time (hours to days), this accumulates 50+ concurrent file descriptors to kanban.db, causing:

1. SQLite disk I/O error when too many connections contend for file locks
2. Eventual database corruption (database disk image is malformed)

Root Cause

# Before (leaks one FD per tick):
conn = _kb.connect(board=slug)
return _kb.dispatch_once(conn, ...)
# Finally block has conn.close() but exception paths skip it

The same function in kanban_db.py line 5475 already uses the correct pattern:

with contextlib.closing(connect()) as conn:
    res = dispatch_once(conn, ...)

Fix

Wrap the connection in contextlib.closing() so it is reliably closed on both normal return and exception paths. Also removes the now-redundant finally: conn.close() block.

Changes: 1 file, +16/-21 lines

Verification

Reproduced in production (Hermes gateway running 2+ hours):

• Before fix: 56 kanban.db FDs → dispatch errors every 60 seconds → DB corruption within hours
• After fix: 9 kanban.db FDs stable → zero dispatch errors after 12+ hours