Load WAL checkpointer outside Zeitwerk reload; harden fork stop

Keep the module under lib/rails_ext, use before_worker_boot, shorten the
SQLite busy timeout under stop's join, and PASSIVE-checkpoint on exit for
short-lived writers.

Co-authored-by: Thomas Klemm <github@tklemm.eu>

config/initializers/sqlite_wal_checkpoint.rb

@@ -1,5 +1,8 @@
-# Start a checkpoint contender in every non-test process. Puma and Resque pool
-# stop before fork and start again in the child so the flock is never inherited.
+# Loaded from lib/rails_ext (autoload_lib ignore list) so reloads do not orphan
+# the contender thread. Non-test processes start here; Puma/Resque stop before
+# fork and start again in the child.
+require Rails.root.join("lib/rails_ext/sqlite_wal_checkpoint")
+
 Rails.application.config.after_initialize do
   SqliteWalCheckpoint.start unless Rails.env.test?
 end

config/puma.rb

@@ -54,7 +54,7 @@ Membership.disconnect_all
 # do not inherit the flock; each worker starts its own contender. Single-process
 # mode never forks, so the initializer's contender keeps running.
 before_fork { SqliteWalCheckpoint.stop }
-on_worker_boot { SqliteWalCheckpoint.start }
+before_worker_boot { SqliteWalCheckpoint.start }
 
 Signal.trap :SIGPROF do
   Thread.list.each do |t|