fix(ai-memory): bound claude-mem sync runs and drop per-row table scans

docs/COMMANDS-GUIDE.md

@@ -72,7 +72,7 @@ Portable AI snapshots exclude skill roots, auth by default, caches, histories, d
 
 ## claude-mem peer sync
 
-`dot ai memory sync [push|pull|sync|sessions|status]` replicates the claude-mem store (`~/.claude-mem/claude-mem.db`) with a peer Mac over ssh; the peer needs only the dot binary, which serves `max`/`export`/`import`/`count` as JSON over stdio via the hidden `--serve` flag — no python or claude-mem tooling on either side. Incremental export uses the receiver's `MAX(created_at)` minus a 7-day overlap (`--full` disables it); dedupe keys are `(created_at, title)` for observations, `(created_at, request)` for summaries, and `(platform_source, content_session_id)`/`memory_session_id` for sessions. Inserts use the column intersection of both sides minus identity/sync columns, so plugin version skew skips a column (logged once) instead of failing; previously copied rows get a one-off attribution backfill (`agent_type`/`agent_id`/`metadata`/`content_hash`/`generated_by_model`). An import that adds rows kicks the local worker's `/api/admin/restart` when it runs. `dot ai memory install --peer <target>` installs the hourly LaunchAgent `com.dotfiles.claude-mem-sync` (ssh pre-check `ssh <target> true || exit 0`, so an away peer is a skipped run, exit 0); each scheduled pass refreshes claude-mem to the latest marketplace version first (also `dot ai memory update` and a `dot ai update` step), keeping both Macs on the same plugin version. Last run/result and both sides' counts live in `~/.claude-mem/sync-state.json` and the `dot ai memory status` "peer sync" section.
+`dot ai memory sync [push|pull|sync|sessions|status]` replicates the claude-mem store (`~/.claude-mem/claude-mem.db`) with a peer Mac over ssh; the peer needs only the dot binary, which serves `max`/`export`/`import`/`count` as JSON over stdio via the hidden `--serve` flag — no python or claude-mem tooling on either side. Incremental export uses the receiver's `MAX(created_at)` minus a 7-day overlap (`--full` disables it); dedupe keys are `(created_at, title)` for observations, `(created_at, request)` for summaries, and `(platform_source, content_session_id)`/`memory_session_id` for sessions. Inserts use the column intersection of both sides minus identity/sync columns, so plugin version skew skips a column (logged once) instead of failing; previously copied rows get a one-off attribution backfill (`agent_type`/`agent_id`/`metadata`/`content_hash`/`generated_by_model`). An import that adds rows kicks the local worker's `/api/admin/restart` when it runs. A run is bounded by `--timeout` (default 10m): past it, or on SIGTERM, the open import rolls back and the run exits non-zero naming the phase it stopped in; each phase logs its start and end with row counts. `dot ai memory install --peer <target>` installs the hourly LaunchAgent `com.dotfiles.claude-mem-sync` (ssh pre-check `ssh <target> true || exit 0`, so an away peer is a skipped run, exit 0; a watchdog sends SIGTERM after 15 min and SIGKILL 30 s later, since launchd starts no new run while one is alive); each scheduled pass refreshes claude-mem to the latest marketplace version first (also `dot ai memory update` and a `dot ai update` step), keeping both Macs on the same plugin version. Last run/result and both sides' counts live in `~/.claude-mem/sync-state.json` and the `dot ai memory status` "peer sync" section.
 
 ## Apps and profile snapshots
 

docs/commands/dot_ai_memory_sync.md

@@ -10,6 +10,10 @@ with a peer Mac. Incremental by default (receiver's MAX(created_at) minus a
 needs only the dot binary: the transport is
 'ssh <target> dot ai memory sync --serve <op>' with JSON over stdio.
 
+A run is bounded by --timeout: past it, or on SIGTERM, the open import
+rolls back and the run exits non-zero naming the phase it stopped in.
+Each phase logs its start and end to stderr.
+
 sessions replicates all sessions both ways. status prints both sides'
 counts plus the last scheduled run. With no action, sync runs both
 directions. --peer defaults to the configured dot peer target.
@@ -25,6 +29,7 @@ dot ai memory sync [push|pull|sync|sessions|status] [flags]
   -h, --help               help for sync
       --peer string        ssh target to sync with (default: the dot peer target)
       --remote-db string   DB path on the peer (default: the peer's ~/.claude-mem/claude-mem.db)
+      --timeout duration   abort the run (and roll back an open import) after this long (default 10m0s)
 ```
 
 ### Options inherited from parent commands