fix: derive active count from activeTaskIds.size, constrain ps scan

Address two additional review concerns:

1. Remove separate 'active' counter from LaneState; derive it from
   activeTaskIds.size instead. This makes negative-underflow impossible
   — the Set is the single source of truth for active task count.
   Previously, a double-reset scenario could drive 'active' negative,
   violating the concurrency check in pump().

2. Replace unbounded 'ps -axo pid=,command=' with targeted pgrep
   pre-filter in orphan scanner. Only fetches full command info for
   candidate PIDs matching 'codex|claude', avoiding O(all-processes)
   overhead on large hosts.

scripts/recover-orphaned-processes.sh

@@ -59,8 +59,21 @@ function resolveCwd(pid) {
   return match ? match[1] : "unknown";
 }
 
-const lines = run("ps -axo pid=,command=").split("\n");
-const includePatterns = [/\bcodex\b/i, /\bclaude\b/i, /claude\s+code/i];
+// Pre-filter candidate PIDs using pgrep to avoid scanning all processes.
+// Falls back to full ps scan if pgrep is unavailable.
+const candidatePids = run("pgrep -f 'codex|claude' 2>/dev/null || true")
+  .split("\n")
+  .map((s) => s.trim())
+  .filter((s) => s.length > 0 && /^\d+$/.test(s));
+
+let lines;
+if (candidatePids.length > 0) {
+  // Fetch command info only for candidate PIDs.
+  lines = run(`ps -o pid=,command= -p ${candidatePids.join(",")}`).split("\n");
+} else {
+  lines = [];
+}
+
 const excludePatterns = [
   /openclaw-gateway/i,
   /signal-cli/i,
@@ -85,9 +98,6 @@ for (const rawLine of lines) {
   if (!Number.isInteger(pid) || pid <= 0 || pid === process.pid) {
     continue;
   }
-  if (!includePatterns.some((pattern) => pattern.test(cmd))) {
-    continue;
-  }
   if (excludePatterns.some((pattern) => pattern.test(cmd))) {
     continue;
   }