drainAll: record that the timeout is a whole-drain budget, not a per-session grace

The javadoc said 'for each session that is BUSY, poll up to timeoutNanos',
which reads as a per-session grace period. The deadline is taken once, before
the loop, so the first BUSY session can spend all of it. That is deliberate and
is the safer of the two designs: the drain is one phase of a shutdown sequence
that must finish inside launchd's exit window, and a per-session grace would
overrun it and get the daemon SIGKILLed part-way through, leaving the sessions
not yet reached with no clean release, no preserved-worktree log and no
snapshot. Found by a read-only hunt that read the code correctly and drew the
opposite conclusion from the wording.

fleetd/src/main/java/dev/ltms/fleet/session/SessionManager.java

@@ -864,10 +864,20 @@ public final class SessionManager implements TurnListener {
     }
 
     /**
-     * Gracefully drain all registered sessions on daemon shutdown. For each session that is
-     * {@code BUSY}, poll up to {@code timeoutNanos} for it to leave {@code BUSY}, then release it
-     * regardless. Non-busy sessions are released immediately. A failure releasing one session is
-     * logged and does not abort the rest.
+     * Gracefully drain all registered sessions on daemon shutdown. Non-busy sessions are released
+     * immediately; a {@code BUSY} one is polled until it leaves {@code BUSY}, then released
+     * regardless. A failure releasing one session is logged and does not abort the rest.
+     *
+     * <p>{@code timeoutNanos} is a budget for the WHOLE drain, not a grace period per session: the
+     * deadline is taken once, before the loop. So the first BUSY session can spend all of it, and a
+     * later BUSY one is then released with no wait at all. That is deliberate. This drain is only
+     * one phase of shutdown — {@code Fleetd} closes the message service, the push loop, the
+     * heartbeat, MCP and the router after it — and the whole sequence has to finish inside
+     * launchd's exit window. A per-session grace would let N busy members drain for N * the
+     * timeout, overrun that window, and get the daemon SIGKILLed part-way through; the members not
+     * yet reached would then get no clean release, no preserved-worktree log, and no snapshot.
+     * Cutting one turn short is the cheaper failure, and it is not silent: an abandoned BUSY
+     * session is preserved, snapshotted, and logged at WARN by {@code logPreservedForShutdown}.
      *
      * <p>CB-544: this is a {@link ReleaseCause#SHUTDOWN} release — the worker's pane is stopped
      * (the process must end) but its worktree is preserved and its path logged. Shutdown is never