Keep redialing a restarted codespace and bring back all its loops

[scgopi]

Summary

Loops on a codespace that restarted no longer stay dead until a human selects one. A down codespace is still dialed once a minute after the 4-minute pause. When it answers again, every loop on it comes back, including finished loops and the workers inside a running composite. Plain ssh hosts are swept every 30 seconds, and their finished loops come back with no pane open.

Why

After a codespace restart, loops did not recover on their own. Someone had to create or select a loop to bring them back. Reading the restore path found three gaps:

Gap	Effect
CodespaceDialBreaker returned .paused forever after 240s of failed dials, and only the reconnect marker cleared it	A restart slower than about 4 minutes stopped every ensure and the 60s liveness sweep for good. Selecting a loop touches the marker, which is why creating a loop "fixed" it.
A finished loop was restored only after a pane redialed (RebootProbeGate)	With no pane open, finished loops never came back, on codespaces and plain ssh hosts alike.
ensureUnattendedSessionsAlive swept only graph.nodes	Workers of a piloted or armed composite live on its sub-graph and were never re-ensured after a reboot.

Changes

• CodespaceDialSchedule.slowRetryInterval (60s). Paused, the breaker lets through one dial per interval for the whole codespace, whichever read or ensure asks first. A failed slow dial leaves the outage clock alone. Enter or selecting a loop still reconnects at once.
• Recovery hook. When a codespace answers after an outage, the breaker touches the host's redial stamp (ZmxSessionLauncher.markRedialed). The next sweep then runs the reboot probe and restores finished loops, as if a pane had redialed.
• Pane. The paused banner now says it retries every 60s and redials on its own after that interval, on the same outage clock.
• Sweep. It also ensures the unresolved unattended children of a piloted or armed composite, the same children pilotComposite starts.
• Plain ssh hosts. The liveness sweep runs every 30s (was 60s). Codespaces are swept on every second tick, so they stay at 60s. RebootProbeGate always lets a plain ssh host through, so its finished loops are probed on every sweep. Those dials are multiplexed over the ControlMaster and spend no quota.

Cost while a codespace stays down past 4 minutes: about one gh dial per minute from the daemon, plus one per open pane of that codespace. A plain ssh host with finished loops costs one multiplexed probe dial every 30s. A dial to a stopped codespace starts it, so a codespace with running loops is brought back up rather than left stopped.

Tests

RED: xcodebuild test -only-testing CodespaceDialScheduleTests CodespaceSelectionTests RemoteSessionResumeTests RemoteRebootRestoreTests with SSHReconnectLoop.swift and GraphStore.swift from origin/main and the breaker's slow retry and recovery hook removed -> 4 failed of 52 (aPausedCodespaceIsStillDialedOncePerSlowInterval, aCodespaceThatAnswersAfterAnOutageAsksForTheRebootProbe, aPausedPaneRedialsOnItsOwnAfterTheSlowInterval, theLivenessSweepRestartsTheLoopsInsideARunningComposite), TEST FAILED exit 65
GREEN: same focused xcodebuild test on this branch -> 52 tests in 4 suites passed, exit 0
GREEN: after the interval change, the same focused suites -> 54 tests in 4 suites passed, exit 0; its new tests (aPlainSSHHostIsProbedOnEverySweepWithNoPaneOpen, plainSSHHostsAreSweptEveryThirtySecondsAndCodespacesEveryMinute) fail to compile on 91a930c -> ProjectRegistry has no member sweeps
REGRESSION: full xcodebuild test of the graphcode scheme -> 2226 tests in 255 suites, 1 failure (MessageDeliveryTests.aSessionWhoseTaskEndedIsNeverTypedInto), which fails identically 3 of 3 runs on base c1febe4 without this change. graphcoded and graphcode-cli schemes build, make check exit 0.

The MessageDeliveryTests failure is pre-existing on main and unrelated: it is a live-zmx send test, and this PR does not touch the send path.