agents: rolling restart + session auto-resume across Pod restarts

[jorisjonkers-dev-agents]

What changed

Five improvements to runner lifecycle, idle policy, and session continuity:

1. Rolling restart endpoint

POST /api/v1/admin/runners/rolling-restart gracefully scales down every
active runner Pod (READY/STARTING/FAILED → IDLE, PVC preserved). On next
session start each workspace re-provisions, pulling :latest and picking
up any image or config changes without manual per-workspace intervention.

Response: { "cycled": N, "workspaceIds": [...] }

Node drain workflow:

# Checkpoint all runners before the drain
curl -X POST -H "X-User-Id: $UID" https://assistant.<domain>/api/v1/admin/runners/rolling-restart
kubectl drain enschede-gtx-960m-1 --ignore-daemonsets --delete-emptydir-data
# <maintenance>
kubectl uncordon enschede-gtx-960m-1
# Workspaces auto-reprovision on next access, sessions resume automatically

2. Claude session auto-resume

Previously every spawn() generated a fresh --session-id <uuid>, so a
re-provisioned Pod started a blank terminal even though conversation history
sat on the shared credentials PVC. The gateway now accepts resumeCliSessionId
in the spawn request and launches Claude with --resume <id> when provided.
StartAgentSessionCommandHandler looks up the latest non-failed Claude
session's cliSessionId for the workspace and forwards it automatically.

3. Codex session auto-resume

Codex session files live in ~/.codex/sessions/ (codex-credentials PVC)
and survive Pod restarts. On re-provision codex resume --last is invoked
which CWD-filters to the most recent session for that workspace — no stored
session ID required.

4. Idle policy: never evict connected users or active AI sessions

IdleScaleDownScheduler now enforces three rules before scaling down:

• Connected client: ConnectedClientTracker counts open WebSocket bridges
  per workspace. A workspace with any attached browser client is never scaled.
• AI output → activity touch: UpstreamHandler (gateway frames → browser)
  now calls activity.touch() on every terminal frame. The 30-min idle timer
  resets as long as Claude/Codex is producing output, even after the user
  disconnects.
• Running sessions use 4-h grace period: when any RUNNING agent session
  exists the idle threshold switches from 30 min to
  agent-runtime.agent-idle-after-seconds (default 4 h). This protects long
  autonomous tasks that outlast the user's browser session.

Combined policy:

client connected                → never scale down
AI outputting (client watching) → never scale down
AI outputting (client gone)     → 4 h grace from last output
no AI, no client, 30 min quiet  → scale to zero

5. CI fixes

• Removed unused Instant import (lint)
• AdminRunnerController annotated @Hidden (excluded from springdoc spec — not part of the frontend contract)
• StartAgentSessionCommandHandlerTest sessions mock changed to relaxed = true
• SessionAttachHandlerTest updated to pass new ConnectedClientTracker param

Files changed

Area	Files
agent-gateway	Dtos.kt, AgentSessionManager.kt, AgentController.kt
assistant-api domain	AgentGatewayClient.kt
assistant-api application	StartAgentSessionCommandHandler.kt, RunnerMaintenanceService.kt (new), IdleScaleDownScheduler.kt, ConnectedClientTracker.kt (new)
assistant-api web	AdminRunnerController.kt (new, @Hidden), SessionAttachHandler.kt
Tests	AgentSessionManagerTest.kt (+2), IdleScaleDownSchedulerTest.kt (+3), RunnerMaintenanceServiceTest.kt (new, 5 tests), StartAgentSessionCommandHandlerTest.kt, SessionAttachHandlerTest.kt