Covers two scenarios on a single Kind cluster:
- Single-repo: deploy create copies commands.py into hooks/, deployment
  start runs it, mutating the stack-source working tree to v2 + deployment
  restart re-copies and re-executes v2.
- Multi-repo: stack with two pod repos produces hooks/commands_0.py +
  commands_1.py, deployment start invokes both pod start() hooks.

The test stages stack files into a temp git clone (bare + working) so
restart's git pull has a real upstream. busybox pods keep the harness
trivial. Phase 2 uses kubectl wait directly because deployment ps's
substring filter (deploy_k8s.py:1366) doesn't list multi-pod stacks.

Also tightens the _copy_hooks docstring to spell out that only
call_stack_deploy_start loads from the copied location.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Rename the stack and script accordingly:

- test-restart-hook stack → test-restart
- test-restart-hook-multi stack → test-restart-multi
- run-restart-hook-test.sh → run-restart-test.sh
- start-hook-marker file → marker
- pod-repo dirs test-restart-hook-pod-{a,b} → test-restart-pod-{a,b}

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…y test

dns_probe.py used 'list[str]' (PEP 585 generic alias for builtins),
which only parses on Python 3.9+. CI runs Python 3.8, so any caller
of 'deployment restart' (which lazy-imports dns_probe) crashed at
module load with 'TypeError: type object is not subscriptable'. Use
'List[str]' from typing to keep 3.8 compatibility, matching the rest
of the file's imports.

run-deploy-test.sh previously ended with --skip-cluster-management,
leaving the Kind cluster running for the next CI step to inherit.
Switch the final stop to --perform-cluster-management so subsequent
tests (e.g. run-restart-test.sh) start from a clean host, and replace
the now-trivial namespace assertion with a real check that the kind
cluster is actually gone.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>