fix(examples): stop launchers when gateway readiness fails

[Bluu (Bluuok)]

Seven example launchers continue to start controllers/trainers after all 60 gateway health probes fail, and each curl request can wait indefinitely. This hides the startup failure and can launch training against an unavailable gateway.

Give each probe a one-second timeout, track readiness explicitly, and exit with a diagnostic before launching the controller/trainer when the retry loop is exhausted. Cover Calc-X local/Minikube, GSM8K, multimodal QA, ScienceWorld, Search-R1 and LLM-in-Sandbox in one focused change.

Validation in the frozen Linux CPU environment:

python -m pytest tests/examples/test_launcher_readiness.py -q -p no:cacheprovider
python -m ruff check tests/examples/test_launcher_readiness.py
python -m ruff format --check tests/examples/test_launcher_readiness.py
python -m pre_commit run --files examples/calc_x/run_local.sh examples/calc_x/run_minikube.sh examples/gsm8k/run_local.sh examples/multimodal_qa/run_local.sh examples/science_world/run_local.sh examples/search_r1/run.sh examples/llm-in-sandbox/run.sh tests/examples/test_launcher_readiness.py
git diff --check

All 15 readiness tests passed: seven failure paths, seven delayed-success paths with intact arguments containing spaces, and one first-attempt success. The original six failure regressions failed on unchanged main; both added multimodal cases exposed missing gating/timeout in the previous branch version. Scoped Pyright and Bash syntax checks on repository-LF content passed. Earlier complete example-directory coverage passed 45 CPU tests before the multimodal addition; final combined verification is tracked separately.

The tests isolate all service, Ray, Minikube, cleanup, HTTP and training commands through a fake PATH. No real services, process termination, cluster, network or GPU training was exercised. Fork CI awaits maintainer approval. AI assistance was used; all seven launcher changes and regression coverage were reviewed.

[Copilot]

Copilot review overview

🟡 Changes recommended

The multimodal QA launcher retains the same readiness bug and is omitted from regression coverage.

Review effort: Balanced
Findings: 1

Open (1)

• Add readiness gating to multimodal QA local launcher · New

What changed in this PR

Adds bounded gateway readiness checks so example launchers fail before starting controllers or trainers.

Changes:

• Adds one-second curl timeouts and explicit readiness gates.
• Adds parameterized regression tests for six launchers.
• Leaves the equivalent multimodal QA launcher unchanged.

File	Description
examples/​calc_x/​run_local.sh	Gates local startup on gateway readiness.
examples/​calc_x/​run_minikube.sh	Gates Minikube startup on gateway readiness.
examples/​gsm8k/​run_local.sh	Adds bounded readiness failure handling.
examples/​science_world/​run_local.sh	Prevents startup after failed probes.
examples/​search_r1/​run.sh	Prevents startup after failed probes.
examples/​llm-in-sandbox/​run.sh	Adds bounded readiness failure handling.
tests/​examples/​test_launcher_readiness.py	Tests failure, delayed success, and argument forwarding.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.