refactor: clean up legacy template fields and surface remaining knobs

[timzsu]

Purpose

PR-2 of RFC #28. Removes template fields the runtime ignores, surfaces remaining legacy knobs under names that match the underlying backend, and unifies YAML scalar quoting across examples/templates/. Also fixes a latent crash in the single-GPU PPO save wrapper that surfaced when validating the PPO config wiring end-to-end. PPO target_kl / early_stopping are deferred to PR-3 because they need a real KL-watching early-stop loop; templates keep the fields and the executor carries a TODO marker so the spec doesn't churn between PRs.

Changes

Template cleanup:

• examples/templates/*.yaml — drop resources.replicas, spec.sloSeconds, and spec.parallel blocks; none of those fields reaches the dispatcher or worker selection.
• examples/templates/ppo_*.yaml — drop generation.do_sample and training.optimize_cuda_cache; TRL's PPOTrainer.train hardcodes do_sample=True, and the cache flag has no step-based contract under the custom PPO loop.
• examples/templates/*.yaml — unify YAML scalars on bare form: drop unnecessary quotes, promote quoted ints (e.g. cpu: "8" → cpu: 8), keep quotes only where YAML grammar requires them (split: "train[:1%]", save_strategy: "no" so it doesn't parse as False, etc.).
• examples/templates/inference_vllm_all_args.yaml — move revision from model.vllm to model.source so all vLLM templates use the same canonical revision location.

Executor surfacing:

• src/worker/executors/agent_executor.py — honour spec.agent.timeout (default 600 s); requires a positive integer, rejecting floats outright so sub-second values like 0.5 don't silently truncate to 0. Threaded through both the single-task and batch streaming paths.
• src/worker/executors/ppo_executor.py + PPO templates — adopt the canonical training.report_to and training.project field names directly (matching PPOConfig); a small helper normalises null / empty values to TRL's "none" sentinel. training.padding_side was already wired through build_hf_load_kwargs; locked down with a test. A TODO(rfc28-pr3) next to the kl_coef read flags training.target_kl / training.early_stopping as intentionally unwired pending PR-3.
• src/worker/executors/vllm_executor.py — stamp revision from spec.model_revision (the canonical model.source.revision); drop the redundant model.vllm.revision lookup and the resolver helper, aligning vLLM with the other executors. tokenizer_revision stays under model.vllm since it's a distinct knob.

PPO save fix:

• src/worker/executors/ppo_executor.py — _wrapped_save_model unconditionally read ppo_trainer.deepspeed, but HF Trainer only sets that attribute when DeepSpeed is enabled. Single-GPU PPO therefore crashed on the first _save_checkpoint with AttributeError. The backup/restore now lives inside the existing is_deepspeed_enabled guard.

Tests:

• tests/worker/test_agent_timeout.py — agent timeout resolution (int-only contract).
• tests/worker/test_ppo_config_mapping.py — report_to / project / padding_side wiring.

Design

The fields handled here split into two groups:

• Dead. Schema accepts them, runtime ignores them. Removing from templates avoids advertising unsupported behaviour to users.
• Live but unwired (or under the wrong name). The backend (TRL PPOConfig, vLLM engine, AgentExecutor) supports the knob, but the worker never threaded the template value through, or the template used a non-canonical name. The fix is a small, local change; where the legacy template names diverged from the backend (log_with, tracker_project_name, vLLM-scoped revision), we align on the backend name rather than maintaining an alias.

target_kl / early_stopping are a separate case — TRL doesn't take them directly, and FlowMesh-owned KL early-stopping is a non-trivial change that warrants its own PR. The TODO marker keeps the connection between templates and executor explicit without dropping the fields.

Test Plan

• uv run pytest tests/server tests/shared tests/sdk tests/cli plus the focused unit tests for the executor wiring and the existing template-parse coverage in tests/server/task/test_template_validation.py.
• Live spot-check on the local stack: echo_local, inference_hf_tiny, inference_vllm_tiny, inference_vllm_all_args (after the revision relocation), lora_sft_llama, ppo_training_llama_1b end-to-end on a single GPU worker.

Test Result

• 474 passed (server + shared + sdk + cli) + unit tests on the touched executor paths.
• Live e2e on local stack (1 GPU worker):
  • echo_local — DONE.
  • inference_hf_tiny — DONE; transformers path unchanged.
  • inference_vllm_tiny — DONE; vLLM picked up the revision from model.source.revision.
  • inference_vllm_all_args — DONE; confirms model.source.revision is the only revision source after the helper was removed.
  • lora_sft_llama — DONE; training_successful=True, final_lora_archive produced (padding_side wiring exercised).
  • ppo_training_llama_1b — DONE; training_successful=True, runtime ~122 s, ran past save_steps=25/50, confirming both the PPOConfig wiring and the _save_checkpoint fix.
  • agent_simple_test — skipped; pre-existing requirement on the worker's secrets.yaml (UTU_LLM_API_KEY), unrelated to this PR.

---

Pre-submission Checklist

• I have read the contribution guidelines.
• I have run pre-commit run --all-files and fixed any issues.
• I have added or updated tests covering my changes (if applicable).
• I have verified that uv run pytest tests/ passes locally.
• If I changed shared schemas or proto definitions, I have checked downstream compatibility across Server and Worker.
• If I changed the SDK or CLI, I have verified the affected packages work (uv sync --all-packages --group ci --frozen).
• If this is a breaking change, I have prefixed the PR title with [BREAKING] and described migration steps above.
• I have updated documentation or config examples if user-facing behavior changed.

[kaiitunnz]

LGTM.