chore: drop runtime guards that can no longer fire

[CjhHa1]

Summary

Follow-up to #560: remove runtime guards that can no longer fire under the current code and dependency pins. Pure deletion (1 line changed, 105 removed).

• vLLM direct rollout pipe — protocol_version. engine.py and runtime.py each hard-coded _PROTOCOL_VERSION = 1 and checked the other side's value on every request/response. The runtime is a spawn child started from the same checkout and is never respawned once the connection breaks, so the two copies always agree; the check can only ever fail when someone bumps one copy and forgets the other — the same shape as the vLLM version guard fixed in fix(vllm): align direct rollout manifest guard with runtime pin #560. Removed the constant, the field, both checks, and its entry in the driver's reserved-payload set. request_id monotonicity and command validation stay; they catch stale or mismatched responses.
• Qwen3-Omni — transformers 5.6.0 s_aux shim. _transformers_compat.py only activated when transformers is exactly 5.6.0, whose flash_attention_forward called s_aux.to(...) unconditionally. The repo now pins transformers>=5.12,<5.13 (build(sglang): migrate runtime to 0.5.19 #471), and 5.12.1 already guards s_aux.to(...) with if s_aux is not None. Deleted the module and its call in bundle.py.
• Qwen3.5 — duplicate transformers floor. bundle.py::_model_class_for repeated the transformers>=5.0 check that validation.py::_MIN_TRANSFORMERS owns; both are below the 5.12 pin. Dropped the bundle copy.

Found while auditing for #560-style drift. The one guard that needs hardware to decide — Cosmos3's H20 cuBLAS workaround, gated to torch 2.10/cu128 while the repo pins torch 2.13/cu130 — is filed as #567 rather than changed here.

Related Issue

N/A (follow-up to #560; related: #567)

Test Plan

• PYTHONPATH=. python3 /tmp/verify_vllm_pipe_protocol.py (kept outside the repo): sends a driver-shaped request through a multiprocessing.Pipe into runtime._recv_request, replies with runtime._send_response, and reads it with VLLMRolloutEngine._recv — round-trip succeeds without protocol_version; non-monotonic request_id, empty command, and a stale response are still rejected.
• Compared transformers wheels: 5.6.0 integrations/flash_attention.py:84 calls s_aux.to(query.dtype) unconditionally; 5.12.1 :85-87 guards it with if s_aux is not None.
• python3 -c "import unirl.models.qwen3_omni.bundle, unirl.models.qwen3_5.bundle" — ok.
• SKIP=no-commit-to-branch pre-commit run --from-ref upstream/main --to-ref HEAD — pass; git diff --check upstream/main — pass.
• GPU checks on H20 (gz), exact PR head 7003606, image qwen38next-runtime-base:ubuntu24-cu130-torch213-v2 (torch 2.13.0+cu130, transformers 5.12.1):
  • Qwen3.5-2B via Qwen3_5Bundle.from_config, eager and meta-init: both resolve Qwen3_5ForConditionalGeneration; greedy reply to "12*7" is 84.
  • Qwen3-Omni-30B-A3B-Instruct thinker via Qwen3OmniBundle.from_config with flash_attention_2: loads with 0 missing keys (only talker.*/code2wav.* unexpected, as for a thinker-only load); forward makes 48 flash-attention calls, all with s_aux=None, logits finite; greedy reply starts with 84.
  • examples/ar/qwen3_moe_grpo_30b_a3b_gsm8k_fsdp_vllm_tp4 on Qwen3-30B-A3B-Base, vLLM 0.28.0 TP4 + FSDP, 2 rollouts with one native IPC weight sync between them: exit 0, reward 0.56 → 0.69. The sync woke the weights, reloaded them layer-wise on all 4 TP workers, and committed without error before rollout 2 generated; every driver/runtime command (startup, generate, sleep/wake, weight update, shutdown) went over the pipe without protocol_version.

Compatibility / Risk

No config, checkpoint, or public API change. The driver/runtime pipe is internal to one process tree, so dropping protocol_version has no cross-version compatibility surface. Running Qwen3-Omni on transformers 5.6.0 (outside the declared pin) would hit the upstream s_aux=None bug again.

Reviewer Notes

AI-assisted change.

• Kept: request_id/command pipe validation; unirl/reward/remote.py protocol_version (separately deployed HTTP service) and VLLM_NATIVE_PROTOCOL_VERSION (single definition) are different protocols and untouched.
• Open PRs touching the same files (feat(inf-train parity): add Qwen3-MoE train-inference parity plugin, part 2/2 — training integration #505, feat(experimental): add bitwise Qwen3-MoE train–inference parity case based on GRPO + vLLM(TP>1) + FSDP #430, feat(qwen3-omni): add Talker direct-TTS training #355) don't modify the deleted lines.
• Independent fresh-context review by gpt-5.6-sol-high and by grok-4.7-high-fast: no findings.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.