feat(sglang_engine): allow PD worker_type on /add_worker registration path

[DavidBellamy]

Problem

SGLangEngine._init_normal asserts worker_type == "regular" before POSTing to the old-style /add_worker?url=... endpoint:

if parse(sglang_router.__version__) <= parse("0.2.1") or self.args.use_miles_router:
    assert (
        self.worker_type == "regular"
    ), "pd disaggregation is not supported in old router or miles router."
    response = requests.post(
        f"http://{self.router_ip}:{self.router_port}/add_worker?url=http://{self.server_host}:{self.server_port}"
    )

Any attempt to stand up prefill/decode workers via the miles-router path (or a gateway that mirrors it, e.g. sgl-model-gateway) fails fast at engine init, even when the receiving router could handle PD if given the information.

This blocks PD disaggregation throughput scaling in Miles-driven rollouts. Observed during LLM360/RL360#76 overnight runs when attempting FAST_ITER_CONFIG=pd-disagg through the gateway's Miles-API shim.

Fix

Relax the assertion. Forward worker_type (and bootstrap_port for prefill workers) as extra query params on the /add_worker URL. Routers that honor them get correct PD registration; routers that only accept the single-arg form ignore the extras and register as regular, with a logger.warning so the fallback behavior is visible to operators.

add_worker_url = (
    f"http://{self.router_ip}:{self.router_port}/add_worker"
    f"?url=http://{self.server_host}:{self.server_port}"
)
if self.worker_type != "regular":
    add_worker_url += f"&worker_type={self.worker_type}"
    if self.worker_type == "prefill":
        bootstrap_port = server_args_dict.get("disaggregation_bootstrap_port")
        if bootstrap_port is not None:
            add_worker_url += f"&bootstrap_port={bootstrap_port}"
    logger.warning("Registering a '%s' worker via /add_worker on the old-style router path...")
response = requests.post(add_worker_url)

Companion server-side change

This PR alone doesn't make PD routing functional through an old-style router that didn't previously understand worker_type; it just removes the fail-fast assert. Operators also need one of:

• sgl-model-gateway: extend the /add_worker handler to accept worker_type= and bootstrap_port= query params (small serde change; the current shim ignores unknown params)
• Newer sglang_router (>0.2.1): switch --no-use-miles-router so Miles takes the POST /workers JSON path that already supports PD

Test

Unit-checked URL construction for all four shapes:

build_url(..., 'regular')               == '.../add_worker?url=http://h:p'
build_url(..., 'decode')                == '.../add_worker?url=http://h:p&worker_type=decode'
build_url(..., 'prefill', 8998)         == '.../add_worker?url=http://h:p&worker_type=prefill&bootstrap_port=8998'
build_url(..., 'prefill', None)         == '.../add_worker?url=http://h:p&worker_type=prefill'

Context

LLM360/RL360#76 Track G (job 1559336) proved that full PD+Mooncake KV transfer works on M2's SR-IOV IB via SGLang's native sglang_router.launch_router --mini-lb. This PR unblocks the same flow through Miles-driven rollouts, so agentic RL can use PD scaling without hand-written sbatch test harnesses.

[DavidBellamy]

Re-opening against radixark/miles:main so the LLM360/miles deploy branch auto-builder picks it up. Same branch, same diff.