feat(rm): add --rm-url and fix small CLI drift in the docs

[Rockdu]

What

• Add --rm-url to add_reward_model_arguments, mirroring the miles-core flag of the same name (placed before --custom-rm-path, same as there). No built-in --rm-type consumes it yet — it is the standard place for --custom-rm-path functions to read a remote service endpoint, which is exactly how the customization guide's api_rm example and the rewards page's remote-RM note already use it.
• Small doc accuracy fixes: the std-normalization switch is --disable-grpo-std-normalization (on by default; --grpo-std-normalization is not a flag), --diffusion-num-steps defaults to 10, and the LoRA sync log line ends with (unmapped=N).

Why

Cross-checked every flag in the six pages against main's arguments.py (AST) and the FSDPArgs dataclass after the recent args refactor (radixark#132/radixark#138/radixark#141/radixark#142/radixark#148). --rm-url was the one flag the pages reference that did not exist — rather than rewriting the (useful) remote-RM examples, this adds the flag the docs assume, keeping the CLI mergeable with miles core. Everything else in the PR already matches post-refactor main.

Files

• miles/utils/arguments.py — new --rm-url argument in the reward group
• docs/user-guide/customization.md — std-normalization wording
• docs/advanced/sde-backend.md — --diffusion-num-steps default
• docs/advanced/lora.md — log-line quote

Checklist

• pre-commit run --all-files — not run locally
• Tests — none added; the new argument has no runtime consumer to test
• pytest -x — not run locally
• Launch flags changed: the argument block is syntax-checked (ast.parse); --help parse not run locally (needs the sglang env)
• New public flag is documented — rewards.md and customization.md in this same PR describe it
• No example added

🤖 Generated with Claude Code

https://claude.ai/code/session_01RXZMxrjDB2V88kSJEGa2cj