docs: fix critical doc-code mismatches across config, endpoints, LoRA, and metrics

[guozhihao-224]

Summary

Fix high-priority documentation issues that can directly cause user errors (runtime ValueError, 404s, or missing files), identified by automated audit in #1165.

• Proxy endpoints (agent_workflow.md): Fix paths from /v1/chat/completions and /v1/responses to /chat/completions and /responses. The OpenAI SDK puts /v1/ in base_url, not in the request path. (Finding nccl error, machine all-reduce communication test passed. #3)
• Rollout config (rollout_workflow.md): Remove non-existent rollout.group_size YAML config — group_size is a runtime API parameter on submit(), not a config field. Fix trajectory dump path to include /logs/{user}/ segment. (Findings Update release v0.1.1 #4, Update release v0.1.2 #5)
• LoRA support (cli_reference.md, lora.md, cli_args.py): Update "Only support FSDP" to include Megatron (requires megatron-bridge). Add note that SGLang LoRA requires weight_update_mode: disk. (Findings V0.2.0 prerelease #8, V0.2.0 prerelease #9)
• Metrics tracking (metrics_tracking.md): Fix SwanLab mode from "online" (invalid, causes ValueError) to "cloud". Add Trackio as a supported logging backend. (Findings Can I use 1*GPU for training by setting 'allocation_made'? #14, Update README.md citation #15)

Verification

Each fix was verified against the actual code:

• Proxy endpoints: confirmed via server.py constants and OpenAI SDK URL construction behavior
• group_size: confirmed InferenceEngineConfig has no such field; all examples use actor.adv_norm.group_size
• Trajectory path: confirmed via StatsLogger.get_log_path() at stats_logger.py:179
• LoRA: confirmed MegatronEngine supports LoRA via megatron-bridge at megatron_engine.py:284
• SwanLab: confirmed valid modes are cloud/local/offline/disabled at cli_args.py:2078
• Trackio: confirmed full integration in stats_logger.py:97-107

Test plan

• pre-commit run --all-files passes
• ./docs/build_all.sh builds without errors
• CLI reference auto-generation reflects updated use_lora help string

Ref: #1165

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the documentation and CLI help text to include support for LoRA with the Megatron backend, requiring the megatron-bridge package, and specifies that SGLang LoRA requires the disk weight update mode. Additionally, it introduces trackio as a new metrics tracking backend, updates API endpoint paths in the agent workflow documentation by removing the /v1 prefix, and modifies the trajectory storage path to include user-specific directories. I have no feedback to provide.