feat(lora): fix LoRA weight-sync deadlock and add stable LoRA RL recipe

[WWWjiahui]

Summary

Makes LoRA RL train end-to-end. Two independent fixes: a weight-sync deadlock in the RaaS engine, and a training recipe that keeps LoRA from collapsing. Validated on Qwen3-1.7B / math with a clean rising eval curve.

1. Weight-sync

On every training step the freshly trained LoRA adapter is pushed to the SGLang inference server. The swap follows a fixed sequence so that no request ever runs against half-updated weights or stale cache:

1. Pause — the server stops accepting new requests (/pause_generation).
2. Drain / abort — it waits out a short grace period for in-flight requests to finish, aborting any still running, so nothing is mid-generation on the old adapter.
3. Load — the new adapter is loaded onto every server.
4. Flush — the KV cache is flushed to drop entries computed with the old adapter's weights.
5. Resume — generation continues; new requests now use the new adapter.

The bug. The old code reloaded the adapter under a fixed name (lora_1), which required unloading the previous one before step 3. SGLang's /unload_lora_adapter calls wait_for_unload, which blocks until the adapter's usage counter reaches zero, But the requests aborted during the drain never released their hold on that counter, so it stayed non-zero and the unload hung forever, freezing the entire pipeline. This only struck LoRA runs under load (full fine-tuning takes a different, unload-free path), which made it look intermittent.

The fix. Never unload. Each sync loads the new adapter under a fresh versioned name (lora_v1, lora_v2, …) and lets SGLang's memory-pool LRU reclaim the stale ones. The deadlock-prone unload is removed entirely, while the pause → drain/abort → load → flush → resume sequence above is unchanged, so the swap stays correct. A lingering old adapter only costs GPU memory, which is bounded by max_loras_per_batch / max_loaded_loras.

Validation

Qwen3-1.7B, math DP-scaled SGLang: clean monotonic rise over the first ~100 steps with importance weight pinned at 1.0 throughout —

metric	base	step 100
eval-avg/math500/avg@4	70.0	77.55
eval-avg/overall_avg	30.2	36.58