txt2sql (BIRD) single-node FSDP baseline crashes silently at Ray startup

[sfc-gh-kganesan]

Summary

The single-node FSDP-native baseline for the txt2sql (BIRD) recipe —
recipes/rl/skyrl/txt2sql/run_qwen3_8b_bird_grpo_fsdp.sh — crashes silently
during Ray startup (worker dies ~30s after ray init, no traceback propagated
to the driver). The Arctic RL sibling (run_qwen3_8b_bird_grpo_arl.sh) comes up
and trains fine on the same box, and the LoongRL FSDP baseline
(long_context_qa/run_qwen3_8b_loongrl_grpo_fsdp.sh) also works — so this is
specific to the BIRD + FSDP path on a single node.

Impact

Blocks the ZoRRo speedup A/B for txt2sql on a single node. Convergence and
the ARL path for BIRD are validated; only the FSDP wall-clock baseline is
missing, so the BIRD speedup number is not yet reproduced end-to-end.

What was tried

• Cleared orphaned Ray processes + leaked /dev/shm IPC segments.
• Confirmed PromptDataset builds correctly in isolation (fork and spawn).
• Ran BasePPOExp inline (no Ray remote) — still a silent exit.
• Combed Ray session logs (worker-*.err/out, raylet.out, monitor.log) —
  no traceback; Ray reports the remote task as never actually started.

Likely area

Worker-side setup (placement group / vLLM engine init) for the BIRD FSDP config
failing without error propagation; possibly environment instability after
repeated force-kills. Needs a clean-box repro with Ray error propagation forced
on.

Acceptance

• Root-cause the silent worker death.
• run_qwen3_8b_bird_grpo_fsdp.sh comes up on 1×8 H200.
• BIRD 8B ARL-vs-FSDP wall-clock A/B captured and posted on PR recipes/rl/skyrl: Arctic RL × SkyRL recipes (simple / txt2sql / long_context_qa) #32.