[biswapanda]

Summary

• Accept a strict positive world size from the single discovered Dynamo sidecar while retaining the one-logical-worker constraint.
• Include world size in the pinned topology identity and reject configured/discovered TP mismatches before NCCL initialization.
• Initialize the broadcaster with the full inference world size and validate one successful collective result per TP rank.

Validation

• uv run --frozen pytest tests/unit/inference/test_dynamo.py -q (17 passed)
• uv run --frozen ruff check src/prime_rl/inference/dynamo.py tests/unit/inference/test_dynamo.py
• uv run --frozen ruff format --check src/prime_rl/inference/dynamo.py tests/unit/inference/test_dynamo.py
• TP=2 Qwen3-0.6B + MathEnv completed 3/3 NCCL training steps on the dev GPU pod at a556ad26e; policy versions v0-v3 transferred successfully with zero rollout errors or truncation. The current tip only inlines the unchanged topology tuple and removes its redundant direct test.