feat: add Qwen3-30B router-replay debug config

[mikasenghaas]

Summary

• Add configs/debug/r3.toml — a Qwen3-30B-A3B (Thinking) RL debug config with router replay enabled (trainer.enable_router_replay = true, paired with inference.enable_return_routed_experts = true).
• Ported from the rlm5/qwen30b-debug.toml config used on the perf/r3 experiment line, updated to the current config schema and cleaned up.

Schema migration (vs. the old config)

• orchestrator.rollouts_per_example → orchestrator.group_size.
• Removed fields that no longer exist in the schema: max_async_level, orchestrator.use_token_client, orchestrator.use_renderer, [orchestrator.buffer] (online_difficulty_filtering).
• Dropped the explicit [orchestrator.client] block — X-Session-ID = trajectory_id is now auto-set by the orchestrator.
• Dropped name from [inference.model] — the shared top-level [model] name now propagates to all sub-configs.
• Eval is disabled by omitting the [orchestrator.eval] block (the schema now rejects env = []).

Debug cleanup

• Removed the one-off perf-experiment fields and their narration comments: use_process_pool (round 4), dump_raw_rollouts (round 5), gather_chunk_size (round 6).
• Removed [trainer.experimental.token_export] (per-token JSONL rollout export — debug-only).

Verification

Validated that the config loads against RLConfig (same path as tests/unit/test_configs.py::test_load_configs):

PARSED OK
trainer.enable_router_replay = True
inference.enable_return_routed_experts = True
orchestrator.group_size = 8
student.client.extra_headers_from_state = {'X-Session-ID': 'trajectory_id'}
inference.model.name (propagated) = Qwen/Qwen3-30B-A3B-Thinking-2507
renderer.name = qwen3

🤖 Generated with Claude Code