feat: add orchestrator debug mode (no-inference, no-trainer)

[samsja]

Summary

Adds an orchestrator debug mode for running without inference or a trainer, plus a deterministic test environment for memory/pipeline ablations.

Changes

Orchestrator debug mode (orchestrator.py, utils.py, watcher.py, client.py)

Adds [orchestrator.debug] with two flags:

• no_inference: Uses a NoOpInferencePool (no model calls), clears the renderer, skips renderer/VLM/pool-size validators, and skips NCCL weight broadcast init and teacher inference pool setup.
• no_trainer: Uses filesystem rollout transport and a DebugWeightWatcher that advances policy.version locally so dispatch gating and off-policy cancellation still run.

update_dispatch_gate is now async and drives local policy updates in no-trainer mode. RSS memory logging via psutil is added after each step when debug mode is enabled.

no-inference-env (tests/debug_envs/no_inference_env/)

A deterministic test environment that synthesizes heavy multi-turn trajectories (configurable sequence length, turns, optional routed experts, rollout delay) without requiring vLLM. Useful for memory/pipeline ablations.

Docs (skills/training/start-run/SKILL.md)

Documents uv run orchestrator ... --debug.no-inference --debug.no-trainer and the no-inference env configuration.

---

Note

Medium Risk
Touches orchestrator startup, weight watching, and dispatch gating on the training hot path, but debug flags default off; production behavior should be unchanged unless debug is enabled.

Overview
Adds orchestrator.debug with no_inference and no_trainer so you can run the orchestrator alone for pipeline/memory ablations without vLLM or a trainer.

With no_inference, setup uses a NoOpInferencePool, clears the renderer, skips teacher/NCCL init, and relaxes renderer/VLM/pool-size validators. With no_trainer, rollout transport is forced to filesystem and a DebugWeightWatcher bumps policy.version locally after each shipped batch (via async update_dispatch_gate) so dispatch gating and off-policy cancellation still behave like production. Debug runs also log process RSS (orchestrator + children) with psutil after each step.

Ships no-inference-env: a deterministic env that builds heavy multi-turn trajectories (long seqs, tools, optional routed-expert payloads, configurable rollout delay) without calling the model. Wired into the envs extra and documented for uv run orchestrator ... --debug.no-inference --debug.no-trainer.

^{Reviewed by Cursor Bugbot for commit f2c79f7. Bugbot is set up for automated code reviews on this repo. Configure here.}

[samsja]

Closing in favor of #2868 which now includes both the memory-saving defaults and the EP auto changes.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bcb4c2a. Configure here.}