[codex] add fake R3 orchestrator debug mode

[samsja]

Summary

Adds a reusable orchestrator debug mode for memory and scheduling ablations without requiring real inference or trainer processes.

Memory reductions are split into the stacked follow-up PR: #2807.

What changed

• Adds debug.no_inference, debug.no_trainer, debug.fake_tokenizer, and debug.log_memory orchestrator flags.
• Lets rl local run the full orchestrator without spawning inference or trainer subprocesses when debug flags are enabled.
• Adds no-op inference, training transport, and weight watcher implementations for debug runs.
• Adds tests/debug_envs/fake_r3_trajectory, a deterministic fake environment that emits heavy pre-tokenized multi-turn trajectories with optional GLM5-shaped R3 payloads.
• Keeps local ablation TOML configs out of git; the committed reusable artifact is the debug environment plus orchestrator debug plumbing.

Fake R3 environment

The fake env is user-configurable through env args:

• turns
• seq_len
• prompt_len
• completion_fraction
• include_r3
• routed_layers
• routed_topk
• n_routed_experts
• num_examples
• vocab_size

Defaults model the GLM5 R3 shape: 78 routed layers, top-k 8, 256 routed experts.

Validation

Validated earlier on this branch with orchestrator unit coverage and fake-R3 orchestrator stress runs. Memory optimizations are intentionally split into #2807.

[cursor]

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

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

^{Reviewed by Cursor Bugbot for commit 56440cc. Configure here.}