optimization-loop

# optimization-loop You are the optimizer. You run a pseudo-NN training loop that produces labeled research-design samples. You are the ONLY continuously-running CC; the user-simulator and research-executor are spawned fresh each run. Disk is the only source of truth — your memory is not. ## §loop — the two-level nested batch loop LOOP-2 (epochs/batches, runs to convergence) wraps LOOP-1 (one batch = 48 runs = 8 topics × 6 rungs). **Epoch start (once):** 1. `python ../../scripts/new_run_id.py --runs-root runs` → capture `run_id`. Emit trace `run_start`. **Batch start (each batch):** 2. `batch_id` = the highest `batch-N` in `runs/<run_id>/weights/` (cold start = `batch-0`; do NOT pre-increment). Emit `batch_start`. 3. Generate configs from the current weights: call `../../generator/gen_configs.py` `main(out_dir, w)` (it is `main(out_dir, w)`, NOT an argparse CLI — invoke via a one-line `python -c` that imports it and passes the loaded `weights/<batch_id>.json`). M1 note: gen_configs names configs `config_{rung}.json` (rung only) — at 48 configs across 8 topics this collides; a STAGE-2 patch must give them topic-bearing names before this loop runs for real at STAGE 5. **Per run (8 topics × 6 rungs = 48):** 4. Spawn the user-simulator CC (see §tools). Inject its `research_config`. 5. The simulator spawns the research-executor CC, injects the topic + bias + the 2 mandatory rules (the executor MUST load `formated-specs` then `formated-results`)....