diagnose-failure

# /diagnose-failure ## When to use - A round's status was `killed_by_safety`, `compile_error`, or `train_error`. - User asks "why did this break?" or provides a bad config. ## Inputs - `run_dir` OR `round_dir` ## Behavior 1. Run `python -m cli.autoqec diagnose <run_dir>` for mechanical stats. 2. Read `train.log`, `config.yaml`, `metrics.json`, the latest 2–3 rounds of `history.jsonl`. 3. LLM-reason over known failure modes: - `killed_by_safety` + wall_clock → params too large for budget → suggest smaller `hidden_dim` / `layers`. - `killed_by_safety` + NaN rate → loss instability → suggest lower learning rate, gradient clipping, or switch `loss` to `focal`. - `compile_error` → DSL syntax → show offending field + schema expected value. - `train_error` + VRAM → OOM → suggest smaller batch_size or profile=dev. 4. Write `round_N/diagnosis.md` with: - Root cause (1 sentence) - Evidence citation (log line numbers) - Recommended fix (as a patched DSL YAML snippet) - **Do NOT apply** the fix — the user runs it themselves. ## Tool-use - Read only. Never Edit or Write outside the `diagnosis.md` file.