post-run-reviewlisted

# /post-run-review — structured Assess → Codify Ports the AIDE `submit_review` pattern (MIT, `aide/agent.py` L19–44 + `parse_exec_result` L296–339). Ours adds the C4 failure-class classification — `bug | bad-hyperparam | bad-hypothesis` — which no other autonomous-research harness currently cleanly separates. ## When to invoke Automatically, as the final step of the Assess phase, after any experiment with a journal node (`journal.py add …` was called at Plan time). Don't invoke on trivial scripts. ## What the skill does Given: - Journal node id (`--node N`) - Log / stdout from the run (`--log PATH`) - Optional metric override (`--metric NAME=VAL`) The agent reads the log, forms a structured verdict, and writes: ```json { "is_buggy": 0 | 1, "failure_class": null | "bug" | "bad-hyperparam" | "bad-hypothesis", "metric_name": "val_loss", "metric_value": 3.42, "lower_is_better": 1, "summary": "one-paragraph what-happened", "learn_block": "<optional [LEARN] emitted separately>", "dead_end_block": "<optional [DEAD-END] emitted separately>" } ``` and then: 1. Calls `journal.py close <node> --metric ... --is-buggy ... --failure-class ... --analysis "..."`. 2. If `is_buggy=1` and it's a clear durable rule: emits a `[LEARN]` block in its response (Stop hook captures it). 3. If `failure_class=bad-hypothesis`: emits a `[DEAD-END]` block. 4. If `actual_delta` for this run was predicted: `calibration.py close <run_id> "<hypothesis>" <actual>`. ## Failure-clas