autoresearch-impllisted

# Bridge: Autoresearch (Implementation-Level) **Announce at start:** "I'm using the bridge:autoresearch-impl skill to iterate on the implementation against a metric before it goes to review." ## Purpose This mirrors karpathy/autoresearch's actual loop more directly than the plan-level version does, because there's now real code and (usually) real tests to run: propose a change to one file/module, run it for a fixed budget, check whether a single metric got better, keep or discard, repeat. The autoresearch repo enforces this by restricting the agent to one file (`train.py`) and running every variant for a fixed 5-minute wall-clock budget so results are comparable. Do the same here: keep each round's diff small and reviewable, and bound the loop by rounds or wall-clock time, not by "iterate until it feels done." Only run this on real, working code — an implementation that already passes its own tests/builds. This skill improves a working baseline; it is not a substitute for finishing the implementation first. ## Step 0 — Confirm there's something to iterate on Run `git status` and `git diff` (or check the relevant worktree/branch) to confirm there's a completed implementation to work from. If the working tree is dirty with unrelated changes, stop and ask before touching anything — per this repo's git safety rules, uncommitted work that isn't yours to discard must be stashed or committed first, never dropped. ## Step 1 — Pin down scope, budget, and metric Same three knob