chore(atlas): scaffold autoresearch campaign for braid_search.py optimization

.atlas/benchmark.sh

@@ -0,0 +1,29 @@
+#!/usr/bin/env bash
+# Generated by `atlas autoresearch` — this file is YOURS, edit it freely.
+#
+# Contract: write {"score": <float>, "examples": [...]} to $ATLAS_OPTIMIZE_RESULT.
+# The candidate artifact has already been written into this worktree at
+#   ${ATLAS_OPTIMIZE_TARGET}   (the repo-relative path: 4-Infrastructure/shim/braid_search.py)
+# so the only thing a candidate can change is that file — the optimizer runs
+# every candidate in its own throwaway git worktree, which keeps the evaluator
+# and data pinned. score.py turns the run's output into the score; .atlas/gate.sh
+# is the Goodhart guard.
+set -uo pipefail
+HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+OUT="$(mktemp)"
+
+# --- the repo's run command (auto-detected) --------------------------------
+# Edit this line if the campaign should run something different.
+( python3 "$ATLAS_OPTIMIZE_TARGET" ) >"$OUT" 2>&1
+RC=$?
+# ---------------------------------------------------------------------------
+
+if [ "$RC" -ne 0 ]; then
+  echo "[benchmark] run command exited $RC — candidate failed" >&2
+  tail -n 40 "$OUT" >&2
+  exit "$RC"
+fi
+
+# Parse the optimized metric out of the run output (edit .atlas/score.py to
+# change which number is read, or to emit richer per-example feedback).
+python3 "$HERE/score.py" --stdout "$OUT"

.atlas/gate.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Generated by `atlas autoresearch` — the Goodhart guard.
+#
+# A candidate may only change the artifact under test. If a run modifies any
+# OTHER tracked file (e.g. rewrites the evaluator or test data to fake a pass),
+# this gate fails (exit non-zero) and the candidate is scored worst, so the
+# Pareto front never elevates a cheat. Relax IGNORE_RE if your run legitimately
+# rewrites tracked files (checkpoints, logs, and .atlas/ are already ignored).
+set -uo pipefail
+WT="${1:-$PWD}"
+TARGET="${2:-${ATLAS_OPTIMIZE_TARGET:-}}"
+REL="${TARGET#"$WT"/}"
+IGNORE_RE='^(outputs/|out/|checkpoints/|runs/|wandb/|\.atlas/|.*\.log$|.*\.ckpt$|.*\.pt$|.*\.bin$|.*\.safetensors$)'
+CHANGED="$(git -C "$WT" status --porcelain --untracked-files=no 2>/dev/null \
+  | sed 's/^...//' \
+  | { [ -n "$REL" ] && grep -vF -- "$REL" || cat; } \
+  | grep -Ev "$IGNORE_RE" || true)"
+if [ -n "$CHANGED" ]; then
+  echo "[gate] the run modified tracked files other than the candidate:" >&2
+  echo "$CHANGED" >&2
+  exit 1
+fi
+exit 0

.atlas/score.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Generated by `atlas autoresearch`. Parse one metric out of a run's output
+into the Atlas benchmark contract.
+
+EDIT the METRIC line if the wrong number is being read, or rewrite parse() for
+a richer evaluator (per-example feedback is what makes GEPA converge — see
+.atlas/benchmark-contract or the atlas-optimize skill)."""
+import argparse
+import json
+import os
+import re
+import sys
+
+# autoresearch guessed this metric name from your --goal.
+METRIC = "find_optimal_crossing"          # <-- EDIT ME if this is the wrong metric
+HIGHER_IS_BETTER = False
+
+# Matches "<metric> = 0.74", "<metric>: 0.74", "<metric> 0.74".
+PATTERN = re.compile(
+    re.escape(METRIC) + r"\s*[:=]?\s*([+-]?[0-9]*\.?[0-9]+(?:[eE][+-]?[0-9]+)?)"
+)
+
+
+def parse(text):
+    matches = PATTERN.findall(text)
+    if not matches:
+        return None
+    return float(matches[-1])  # last occurrence == final epoch / final eval
+
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--stdout", required=True, help="file holding the run's captured output")
+    args = ap.parse_args()
+    text = open(args.stdout, encoding="utf-8", errors="replace").read()
+    score = parse(text)
+    if score is None:
+        sys.stderr.write(
+            f"[score] could not find metric '{METRIC}' in the run output.\n"
+            "        Fix: edit .atlas/score.py (METRIC=...), or pass an explicit\n"
+            "        --benchmark to 'atlas autoresearch'.\n"
+        )
+        sys.exit(3)
+    direction = "higher is better" if HIGHER_IS_BETTER else "lower is better"
+    result = {
+        "score": score,
+        "examples": [{
+            "id": METRIC,
+            "score": score,
+            "pass": True,
+            "feedback": f"{METRIC} = {score} (parsed from run output; {direction})",
+        }],
+        "feedback": f"Final {METRIC} = {score}.",
+    }
+    out = os.environ.get("ATLAS_OPTIMIZE_RESULT")
+    payload = json.dumps(result)
+    if out:
+        with open(out, "w", encoding="utf-8") as fh:
+            fh.write(payload)
+    else:
+        sys.stdout.write(payload)
+
+
+if __name__ == "__main__":
+    main()

atlas.json

@@ -0,0 +1,11 @@
+{
+  "project_id": "b4390de5-14d9-46dc-8be7-9bb36291e0b4",
+  "auto": {
+    "enabled": true,
+    "log": true,
+    "web": "prefer_atlas",
+    "capture": "suggest"
+  },
+  "privacy": "private",
+  "source_ids": []
+}