fix(eval): end each A/B session with its run, and attribute runs to arms

Two defects made the earlier A/B numbers meaningless:

  * Without `--stop-on-run-end` a session contained several runs, so runs could
    not be attributed to an arm.
  * The script then read "the last 2N runs", which silently compared runs from
    earlier experiments rather than the two arms.

Pass `--stop-on-run-end`, and attribute each run record to its arm by diffing
the history directory around each session.

Also raise the step budget and warn on a thin sample: measured, a run still
going at step 600 produces NO run record, so a short budget biases the sample
toward runs that died early -- exactly the wrong bias for this question.

ab_card_skip.sh

@@ -3,74 +3,208 @@
 #
 #   ./ab_card_skip.sh <sessions-per-arm> <steps-per-session>
 #
-# Runs N sessions with --card-skip-policy jev, then N with --card-skip-policy
+# Everything except --card-skip-policy is identical between the arms.
-# combined, and prints a comparison of deck size and progress.
 #
-# Everything else is identical between the arms, so any difference is the
+# Three things this script does that earlier versions did not, each because an
-# policy. (The earlier before/after comparison was confounded: several bug
+# earlier attempt produced meaningless numbers:
-# fixes landed at the same time as the policy change.)
+#
+#   1. --stop-on-run-end, so ONE SESSION = ONE RUN. Without it the bot dies,
+#      returns to the menu and starts a fresh run inside the same session, and a
+#      session yields several run records that cannot be attributed to an arm.
+#   2. Each run record is attributed to the arm that produced it, by diffing the
+#      history directory around each session. The old version just took "the
+#      last 2N runs", which silently compared old runs from earlier experiments.
+#   3. Every attributed run carries a hash of the decision layer, and the report
+#      compares only runs from ONE revision. A change to brain.py / facts.py /
+#      jev.py / run.py changes how both arms behave, so runs collected before it
+#      are not comparable with runs collected after it -- and the working copy is
+#      usually dirty, so a commit id would not distinguish the revisions.
 set -u
 
+N="${1:-5}"
+# Steps per session. This must be generous: measured, a run that is still going
+# at step 600 produces NO run record, so the sample silently fills with runs
+# that died early -- exactly the wrong bias. --stop-on-run-end is the real stop
+# condition; this is only a safety cap.
+STEPS="${2:-4000}"
+HIST="$HOME/Library/Application Support/SlayTheSpire2/steam/76561198141226155/modded/profile1/saves/history"
+RESULTS=/tmp/ab_results.tsv
+
+# The revision every arm in this invocation is collected from. Content hash, not
+# a commit id: the decision layer is normally edited in place, uncommitted.
+#
+# Recomputed BEFORE EVERY SESSION, not once: this file is edited in place while
+# experiments run, which is the whole reason the guard exists. A hash taken once
+# at startup would be stamped onto every row regardless, and because $RESULTS is
+# truncated below, the mixed-revision warning could never fire.
+CODE_FILES=(brain.py facts.py jev.py run.py sts2.py)
+code_hash() {
+  cat "${CODE_FILES[@]}" | { shasum -a 256 2>/dev/null || sha256sum; } | cut -c1-12
+}
+CODE_HASH=$(code_hash)
+ABORTED=0
+echo "code revision: $CODE_HASH  (${CODE_FILES[*]})"
+
 # Fail fast if the bot cannot start a run (e.g. a pending Timeline epoch).
 if ! python3 run.py --steps 1 >/tmp/preflight.log 2>&1; then
   cat /tmp/preflight.log
   exit 2
 fi
 
-N="${1:-5}"
+: > "$RESULTS"
-STEPS="${2:-600}"
+
+snapshot() { ls "$HIST" 2>/dev/null | sort; }
 
 run_arm() {
-  local policy="$1" tag="$2"
+  local policy="$1"
   for i in $(seq 1 "$N"); do
-    echo "--- $tag $i/$N ---"
+    local before after attempt=0 rc=1 session_hash
+    before=$(snapshot)
+
+    # An arm collected across two revisions is not an arm. Stop the experiment
+    # rather than stamp rows with a revision that did not produce them.
+    session_hash=$(code_hash)
+    if [ "$session_hash" != "$CODE_HASH" ]; then
+      echo
+      echo "!! ABORT: the decision layer changed mid-experiment."
+      echo "!!   started on $CODE_HASH, now $session_hash"
+      echo "!!   Rows already collected are in $RESULTS and are NOT comparable"
+      echo "!!   with anything collected on the other revision."
+      echo "!!   Re-run BOTH arms from one revision."
+      ABORTED=1
+      return 1
+    fi
+
+    while [ "$attempt" -lt 3 ]; do
+      attempt=$((attempt + 1))
+      echo "--- $policy $i/$N (attempt $attempt) rev=$session_hash ---"
       python3 run.py --steps "$STEPS" --pause 0.3 \
-      --card-skip-policy "$policy" > "/tmp/ab_${tag}_${i}.log" 2>&1
+        --card-skip-policy "$policy" --stop-on-run-end \
-    echo "    exit=$? takes=$(grep -c 'select_card_reward' "/tmp/ab_${tag}_${i}.log") skips=$(grep -c 'skip_card_reward' "/tmp/ab_${tag}_${i}.log")"
+        > "/tmp/ab_${policy}_${i}.log" 2>&1
+      rc=$?
+      [ "$rc" -eq 0 ] && break
+      [ "$rc" -ne 3 ] && break          # 2 = blocked, not worth retrying
+      echo "    retrying after model failures..."
+      sleep 20
     done
+
+    after=$(snapshot)
+    local new_runs session
+    new_runs=$(comm -13 <(echo "$before") <(echo "$after") | tr '\n' ' ')
+    # The session id the bot printed, so an attribution row joins to the same
+    # decisions.jsonl rows and to sessions.jsonl. Without it the two ends of the
+    # join carry different identifiers and nothing lines up.
+    session=$(sed -n 's/^session: //p' "/tmp/ab_${policy}_${i}.log" | head -1)
+    echo "    exit=$rc session=${session:-?} takes=$(grep -c 'select_card_reward' "/tmp/ab_${policy}_${i}.log") skips=$(grep -c 'skip_card_reward' "/tmp/ab_${policy}_${i}.log") runs='${new_runs}'"
+
+    # Stamped with the hash that produced THIS run, not the startup hash.
+    for r in $new_runs; do
+      printf '%s\t%s\t%s\t%s\n' "$policy" "$r" "$session_hash" "${session:-?}" >> "$RESULTS"
+    done
+  done
+  return 0
 }
 
 echo "### ARM A: jev decides skip"
-run_arm jev jev
+run_arm jev
+echo
+if [ "$ABORTED" -eq 0 ]; then
+  echo "### ARM B: combined (jev + a floor)"
+  run_arm combined
+else
+  echo "### ARM B: skipped (aborted mid-experiment)"
+fi
 
 echo
-echo "### ARM B: combined (jev + a floor)"
+POLICY=combined RESULTS="$RESULTS" CODE_HASH="$CODE_HASH" HIST="$HIST" python3 - <<'PY'
-run_arm combined combined
+import json, os, statistics as st
 
-echo
+# The same directory the shell snapshotted. A second hardcoded copy of this path
-python3 - <<'PY'
+# silently reported "no runs recorded" for sessions that did produce runs.
-import json, glob, os, statistics as st
+hist = os.environ["HIST"]
-
+results = os.environ["RESULTS"]
-H = os.path.expanduser(
+this_hash = os.environ["CODE_HASH"]
-    "~/Library/Application Support/SlayTheSpire2/steam/"
-    "76561198141226155/modded/profile1/saves/history"
-)
 
 rows = []
-for f in glob.glob(H + "/*.run"):
+with open(results) as fh:
-    d = json.load(open(f))
+    for line in fh:
+        parts = line.rstrip("\n").split("\t")
+        if len(parts) < 2 or not parts[1]:
+            continue
+        rows.append((parts[0], parts[1],
+                     parts[2] if len(parts) > 2 and parts[2] else "unstamped",
+                     parts[3] if len(parts) > 3 and parts[3] else "?"))
+
+hashes = sorted({h for _, _, h, _ in rows})
+print(f"code revision(s) in {results}: {', '.join(hashes) if hashes else '(none)'}")
+if len(hashes) > 1:
+    print()
+    print("!! MIXED REVISIONS. A change to the decision layer changes how BOTH")
+    print("!! arms behave, so a run collected before it is not comparable with")
+    print("!! one collected after it. Only ONE revision is scored below; re-run")
+    print("!! BOTH arms on one revision before interpreting anything.")
+# Score the revision this invocation collected from when it has rows; otherwise
+# the most recently appended one. Never "the largest hash": hex is not ordered.
+if this_hash in hashes:
+    current = this_hash
+elif rows:
+    current = rows[-1][2]
+else:
+    current = this_hash
+print(f"scoring revision: {current}")
+print()
+
+arms = {"jev": [], "combined": []}
+sessions = {"jev": 0, "combined": 0}
+for policy, run, code, session in rows:
+    if policy not in arms or code != current:
+        continue
+    sessions[policy] += 1
+    path = os.path.join(hist, run)
+    if not os.path.exists(path):
+        continue
+    d = json.load(open(path))
     pl = (d.get("players") or [{}])[0]
     mph = d.get("map_point_history") or []
-    rows.append({
+    arms[policy].append({
-        "t": d.get("start_time", 0),
         "killed": d.get("killed_by_encounter", "?").replace("ENCOUNTER.", ""),
         "deck": len(pl.get("deck") or []),
         "pts": sum(len(a) for a in mph if isinstance(a, list)),
+        "session": session,
     })
-rows.sort(key=lambda r: r["t"])
 
-# The last 2N runs are this experiment's.
+print(f"{'arm':<10} {'sess':>4} {'runs':>4} {'deck':>6} {'pts':>6} {'added/pt':>9} {'act2':>6}")
-total = 2 * int(os.environ.get("AB_N", "5"))
+print("-" * 54)
-tail = rows[-total:] if total <= len(rows) else rows
+for name, grp in arms.items():
-armA, armB = tail[:len(tail)//2], tail[len(tail)//2:]
-
-print(f"{'arm':<12} {'n':>2} {'deck avg':>9} {'pts avg':>8} {'reached act 2':>14}")
-for name, grp in (("jev", armA), ("combined", armB)):
     if not grp:
+        print(f"{name:<10} {sessions[name]:>4} {0:>4}   (no runs recorded)")
         continue
     decks = [r["deck"] for r in grp]
     pts = [r["pts"] for r in grp]
     act2 = sum(1 for r in grp if r["pts"] > 17)
-    print(f"{name:<12} {len(grp):>2} {st.mean(decks):>9.1f} {st.mean(pts):>8.1f} "
+    added = max(0.0, st.mean(decks) - 10) / max(1e-9, st.mean(pts))
-          f"{act2}/{len(grp):>12}")
+    print(f"{name:<10} {sessions[name]:>4} {len(grp):>4} {st.mean(decks):>6.1f} "
+          f"{st.mean(pts):>6.1f} {added:>9.2f} {act2:>4}/{len(grp)}")
+
+print()
+print("sess     = sessions that produced a COMPLETED run record")
+print("runs     = run records attributed to this arm, on the scored revision")
+print("added/pt = non-starting cards per map point reached  <- the metric that")
+print("           matters, because deck size at death scales with run length")
+print("act2     = runs past the Act 1 boss (17 map points)")
+
+thin = [n for n, g in arms.items() if len(g) < 5]
+if thin:
+    print()
+    print(f"!! THIN SAMPLE: {', '.join(thin)} have fewer than 5 runs.")
+    print("!! Sessions that hit the step cap mid-run produce NO record, so a thin")
+    print("!! sample is biased toward runs that died early. Raise STEPS and re-run.")
+print()
+for name, grp in arms.items():
+    if grp:
+        print(f"{name} killed by: {[r['killed'] for r in grp]}")
+        print(f"{name} sessions : {[r['session'] for r in grp]}")
+print()
+print("sessions are the join key: grep one in capture/decisions.jsonl for that")
+print("session's decisions, and in capture/sessions.jsonl for its outcome row.")
 PY