@mcptoolshop/style-dataset-lab 3.4.0 → 3.5.0

package/scripts/colab/ai-toolkit-override-py315.txt

@@ -0,0 +1,45 @@
+# CPython 3.15.1 dependency override file for ostris/ai-toolkit (Lane 3 LoRA training)
+# Upstream tracking issue: https://github.com/ostris/ai-toolkit/issues/1087
+# Reference: DESIGN-v3.md section 3.5 (CPython 3.15.1 wheel audit)
+# Note: av==16.0.1 and torchcodec==0.9.1 are dropped for image LoRA lanes.
+scipy==1.18.1
+matplotlib==3.11.2
+torchao==0.18.0
+safetensors==0.7.0
+transformers==5.5.3
+lycoris-lora==1.8.3
+flatten_json==0.1.14
+pyyaml==6.0.3
+oyaml==1.0
+tensorboard==2.20.0
+kornia==0.9.0
+invisible-watermark==0.2.0
+einops==0.8.2
+accelerate==1.13.0
+toml==0.10.2
+albumentations==1.4.15
+albucore==0.0.16
+pydantic==2.13.2
+omegaconf==2.3.0
+k-diffusion==0.1.1.post1
+open_clip_torch==3.3.0
+timm==1.0.22
+prodigyopt==1.1.2
+controlnet_aux==0.0.10
+python-dotenv==1.2.2
+bitsandbytes==0.49.2
+hf_transfer==0.1.9
+lpips==0.1.4
+pytorch_fid==0.3.0
+optimum-quanto==0.2.4
+sentencepiece==0.2.1
+huggingface_hub==1.10.1
+peft==0.18.1
+python-slugify==8.0.4
+opencv-python==4.13.0.92
+pytorch-wavelets==1.3.0
+setuptools==82.0.1
+librosa==0.11.0
+mutagen==1.47.0
+pillow==12.2.0
+numpy==2.4.4

package/scripts/colab/colab_window.sh

@@ -0,0 +1,587 @@
+#!/usr/bin/env bash
+# Shared Colab window lifecycle, lock, live wall/cap/vitals/RAM guards, token scrubber, and EXIT trap.
+set -euo pipefail
+
+colab_resolve_python() {
+  if [ -n "${SDLAB_PYTHON:-}" ] && [ -x "${SDLAB_PYTHON}" ]; then
+    echo "${SDLAB_PYTHON}"
+    return 0
+  fi
+  if command -v python3 >/dev/null 2>&1; then
+    command -v python3
+    return 0
+  fi
+  command -v python
+}
+
+colab_refuse_nonempty_out_dir() {
+  local out_dir="$1"
+  if [ -d "${out_dir}" ] && [ -n "$(ls -A "${out_dir}" 2>/dev/null)" ]; then
+    echo "REFUSED: OUT_DIR_NOT_EMPTY: ${out_dir} already exists and is not empty" >&2
+    return 97
+  fi
+}
+
+colab_guard_fake_cli_if_test() {
+  local colab_bin
+  colab_bin="$(command -v colab || true)"
+  if [ -z "${colab_bin}" ]; then
+    echo "REFUSED: colab CLI not found in PATH" >&2
+    return 98
+  fi
+  if [ "${SDLAB_TEST_FAKE_COLAB:-0}" = "1" ]; then
+    if ! grep -q "^#SDLAB_FAKE_COLAB_CLI=1$" "${colab_bin}" 2>/dev/null; then
+      echo "REFUSED: REAL_COLAB_FORBIDDEN: colab resolved to ${colab_bin} without #SDLAB_FAKE_COLAB_CLI=1 marker" >&2
+      return 99
+    fi
+  fi
+}
+
+colab_verify_input_shas() {
+  local py_bin="$1"
+  local guards_py="$2"
+  shift 2
+  local args=()
+  for pair in "$@"; do
+    args+=("--verify-sha" "${pair}")
+  done
+  "${py_bin}" "${guards_py}" "${args[@]}" </dev/null
+}
+
+colab_poll_usage_and_ram() {
+  local py_bin="$1"
+  local out_dir="$2"
+  local fault_mode="${SDLAB_FAULT_POLL_HELPER:-}"
+  if [ -f "${out_dir}/start_balance.txt" ] && [ -n "${SDLAB_FAULT_POLL_HELPER_MIDRUN:-}" ]; then
+    fault_mode="${SDLAB_FAULT_POLL_HELPER_MIDRUN}"
+  fi
+  if [ "${fault_mode}" = "crash" ]; then
+    return 137
+  elif [ "${fault_mode}" = "empty" ]; then
+    return 0
+  elif [ "${fault_mode}" = "short" ]; then
+    echo "9.85 9.85"
+    return 0
+  fi
+  local default_rate="${RATE_CCU_PER_HOUR:-0.26}"
+  local guards_py
+  guards_py="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/launch_guards.py"
+
+  local usage_text=""
+  local usage_rc=0
+  usage_text="$(colab usage </dev/null 2>>"${out_dir}/colab.log")" || usage_rc=$?
+  printf '%s\n' "${usage_text}" >> "${out_dir}/usage.log"
+
+  "${py_bin}" - "${out_dir}" "${default_rate}" "${usage_rc}" "${usage_text}" "${guards_py}" << 'PYEOF'
+import os
+import re
+import sys
+from pathlib import Path
+
+out_dir = Path(sys.argv[1])
+default_rate = float(sys.argv[2])
+usage_rc = int(sys.argv[3])
+raw_usage = sys.argv[4]
+guards_py = Path(sys.argv[5])
+
+sys.path.insert(0, str(guards_py.parent))
+import launch_guards as lg
+
+
+def _read_float(p: Path) -> float | None:
+    if not p.exists():
+        return None
+    try:
+        return float(p.read_text(encoding="utf-8").strip())
+    except Exception:
+        return None
+
+
+start_f = out_dir / "start_balance.txt"
+end_f = out_dir / "end_balance.txt"
+rate_f = out_dir / "rate_measured.txt"
+rate_pre_f = out_dir / "rate_measured_pre_open.txt"
+rate_post_f = out_dir / "rate_measured_post_open.txt"
+rate_proj_f = out_dir / "rate_for_projection.txt"
+rate_src_f = out_dir / "rate_source.txt"
+err_f = out_dir / "usage_parse_error.txt"
+opened_marker = out_dir / "session_opened.txt"
+
+session_opened = opened_marker.exists()
+existing_pre = _read_float(rate_pre_f)
+existing_post = _read_float(rate_post_f)
+
+parsed = lg.parse_usage_text(
+    raw_usage,
+    usage_rc=usage_rc,
+    configured_rate=default_rate,
+    session_opened=session_opened,
+    post_open_rate=existing_post,
+    pre_open_rate=existing_pre,
+)
+
+if parsed["ok"]:
+    bal = parsed["balance"]
+    if not start_f.exists():
+        start_f.write_text(f"{bal}\n", encoding="utf-8", newline="\n")
+    start_bal_str = start_f.read_text(encoding="utf-8").strip()
+    end_f.write_text(f"{bal}\n", encoding="utf-8", newline="\n")
+    cur_bal_str = f"{bal}"
+    if parsed["rate_measured_pre_open"] is not None:
+        rate_pre_f.write_text(f"{parsed['rate_measured_pre_open']}\n", encoding="utf-8", newline="\n")
+    if parsed["rate_measured_post_open"] is not None:
+        rate_post_f.write_text(f"{parsed['rate_measured_post_open']}\n", encoding="utf-8", newline="\n")
+    if parsed["rate_measured_post_open"] is not None:
+        rate_f.write_text(f"{parsed['rate_measured_post_open']}\n", encoding="utf-8", newline="\n")
+    elif parsed["rate_measured_pre_open"] is not None:
+        rate_f.write_text(f"{parsed['rate_measured_pre_open']}\n", encoding="utf-8", newline="\n")
+    rate_proj_f.write_text(f"{parsed['rate_for_projection']}\n", encoding="utf-8", newline="\n")
+    rate_src_f.write_text(f"{parsed['rate_source']}\n", encoding="utf-8", newline="\n")
+    err_tok = "NONE"
+else:
+    err_msg = str(parsed["error"] or "COLAB_USAGE_PARSE_FAILED")
+    err_f.write_text(f"{err_msg}\n", encoding="utf-8", newline="\n")
+    if end_f.exists():
+        end_f.unlink()
+    start_bal_str = start_f.read_text(encoding="utf-8").strip() if start_f.exists() else "NONE"
+    cur_bal_str = "NONE"
+    rate_proj_f.write_text(f"{parsed['rate_for_projection']}\n", encoding="utf-8", newline="\n")
+    rate_src_f.write_text(f"{parsed['rate_source']}\n", encoding="utf-8", newline="\n")
+    err_tok = "COLAB_USAGE_PARSE_FAILED"
+
+# Parse VM /proc/meminfo and cgroup memory from vitals.log when present
+anon_mb = float(os.environ.get("CURRENT_ANON_MB", "512")) if "os" in sys.modules else 512.0
+mem_total_mb = 12288.0
+cgroup_max_mb = ""
+vitals_f = out_dir / "vitals.log"
+if vitals_f.exists():
+    vtxt = vitals_f.read_text(encoding="utf-8", errors="replace")
+    m_tot = re.findall(r"MemTotal:\s*(\d+)\s*kB", vtxt)
+    if m_tot:
+        mem_total_mb = round(int(m_tot[-1]) / 1024.0, 1)
+    m_anon = re.findall(r"AnonPages:\s*(\d+)\s*kB", vtxt)
+    if m_anon:
+        anon_mb = round(int(m_anon[-1]) / 1024.0, 1)
+    m_cg_cur = re.findall(r"CGROUP_MEM_CURRENT:\s*(\d+)", vtxt)
+    if m_cg_cur:
+        anon_mb = round(int(m_cg_cur[-1]) / (1024.0 * 1024.0), 1)
+    m_cg_max = re.findall(r"CGROUP_MEM_MAX:\s*(\d+)", vtxt)
+    if m_cg_max:
+        cgroup_max_mb = str(round(int(m_cg_max[-1]) / (1024.0 * 1024.0), 1))
+
+print(f"{start_bal_str} {cur_bal_str} {parsed['rate_for_projection']} {err_tok} {anon_mb} {mem_total_mb} {cgroup_max_mb}")
+PYEOF
+}
+
+colab_check_cap_and_ram() {
+  local py_bin="$1"
+  local guards_py="$2"
+  local mode="$3"
+  local elapsed_s="$4"
+  local wall_s="$5"
+  local grace_s="$6"
+  local cap="$7"
+  local margin="$8"
+  local out_dir="$9"
+
+  local metrics=""
+  local poll_rc=0
+  metrics="$(colab_poll_usage_and_ram "${py_bin}" "${out_dir}")" || poll_rc=$?
+  local start_bal="" cur_bal="" rate_val="" err_tok="" anon_mb="" mem_total_mb="" cgroup_max_mb=""
+  if [ "${poll_rc}" -ne 0 ]; then
+    err_tok="COLAB_USAGE_PARSE_FAILED"
+    printf 'COLAB_USAGE_PARSE_FAILED: usage poll helper exited with status %s\n' "${poll_rc}" > "${out_dir}/usage_parse_error.txt"
+  else
+    read -r start_bal cur_bal rate_val err_tok anon_mb mem_total_mb cgroup_max_mb <<< "${metrics}"
+    if [ -z "${metrics}" ] || [ -z "${start_bal}" ] || [ -z "${cur_bal}" ] || [ -z "${rate_val}" ] || [ -z "${err_tok}" ] || [ "${start_bal}" = "NONE" ] || [ "${cur_bal}" = "NONE" ]; then
+      err_tok="COLAB_USAGE_PARSE_FAILED"
+      if [ ! -s "${out_dir}/usage_parse_error.txt" ]; then
+        printf 'COLAB_USAGE_PARSE_FAILED: incomplete or missing usage metrics (%s)\n' "${metrics}" > "${out_dir}/usage_parse_error.txt"
+      fi
+    fi
+  fi
+
+  local guard_args=(
+    "--mode" "${mode}"
+    "--elapsed-s" "${elapsed_s}"
+    "--wall-s" "${wall_s}"
+    "--grace-s" "${grace_s}"
+    "--cap" "${cap}"
+    "--margin" "${margin}"
+    "--rate" "${rate_val:-0.26}"
+  )
+  if [ "${start_bal}" != "NONE" ] && [ -n "${start_bal}" ]; then
+    guard_args+=("--start-balance" "${start_bal}")
+  fi
+  if [ "${cur_bal}" != "NONE" ] && [ -n "${cur_bal}" ]; then
+    guard_args+=("--current-balance" "${cur_bal}")
+  fi
+  if [ "${err_tok}" != "NONE" ] && [ -n "${err_tok}" ]; then
+    local err_detail="${err_tok}"
+    if [ -f "${out_dir}/usage_parse_error.txt" ]; then
+      err_detail="$(cat "${out_dir}/usage_parse_error.txt")"
+    fi
+    guard_args+=("--usage-parse-error" "${err_detail}")
+  fi
+
+  local guard_rc=0
+  "${py_bin}" "${guards_py}" "${guard_args[@]}" </dev/null >> "${out_dir}/guards.log" || guard_rc=$?
+  if [ "${guard_rc}" -ne 0 ]; then
+    if [ "${err_tok}" != "NONE" ] && [ -n "${err_tok}" ]; then
+      echo "REFUSED: COLAB_USAGE_PARSE_FAILED" >&2
+    fi
+    return 1
+  fi
+
+  local ram_args=(
+    "--check-ram"
+    "--current-anon-mb" "${CURRENT_ANON_MB:-${anon_mb:-512}}"
+    "--projected-delta-mb" "${PROJECTED_DELTA_MB:-2048}"
+    "--mem-total-mb" "${MEM_TOTAL_MB:-${mem_total_mb:-12288}}"
+  )
+  if [ -n "${cgroup_max_mb:-}" ]; then
+    ram_args+=("--cgroup-max-mb" "${cgroup_max_mb}")
+  fi
+  local ram_rc=0
+  "${py_bin}" "${guards_py}" "${ram_args[@]}" </dev/null >> "${out_dir}/guards.log" || ram_rc=$?
+  if [ "${ram_rc}" -ne 0 ]; then
+    return "${ram_rc}"
+  fi
+}
+
+colab_run_exec_with_live_guards() {
+  local py_bin="$1"
+  local guards_py="$2"
+  local session="$3"
+  local runner_py="$4"
+  local start_epoch="$5"
+  local wall_s="$6"
+  local grace_s="$7"
+  local cap="$8"
+  local margin="$9"
+  local out_dir="${10}"
+  shift 10
+  local env_pairs=("$@")
+
+  local poll_s="${GUARD_POLL_INTERVAL_S:-2}"
+  local stall_s="${STALL_S:-300}"
+  local setup_stall_s="${SETUP_STALL_S:-600}"
+  local last_size=0
+  local last_progress_epoch
+  last_progress_epoch="$(date -u +%s)"
+  local last_setup_bytes=0
+  local last_setup_progress_epoch="${last_progress_epoch}"
+
+  local wall_int="${wall_s%.*}"
+  local grace_int="${grace_s%.*}"
+  local exec_timeout="$(( wall_int + grace_int + 60 ))"
+
+  local exec_args=("-s" "${session}" "-f" "${runner_py}" "--timeout" "${exec_timeout}")
+  for kv in "${env_pairs[@]}"; do
+    exec_args+=("--env" "${kv}")
+  done
+
+  colab exec "${exec_args[@]}" </dev/null > "${out_dir}/vitals.log" 2>&1 &
+  EXEC_PID="$!"
+
+  while kill -0 "${EXEC_PID}" 2>/dev/null; do
+    sleep "${poll_s}"
+    if ! kill -0 "${EXEC_PID}" 2>/dev/null; then
+      break
+    fi
+    local now_epoch elapsed_s cur_size idle_s cur_setup_bytes setup_idle_s run_only_bytes
+    now_epoch="$(date -u +%s)"
+    elapsed_s="$((now_epoch - start_epoch))"
+
+    local check_rc=0
+    colab_check_cap_and_ram "${py_bin}" "${guards_py}" "live" "${elapsed_s}" "${wall_s}" "${grace_s}" "${cap}" "${margin}" "${out_dir}" || check_rc=$?
+    if [ "${check_rc}" -ne 0 ]; then
+      pkill -P "${EXEC_PID}" 2>/dev/null || true
+      kill -9 "${EXEC_PID}" 2>/dev/null || true
+      wait "${EXEC_PID}" 2>/dev/null || true
+      EXEC_PID=""
+      return 91
+    fi
+
+    run_only_bytes="$(grep -Ev '^(===[[:space:]]*VITALS|CGROUP_MEM(_CURRENT|_MAX)?:|MEM:|SETUP_BYTES:|PROC:|HARNESS:|[[:space:]]*[0-9]+[[:space:]]+.*(comfy|main\.py|ai-toolkit|run\.py|llama-server|ollama|lane1_render\.py))' "${out_dir}/vitals.log" 2>/dev/null | wc -c | tr -d '[:space:]' || echo 0)"
+    cur_size="${run_only_bytes:-0}"
+    if [ "${cur_size:-0}" != "${last_size}" ]; then
+      last_size="${cur_size:-0}"
+      last_progress_epoch="${now_epoch}"
+    fi
+    idle_s="$((now_epoch - last_progress_epoch))"
+
+    cur_setup_bytes="$(grep -E '^SETUP_BYTES:[[:space:]]*[0-9]+[[:space:]]*$' "${out_dir}/vitals.log" 2>/dev/null | tail -n 1 | awk -F: '{gsub(/[[:space:]]/, "", $2); print $2}' || true)"
+    if [ -n "${cur_setup_bytes}" ] && [ "${cur_setup_bytes}" != "${last_setup_bytes}" ]; then
+      last_setup_bytes="${cur_setup_bytes}"
+      last_setup_progress_epoch="${now_epoch}"
+    fi
+    setup_idle_s="$((now_epoch - last_setup_progress_epoch))"
+
+    local vitals_rc=0
+    "${py_bin}" "${guards_py}" \
+      --vitals-file "${out_dir}/vitals.log" \
+      --idle-seconds "${idle_s}" \
+      --setup-idle-seconds "${setup_idle_s}" \
+      --stall-s "${stall_s}" \
+      --setup-stall-s "${setup_stall_s}" </dev/null >> "${out_dir}/guards.log" || vitals_rc=$?
+    if [ "${vitals_rc}" -ne 0 ]; then
+      pkill -P "${EXEC_PID}" 2>/dev/null || true
+      kill -9 "${EXEC_PID}" 2>/dev/null || true
+      wait "${EXEC_PID}" 2>/dev/null || true
+      EXEC_PID=""
+      return 93
+    fi
+  done
+
+  local exec_rc=0
+  wait "${EXEC_PID}" || exec_rc=$?
+  EXEC_PID=""
+
+  colab_scrub_all_logs "${py_bin}" "${guards_py}" "${out_dir}"
+
+  local marker_line=""
+  marker_line="$(grep -E '^SDLAB_REMOTE_RC=-?[0-9]+$' "${out_dir}/vitals.log" 2>/dev/null | tail -n 1 | tr -d '\r' || true)"
+  if [ -z "${marker_line}" ]; then
+    echo "REFUSED: SDLAB_REMOTE_RC marker missing from vitals.log (colab exec_rc=${exec_rc})" >> "${out_dir}/guards.log"
+    if [ "${exec_rc}" -ne 0 ]; then
+      return "${exec_rc}"
+    fi
+    return 94
+  fi
+  local remote_rc="${marker_line#SDLAB_REMOTE_RC=}"
+  if [ "${remote_rc}" -ne 0 ]; then
+    echo "REFUSED: remote runner reported SDLAB_REMOTE_RC=${remote_rc}" >> "${out_dir}/guards.log"
+    return "${remote_rc}"
+  fi
+
+  "${py_bin}" "${guards_py}" --vitals-file "${out_dir}/vitals.log" --idle-seconds 0 --setup-idle-seconds 0 --stall-s "${stall_s}" --setup-stall-s "${setup_stall_s}" </dev/null >> "${out_dir}/guards.log"
+  local end_epoch elapsed_final
+  end_epoch="$(date -u +%s)"
+  elapsed_final="$((end_epoch - start_epoch))"
+  colab_check_cap_and_ram "${py_bin}" "${guards_py}" "live" "${elapsed_final}" "${wall_s}" "${grace_s}" "${cap}" "${margin}" "${out_dir}"
+}
+
+colab_stop_and_log_closing() {
+  local py_bin="$1"
+  local session="$2"
+  local out_dir="$3"
+  colab stop -s "${session}" </dev/null >> "${out_dir}/colab.log" 2>&1 || true
+  colab sessions </dev/null >> "${out_dir}/colab.log" 2>&1 || true
+  colab_poll_usage_and_ram "${py_bin}" "${out_dir}" >/dev/null 2>&1 || true
+}
+
+colab_open_session_locked() {
+  local session="$1"
+  local card="$2"
+  local out_dir="$3"
+  local lock_file="${COLAB_OPEN_LOCK:-/tmp/colab-window-open.lock}"
+  mkdir -p "$(dirname "${lock_file}")"
+
+  local new_args=()
+  if [ "${card}" = "CPU-high-mem" ] || [ "${card}" = "CPU" ]; then
+    new_args=("new" "--high-mem" "-s" "${session}")
+  else
+    new_args=("new" "--gpu" "${card}" "--high-mem" "-s" "${session}")
+  fi
+
+  if command -v flock >/dev/null 2>&1; then
+    (
+      flock -x 9
+      printf '1\n' > "${lock_file}.held"
+      colab "${new_args[@]}" </dev/null >> "${out_dir}/colab.log" 2>&1
+      rm -f "${lock_file}.held"
+    ) 9>"${lock_file}"
+  else
+    local lock_dir="${lock_file}.lockdir"
+    while ! mkdir "${lock_dir}" 2>/dev/null; do
+      sleep 0.1
+    done
+    printf '1\n' > "${lock_file}.held"
+    local rc=0
+    colab "${new_args[@]}" </dev/null >> "${out_dir}/colab.log" 2>&1 || rc=$?
+    rm -f "${lock_file}.held"
+    rmdir "${lock_dir}" 2>/dev/null || true
+    if [ "${rc}" -ne 0 ]; then
+      return "${rc}"
+    fi
+  fi
+  printf '1\n' > "${out_dir}/session_opened.txt"
+}
+
+colab_scrub_all_logs() {
+  local py_bin="$1"
+  local guards_py="$2"
+  local out_dir="$3"
+  if [ ! -d "${out_dir}" ]; then
+    return 0
+  fi
+  local scrub_args=()
+  while IFS= read -r -d '' f; do
+    scrub_args+=("--scrub-file" "${f}")
+  done < <(find "${out_dir}" -type f \( -name "*.log" -o -name "*.json" -o -name "*.txt" -o -name "*.sh" \) -print0 2>/dev/null || true)
+  if [ "${#scrub_args[@]}" -gt 0 ]; then
+    "${py_bin}" "${guards_py}" "${scrub_args[@]}" </dev/null || true
+  fi
+}
+
+colab_write_window_and_outcome() {
+  local py_bin="$1"
+  local out_dir="$2"
+  local lane="$3"
+  local session="$4"
+  local card="$5"
+  local cap="$6"
+  local wall_s="$7"
+  local grace_s="$8"
+  local margin="$9"
+  local opened_at="${10}"
+  local closed_at="${11}"
+  local exit_code="${12}"
+  local shas_json="${13}"
+
+  "${py_bin}" - "${out_dir}" "${lane}" "${session}" "${card}" "${cap}" "${wall_s}" "${grace_s}" "${margin}" "${opened_at}" "${closed_at}" "${exit_code}" "${shas_json}" << 'PYEOF'
+import hashlib
+import json
+import sys
+import tarfile
+from pathlib import Path
+
+out_dir = Path(sys.argv[1])
+lane = sys.argv[2]
+session = sys.argv[3]
+card = sys.argv[4]
+cap = float(sys.argv[5])
+wall_s = int(float(sys.argv[6]))
+grace_s = int(float(sys.argv[7]))
+margin = float(sys.argv[8])
+opened_at = sys.argv[9]
+closed_at = sys.argv[10]
+exit_code = int(sys.argv[11])
+shas_raw = json.loads(sys.argv[12])
+
+out_dir.mkdir(parents=True, exist_ok=True)
+
+
+def sha256_path(p: Path) -> str:
+    h = hashlib.sha256()
+    with p.open("rb") as f:
+        for chunk in iter(lambda: f.read(1 << 20), b""):
+            h.update(chunk)
+    return "sha256:" + h.hexdigest()
+
+
+def read_float_file(p: Path) -> float | None:
+    if not p.exists():
+        return None
+    try:
+        return float(p.read_text(encoding="utf-8").strip())
+    except Exception:
+        return None
+
+
+def read_text_file(p: Path) -> str | None:
+    if not p.exists():
+        return None
+    val = p.read_text(encoding="utf-8", errors="replace").strip()
+    return val if val else None
+
+
+vm_facts: dict = {}
+vm_win_file = out_dir / "vm_window.json"
+if vm_win_file.exists():
+    try:
+        vm_facts.update(json.loads(vm_win_file.read_text(encoding="utf-8")))
+    except Exception:
+        pass
+
+for tar_p in sorted(out_dir.glob("*.tar.gz")):
+    if tarfile.is_tarfile(tar_p):
+        try:
+            with tarfile.open(tar_p, "r:gz") as tf:
+                for member in tf.getmembers():
+                    if Path(member.name).name == "vm_window.json":
+                        ext = tf.extractfile(member)
+                        if ext is not None:
+                            vm_facts.update(json.loads(ext.read().decode("utf-8")))
+        except Exception:
+            pass
+
+vitals_p = out_dir / "vitals.log"
+if vitals_p.exists():
+    for line in vitals_p.read_text(encoding="utf-8", errors="replace").splitlines():
+        if line.startswith("WINDOW_JSON "):
+            try:
+                vm_facts.update(json.loads(line[len("WINDOW_JSON ") :].strip()))
+            except Exception:
+                pass
+
+downloaded_files: dict[str, str] = {}
+for artifact in sorted(list(out_dir.glob("*.tar.gz")) + list(out_dir.glob("*.safetensors"))):
+    if artifact.is_file():
+        downloaded_files[artifact.name] = sha256_path(artifact)
+
+usage_parse_error = read_text_file(out_dir / "usage_parse_error.txt")
+rate_measured_pre_open = read_float_file(out_dir / "rate_measured_pre_open.txt")
+rate_measured_post_open = read_float_file(out_dir / "rate_measured_post_open.txt")
+rate_measured = (
+    rate_measured_post_open
+    if rate_measured_post_open is not None
+    else read_float_file(out_dir / "rate_measured.txt")
+)
+rate_for_projection = read_float_file(out_dir / "rate_for_projection.txt")
+rate_used_for_cap_and_projection = rate_for_projection
+rate_source = read_text_file(out_dir / "rate_source.txt") or (
+    "post_open_measured"
+    if (rate_measured_post_open is not None and rate_measured_post_open > 0.0)
+    else (
+        "configured_pending_post_open_rate"
+        if (rate_measured_post_open is not None and rate_measured_post_open == 0.0)
+        else (
+            "max_of_pre_open_and_configured"
+            if rate_measured_pre_open is not None
+            else "configured"
+        )
+    )
+)
+
+window_payload = {
+    **vm_facts,
+    "window_id": session,
+    "lane": lane,
+    "session": session,
+    "card": card,
+    "cap": cap,
+    "wall_s": wall_s,
+    "grace_s": grace_s,
+    "margin": margin,
+    "start_balance": read_float_file(out_dir / "start_balance.txt"),
+    "end_balance": read_float_file(out_dir / "end_balance.txt"),
+    "rate_measured": rate_measured,
+    "rate_measured_pre_open": rate_measured_pre_open,
+    "rate_measured_post_open": rate_measured_post_open,
+    "rate_used_for_cap_and_projection": rate_used_for_cap_and_projection,
+    "rate_for_projection": rate_for_projection,
+    "rate_source": rate_source,
+    "usage_parse_error": usage_parse_error,
+    "input_shas": shas_raw.get("verified_shas", shas_raw),
+    "downloaded_files": downloaded_files,
+    "opened_at": opened_at,
+    "closed_at": closed_at,
+    "exit_code": exit_code,
+}
+(out_dir / "window.json").write_text(json.dumps(window_payload, indent=2) + "\n", encoding="utf-8", newline="\n")
+
+outcome_payload = {
+    "window_id": session,
+    "lane": lane,
+    "session": session,
+    "card": card,
+    "status": "ok" if exit_code == 0 else "failed",
+    "exit_code": exit_code,
+    "opened_at": opened_at,
+    "closed_at": closed_at,
+}
+(out_dir / "outcome.draft.json").write_text(json.dumps(outcome_payload, indent=2) + "\n", encoding="utf-8", newline="\n")
+PYEOF
+}