start.sh: gate 5 heavy harvest launchers behind LOW_MEM

CPU-Basic Space cap is 16 GB. With uvicorn + 5 nohup'd bash launchers
(dedup-bootstrap, dataset-enrich, kaggle-trainer, lightning-trainer,
dataset-mirror) running concurrently while harvest workers hammered
/cursor/* at ~100 req/min, the container blew through 16 GB roughly
5 min after every boot and HF auto-killed it. The watchdog correctly
flagged 'HTTP layer hung' but the actual signal was OOM_KILL.

Fix: gate the 5 launchers behind LOW_MEM=0. LOW_MEM defaults to 1 (set
in the global env at line ~20 BEFORE any reference, so 'set -u' is
happy) so on CPU-Basic the launchers no-op and the Space stays
under ~2 GB resident — just FastAPI + Redis + status server.

Harvest pipeline: the 3 recurring launchers (dataset-enrich every 30m,
dataset-mirror every 6h, lightning-trainer daily 02:00 UTC) move to
GCP scheduler via hermes-jobs.json entries. dedup-bootstrap is
one-shot so doesn't need a cron. kaggle-trainer is launched manually
from Mac for V19 runs.

Re-enable in-Space launchers by setting LOW_MEM=0 (Space variable)
once we upgrade to a paid tier (cpu-upgrade ≥ 32 GB) or migrate to a
larger anchor like OCI A1.Flex 24 GB ARM.

start.sh

@@ -15,6 +15,13 @@ set -x
 # Echo stdout so HF run-logs see progress (safe steps before .env is loaded)
 exec > >(tee -a "$LOG_DIR/boot.log") 2>&1
 
+# ── Memory mode (must be set BEFORE any reference; we use `set -u`) ───────
+# CPU-Basic Space = 16 GB cap. With LOW_MEM=1 we skip the heavy harvest
+# launchers (dataset-enrich, dataset-mirror, kaggle-trainer, lightning-trainer,
+# dedup-bootstrap) — those run on GCP daemons instead. Set LOW_MEM=0 only
+# on a paid Space tier (cpu-upgrade ≥32 GB).
+LOW_MEM="${LOW_MEM:-1}"
+
 # ── 1. Persistent data — symlink state subdirs to /data (HF persistent mount) ──
 # bin/ is NOT persisted (baked into image, refreshed on every push).
 # Persisted: state (DBs), logs, memory, skills, sessions, training pairs,
@@ -86,34 +93,38 @@ if [[ -d "$DATA" ]] && [[ -w "$DATA" ]]; then
         fi
     fi
 
-    # ── One-time central dedup bootstrap from existing data ──────────────
-    if [[ ! -f "${HOME}/.surrogate/.dedup-bootstrap-done" ]]; then
-        echo "[$(date +%H:%M:%S)] running central dedup bootstrap (one-time)" >> "$LOG_DIR/boot.log"
-        nohup bash "${HOME}/.surrogate/bin/dedup-bootstrap.sh" > "$LOG_DIR/dedup-bootstrap.log" 2>&1 &
-    fi
+    # ── Heavy harvest launchers — only on HIGH_MEM (LOW_MEM=0) ───────────
+    # On CPU-Basic (16 GB cap) launching 5 background bash + uvicorn + 5 harvest
+    # workers blew through the cap and HF auto-killed the container ~5 min after
+    # boot. These launchers are now scheduled on GCP via hermes-scheduler-daemon
+    # (entries in data/hermes-jobs.json) so harvest still runs — just not from
+    # inside the Space's RAM. Re-enable in-Space by setting LOW_MEM=0 once we
+    # upgrade to a ≥32 GB tier.
+    if [[ "$LOW_MEM" != "1" ]]; then
+        # ── One-time central dedup bootstrap from existing data ──────────
+        if [[ ! -f "${HOME}/.surrogate/.dedup-bootstrap-done" ]]; then
+            echo "[$(date +%H:%M:%S)] running central dedup bootstrap (one-time)" >> "$LOG_DIR/boot.log"
+            nohup bash "${HOME}/.surrogate/bin/dedup-bootstrap.sh" > "$LOG_DIR/dedup-bootstrap.log" 2>&1 &
+        fi
+
+        # ── BOOT-TIME enrich kickoff (trigger immediate pull, don't wait for cron)
+        nohup bash "${HOME}/.surrogate/bin/dataset-enrich.sh" >> "$LOG_DIR/dataset-enrich.log" 2>&1 &
+        echo "[$(date +%H:%M:%S)] boot-time dataset-enrich kicked off" >> "$LOG_DIR/boot.log"
 
-    # ── BOOT-TIME enrich kickoff (trigger immediate pull, don't wait for cron) ──
-    # User feedback: 'ตั้งแต่ 7 โมงเช้า แต่ data ไม่เพิ่ม' — cron M%60 may have
-    # been mis-aligned with rebuilds. Force one enrich run on every boot.
-    nohup bash "${HOME}/.surrogate/bin/dataset-enrich.sh" >> "$LOG_DIR/dataset-enrich.log" 2>&1 &
-    echo "[$(date +%H:%M:%S)] boot-time dataset-enrich kicked off" >> "$LOG_DIR/boot.log"
-
-    # ── BOOT-TIME kaggle-trainer kickoff (don't wait for 90-min cron) ─────
-    # Submits the LoRA training notebook to Kaggle T4 immediately on every
-    # Space boot. Kaggle CLI gracefully handles 'already running' state.
-    nohup bash "${HOME}/.surrogate/bin/kaggle-trainer.sh" >> "$LOG_DIR/kaggle-trainer.log" 2>&1 &
-    echo "[$(date +%H:%M:%S)] boot-time kaggle-trainer kicked off" >> "$LOG_DIR/boot.log"
-
-    # ── BOOT-TIME lightning-trainer kickoff — H200 4 hr free for big model ─
-    nohup bash "${HOME}/.surrogate/bin/lightning-trainer.sh" >> "$LOG_DIR/lightning-trainer.log" 2>&1 &
-    echo "[$(date +%H:%M:%S)] boot-time lightning-trainer kicked off (H200 4hr quota)" >> "$LOG_DIR/boot.log"
-
-    # ── BOOT-TIME dataset-mirror — bulk-clone top community SFT mixes ──────
-    # Far faster than streaming-and-normalize — 1 commit per parquet file
-    # = millions of pairs landing as raw mirrors/<slug>/<file>. Idempotent
-    # via stamp file so we don't re-mirror what's already been pulled.
-    nohup bash "${HOME}/.surrogate/bin/dataset-mirror.sh" >> "$LOG_DIR/dataset-mirror.log" 2>&1 &
-    echo "[$(date +%H:%M:%S)] boot-time dataset-mirror kicked off (30 community sources)" >> "$LOG_DIR/boot.log"
+        # ── BOOT-TIME kaggle-trainer kickoff (don't wait for 90-min cron) ─
+        nohup bash "${HOME}/.surrogate/bin/kaggle-trainer.sh" >> "$LOG_DIR/kaggle-trainer.log" 2>&1 &
+        echo "[$(date +%H:%M:%S)] boot-time kaggle-trainer kicked off" >> "$LOG_DIR/boot.log"
+
+        # ── BOOT-TIME lightning-trainer kickoff — H200 4 hr free for big model
+        nohup bash "${HOME}/.surrogate/bin/lightning-trainer.sh" >> "$LOG_DIR/lightning-trainer.log" 2>&1 &
+        echo "[$(date +%H:%M:%S)] boot-time lightning-trainer kicked off (H200 4hr quota)" >> "$LOG_DIR/boot.log"
+
+        # ── BOOT-TIME dataset-mirror — bulk-clone top community SFT mixes ─
+        nohup bash "${HOME}/.surrogate/bin/dataset-mirror.sh" >> "$LOG_DIR/dataset-mirror.log" 2>&1 &
+        echo "[$(date +%H:%M:%S)] boot-time dataset-mirror kicked off (30 community sources)" >> "$LOG_DIR/boot.log"
+    else
+        echo "[$(date +%H:%M:%S)] LOW_MEM=1 → skipped 5 heavy harvest launchers (delegated to GCP daemons)" >> "$LOG_DIR/boot.log"
+    fi
 
     echo "[$(date +%H:%M:%S)] persistent /data linked (state, logs, memory, skills, sessions, workspace, ollama, training-pairs)" >> "$LOG_DIR/boot.log"
 else