v3_scale: bg1d lancé (PID 838852) — tranches par segment (cm_bg1d_m8.cov.json) + compteur lourds par bot (heavy_checker.sh) + gardes bg1d/div1d

v3_scale/scale/cm_bg1d_m8.cov.json

@@ -0,0 +1,615 @@
+{
+ "N": 78364074,
+ "segments": [
+  {
+   "name": "broad_local_v1",
+   "byte_start": 0,
+   "first_patch": 0,
+   "n_patches": 8680605
+  },
+  {
+   "name": "gutenberg_v2",
+   "byte_start": 34835259,
+   "first_patch": 8680605,
+   "n_patches": 69683469
+  }
+ ],
+ "snn": {
+  "n_slices": 64,
+  "patches": 600000,
+  "patches_by_segment": {
+   "broad_local_v1": 65625,
+   "gutenberg_v2": 534375
+  },
+  "slices_by_segment": {
+   "broad_local_v1": 7,
+   "gutenberg_v2": 57,
+   "CHEVAUCHE": 0
+  },
+  "slices_md5": "bab55e5d9d44fea499e9049b52c5af99",
+  "byte_offsets_md5": "3a50468e31f7cab3d9b0d4abceed506a",
+  "bytes_covered": 2211450,
+  "first_last_bytes": [
+   [
+    4661151,
+    4703859
+   ],
+   [
+    290179586,
+    290210348
+   ]
+  ]
+ },
+ "am": {
+  "n_slices": 64,
+  "patches": 1200000,
+  "patches_by_segment": {
+   "broad_local_v1": 131250,
+   "gutenberg_v2": 1068750
+  },
+  "slices_by_segment": {
+   "broad_local_v1": 7,
+   "gutenberg_v2": 57,
+   "CHEVAUCHE": 0
+  },
+  "slices_md5": "7276f94960d7bb200b14230955f2c07d",
+  "byte_offsets_md5": "26cf65b8a25a17431d2d6571565c9c7b",
+  "bytes_covered": 4526250,
+  "first_last_bytes": [
+   [
+    2570891,
+    2655272
+   ],
+   [
+    288674578,
+    288735559
+   ]
+  ]
+ },
+ "snn_am_disjoint": true,
+ "bench_seen": {
+  "slice": 32,
+  "patch_range": [
+   40215598,
+   40224973
+  ],
+  "n_patches": 9375,
+  "bytes": [
+   153380822,
+   153414962
+  ],
+  "n_bytes": 34140,
+  "segment": "gutenberg_v2",
+  "outside_ept_fit_calib_0_2400000": true,
+  "outside_fusion_calib_0_60000": true,
+  "in_am_slices": false,
+  "ids_match_build": true,
+  "sha256": "41ff26b784f2614b4789504ab000571bae282223234312210e66096097ac24ff"
+ },
+ "snn_byte_offsets": [
+  [
+   4661151,
+   4703859
+  ],
+  [
+   10199896,
+   10242465
+  ],
+  [
+   15735566,
+   15778296
+  ],
+  [
+   20811473,
+   20841329
+  ],
+  [
+   25042673,
+   25079787
+  ],
+  [
+   29553342,
+   29583904
+  ],
+  [
+   33792551,
+   33820483
+  ],
+  [
+   38479114,
+   38516880
+  ],
+  [
+   42846511,
+   42885821
+  ],
+  [
+   47507622,
+   47539716
+  ],
+  [
+   52020526,
+   52055819
+  ],
+  [
+   56629257,
+   56651883
+  ],
+  [
+   61132303,
+   61162345
+  ],
+  [
+   65491739,
+   65529941
+  ],
+  [
+   70189736,
+   70228207
+  ],
+  [
+   74904914,
+   74942555
+  ],
+  [
+   78726022,
+   78747244
+  ],
+  [
+   83041979,
+   83077029
+  ],
+  [
+   87649342,
+   87684434
+  ],
+  [
+   92506224,
+   92544180
+  ],
+  [
+   95860260,
+   95893743
+  ],
+  [
+   100786039,
+   100824734
+  ],
+  [
+   105671517,
+   105709584
+  ],
+  [
+   110619467,
+   110658011
+  ],
+  [
+   115564726,
+   115602877
+  ],
+  [
+   120481141,
+   120518641
+  ],
+  [
+   125417740,
+   125455012
+  ],
+  [
+   129903316,
+   129933016
+  ],
+  [
+   134304736,
+   134340886
+  ],
+  [
+   138821168,
+   138855728
+  ],
+  [
+   143697600,
+   143731086
+  ],
+  [
+   148465751,
+   148501464
+  ],
+  [
+   153380822,
+   153414965
+  ],
+  [
+   157768698,
+   157805395
+  ],
+  [
+   162615026,
+   162652898
+  ],
+  [
+   167575577,
+   167612934
+  ],
+  [
+   172372376,
+   172411090
+  ],
+  [
+   177221143,
+   177259522
+  ],
+  [
+   182207525,
+   182245607
+  ],
+  [
+   187167439,
+   187204964
+  ],
+  [
+   192068868,
+   192107179
+  ],
+  [
+   196975361,
+   197013957
+  ],
+  [
+   201929058,
+   201967540
+  ],
+  [
+   206817086,
+   206854830
+  ],
+  [
+   211753264,
+   211790770
+  ],
+  [
+   216601348,
+   216637899
+  ],
+  [
+   221434517,
+   221466057
+  ],
+  [
+   225475136,
+   225506493
+  ],
+  [
+   229505602,
+   229536375
+  ],
+  [
+   233544279,
+   233575369
+  ],
+  [
+   237596238,
+   237627530
+  ],
+  [
+   241659330,
+   241690495
+  ],
+  [
+   245737072,
+   245767749
+  ],
+  [
+   249778454,
+   249809846
+  ],
+  [
+   253832910,
+   253863885
+  ],
+  [
+   257911349,
+   257943813
+  ],
+  [
+   261994595,
+   262025821
+  ],
+  [
+   266011266,
+   266042600
+  ],
+  [
+   270054196,
+   270085125
+  ],
+  [
+   274077252,
+   274107970
+  ],
+  [
+   278084137,
+   278114541
+  ],
+  [
+   282139067,
+   282169557
+  ],
+  [
+   286168424,
+   286199770
+  ],
+  [
+   290179586,
+   290210348
+  ]
+ ],
+ "am_byte_offsets": [
+  [
+   2570891,
+   2655272
+  ],
+  [
+   8107954,
+   8193378
+  ],
+  [
+   13642088,
+   13725858
+  ],
+  [
+   19189010,
+   19274157
+  ],
+  [
+   23341427,
+   23415764
+  ],
+  [
+   27865007,
+   27924864
+  ],
+  [
+   32255536,
+   32330382
+  ],
+  [
+   36655964,
+   36729895
+  ],
+  [
+   41110314,
+   41169156
+  ],
+  [
+   45876236,
+   45948735
+  ],
+  [
+   50243341,
+   50315051
+  ],
+  [
+   54911769,
+   54985092
+  ],
+  [
+   59467267,
+   59543010
+  ],
+  [
+   63716962,
+   63792912
+  ],
+  [
+   68328640,
+   68405486
+  ],
+  [
+   73052407,
+   73127415
+  ],
+  [
+   77491885,
+   77552374
+  ],
+  [
+   81191902,
+   81268618
+  ],
+  [
+   85908626,
+   85979158
+  ],
+  [
+   90642726,
+   90718027
+  ],
+  [
+   94749191,
+   94791298
+  ],
+  [
+   98903725,
+   98979507
+  ],
+  [
+   103816924,
+   103892687
+  ],
+  [
+   108746865,
+   108822448
+  ],
+  [
+   113687624,
+   113764609
+  ],
+  [
+   118620827,
+   118697357
+  ],
+  [
+   123559371,
+   123635577
+  ],
+  [
+   128420466,
+   128480598
+  ],
+  [
+   132571234,
+   132643498
+  ],
+  [
+   137083957,
+   137153092
+  ],
+  [
+   141843729,
+   141919608
+  ],
+  [
+   146584615,
+   146661191
+  ],
+  [
+   151521829,
+   151598991
+  ],
+  [
+   156078149,
+   156147060
+  ],
+  [
+   160755919,
+   160831418
+  ],
+  [
+   165705158,
+   165783081
+  ],
+  [
+   170644392,
+   170721731
+  ],
+  [
+   175378152,
+   175452879
+  ],
+  [
+   180319501,
+   180395370
+  ],
+  [
+   185284699,
+   185361295
+  ],
+  [
+   190181814,
+   190257993
+  ],
+  [
+   195116875,
+   195194373
+  ],
+  [
+   200051726,
+   200127725
+  ],
+  [
+   204964345,
+   205038883
+  ],
+  [
+   209896285,
+   209971124
+  ],
+  [
+   214807852,
+   214882337
+  ],
+  [
+   219621117,
+   219696715
+  ],
+  [
+   223969391,
+   224030774
+  ],
+  [
+   227994712,
+   228056625
+  ],
+  [
+   232016673,
+   232079200
+  ],
+  [
+   236056650,
+   236119008
+  ],
+  [
+   240119667,
+   240182560
+  ],
+  [
+   244195500,
+   244259073
+  ],
+  [
+   248248776,
+   248310002
+  ],
+  [
+   252308548,
+   252369966
+  ],
+  [
+   256323790,
+   256388835
+  ],
+  [
+   260445249,
+   260508269
+  ],
+  [
+   264495679,
+   264556066
+  ],
+  [
+   268538320,
+   268600619
+  ],
+  [
+   272559063,
+   272620479
+  ],
+  [
+   276575914,
+   276637453
+  ],
+  [
+   280604816,
+   280666480
+  ],
+  [
+   284663887,
+   284725739
+  ],
+  [
+   288674578,
+   288735559
+  ]
+ ],
+ "cov": {
+  "arm": "a",
+  "N": 78364074,
+  "seed": 0,
+  "k": 64
+ },
+ "snn_tokens": 600000,
+ "am_tokens": 1200000,
+ "train_bin": "/workspace/spira_shared/broad/broad_gutenberg_v1/train.bin",
+ "train_bytes": 290810502
+}

v3_scale/scale/eval_bench_v4.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+"""Checker — banc scale v4. Usage : eval_bench_v4.py CKPT CORPUS_NAME TAG   env MAXPOS (cap/fenêtre, smoke) WINDOWS HITS OUT
+Fenêtres : seen (corpus propre) + fc2, fresh_en, fresh_fr, neither_code, test20k, queries, test_v3 (octets fixes, identiques pour tous).
+Boucle par position = eval_hits() de checker/eval_ref_v3.py (copie ligne à ligne de SPIRA.eval_stream), importée telle quelle.
+Enregistre wall-time et pic RSS (ru_maxrss) par fenêtre."""
+import os, sys, json, time, resource, hashlib
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+import numpy as np
+from bench_common import *
+ckpt, cname, tag = sys.argv[1:4]; D = os.path.dirname(os.path.abspath(__file__))
+HITS = os.environ.get("HITS", f"{D}/hits_v4"); OUT = os.environ.get("OUT", f"{D}/results_v4.jsonl"); MAXPOS = int(os.environ.get("MAXPOS", "0")); SEL = os.environ.get("WINDOWS")
+src = open(f"{S}/checker/eval_ref_v3.py").read(); g = {"np": np, "fnv1a64": fnv1a64}
+exec(src[src.index("def eval_hits"):src.index("m = SPIRA.load(ckpt)")], g); eval_hits = g["eval_hits"]
+t0 = time.perf_counter(); train = rd(f"{CORPORA[cname]}/train.bin")
+m, kind, saved, cal = load_model(ckpt, train); tload = time.perf_counter() - t0
+SEEN_DEF = "train[2600000:2800000]"; SEEN = train[2_600_000:2_800_000]  # Checker 23 h 50 : seen par défaut (inchangé pour les modèles préfixe)
+CJ = ckpt.replace(".npz", ".cov.json")  # bg1d/div1d (tranches stratifiées) : seen = octets d'une tranche SNN RÉELLEMENT entraînée (cov_bytes_m8.slice_report)
+if os.path.exists(CJ):
+    bs = json.load(open(CJ))["bench_seen"]; w0, w1 = bs["bytes"]; SEEN = train[w0:w1]
+    assert hashlib.sha256(SEEN).hexdigest() == bs["sha256"] and bs["n_patches"] > 8000 and bs["outside_ept_fit_calib_0_2400000"], bs
+    SEEN_DEF = f"train[{w0}:{w1}] = tranche SNN {bs['slice']} (patches {bs['patch_range']}, segment {bs['segment']}, hors AM)"
+W = {"seen": (SEEN, 8000), **fixed_windows()}; del train, SEEN
+os.makedirs(HITS, exist_ok=True)
+for name, (data, mp) in W.items():
+    if SEL and name not in SEL.split(","): continue
+    mp = min(mp, MAXPOS) if MAXPOS else mp; t1 = time.perf_counter(); r, H = eval_hits(m, data, max_positions=mp)
+    pl = np.fromiter((len(p) for p in m.ept.segment(data)), dtype=np.int64); off = np.concatenate([[0], np.cumsum(pl)])
+    n = len(H["fused_hit"]); bb = hashlib.sha256(data).hexdigest()  # v4.1 : offsets octets de la cible t+1 → appariement McNemar par octet entre tokenizers différents
+    np.savez_compressed(f"{HITS}/{tag}__{name}.npz", **H, tgt_start=off[1:n + 1], tgt_len=pl[1:n + 1], window_sha256=np.array(bb), window_bytes=np.array(len(data)), ept_budget_bits=np.array(float(m.ept.budget_bits)))
+    rec = {"tag": tag, "corpus": cname, "ckpt": ckpt, "kind": kind, "window": name, "n_positions": r["n_positions"],
+           "fused_invocab_pct": round(100 * r["fused_acc_invocab"], 3), "snn_invocab_pct": round(100 * r["snn_acc_invocab"], 3),
+           "vocab_coverage": r["vocab_coverage"], "fusion_saved": saved, "fusion_cal": cal, "load_s": round(tload, 1),
+           "eval_s": round(time.perf_counter() - t1, 1), "peak_rss_mb": round(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss / 1024, 1),
+           "maxpos": MAXPOS, "window_sha256": bb, "window_bytes": len(data), "ept_budget_bits": float(m.ept.budget_bits), "script": "eval_bench_v4.1", **({"seen_def": SEEN_DEF} if name == "seen" else {})}
+    open(OUT, "a").write(json.dumps(rec) + "\n"); print(name, rec["fused_invocab_pct"], rec["eval_s"], rec["peak_rss_mb"], flush=True)

v3_scale/scale/gate_bg1d_m8.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Checker 23 h 55 (Toronto) — garde M8 broad_gutenberg_v1 bg1d (64 tranches stratifiées sur TOUT le flux ; diag_m8_coverage_2315.md).
+# COPIE de gate_bg1c_m8_v2.sh ; changements : script train_cm_m8_bg1d.py, sortie cm_bg1d_m8.npz, journaux *_bg1d_*, fichier .ok =
+# logs/bg1d_test.ok (test petit cas PASS : tranches == train_cm_m8_full.py, build bit-identique bg1c, part Gutenberg, pic RSS),
+# PID écrit dans logs/bg1d.pid, pas d'ajout au relais au lancement (section relais unique écrite par le Checker).
+# Prioritaire : gate_ok.sh (motif adapté gate_bg1[cd]/train_cm_m8_bg1[cd]) répond NON aux autres tant que cette garde attend.
+# Lance UNE fois quand : .ok présent ET gate_ok.sh --bg1c = OUI (lourds < 4, MemAvailable >= 4 Go) ET MemAvailable >= 4,5 Go ET pas de sortie.
+# ulimit -v (ULIMIT_KB du .ok, 4 Go) : MemoryError propre plutôt qu'un OOM-kill des autres. nice 10, OMP=1, sleep 60. Ne tue rien.
+export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 TZ=America/Toronto
+D=/workspace/spira_shared/checker/scale; PY=/workspace/SPIRA-venv/bin/python; cd $D
+OUT=$D/cm_bg1d_m8.npz; L=$D/logs/train_cm_bg1d_m8.log; OK=$D/logs/bg1d_test.ok
+log(){ echo "$(TZ=America/Toronto date '+%F %H:%M:%S') $*" >> $D/logs/gate_bg1d_m8.log; }
+log "garde bg1d (seuil 4,5 Go) démarrée pid $$"
+while true; do
+  if [ -f "$OUT" ]; then log "sortie $OUT existe déjà : rien à lancer"; exit 0; fi
+  g=$(bash /workspace/spira_shared/checker/gate_ok.sh --bg1c); mem=$(awk '/MemAvailable/{print $2}' /proc/meminfo)
+  if [ -f "$OK" ] && echo "$g" | grep -q '^OUI' && [ "$mem" -ge 4718592 ]; then
+    ULIMIT_KB=$(awk -F= '/^ULIMIT_KB=/{print $2}' "$OK"); [ -n "$ULIMIT_KB" ] || ULIMIT_KB=4194304
+    cd $D/logs; ( ulimit -v $ULIMIT_KB; exec nice -n 10 $PY -u $D/train_cm_m8_bg1d.py concat $OUT ) > $L 2>&1 & TP=$!; cd $D
+    echo $TP > $D/logs/bg1d.pid
+    log "LANCÉ entraînement PID $TP ($g ; mem $((mem/1024)) Mo ; ulimit -v $ULIMIT_KB kB)"
+    wait $TP; rc=$?
+    pk=$(grep -o 'pic_RSS=[0-9.]*Go' $L | tail -1)
+    log "FIN entraînement PID $TP : code $rc, dernier $pk, sortie $( [ -f $OUT ] && echo présente || echo ABSENTE )"
+    exit 0
+  fi
+  log "attente : $g ; mem $((mem/1024)) Mo ; test_ok=$( [ -f $OK ] && echo oui || echo non )"; sleep 60
+done

v3_scale/scale/gate_div1d_m8.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Checker 23 h 40 (Toronto) — garde div1d = COPIE de gate_div1c_m8.sh (garde 725157 arrêtée) ; changements : script train_cm_m8_div1d.py
+#  (64 tranches stratifiées sur TOUT corpus_diverse_v1, test petit cas test_div1d/ PASS), attend bg1d (gate_bg1d_m8.sh / train_cm_m8_bg1d.py /
+#  cm_bg1d_m8.npz) au lieu de bg1c, sortie cm_div1d_m8.npz, journaux *_div1d_*, .ok = logs/div1d_test.ok, tag div1_m8 déjà dans jobs_v4.tsv.
+# --- en-tête div1c ---
+# Checker 22 h 11 (Toronto) — garde M8 corpus_diverse_v1 (concat) div1c, APRÈS M8 bg1c (prioritaire). nice 10, sleep 60. Ne tue rien. Lance UNE fois quand :
+#  1) bg1c terminé AVEC succès : aucune gate_bg1c_m8.sh ni train_cm_m8_bg1c.py vivante ET cm_bg1c_m8.npz présent
+#     (bg1c fini SANS sortie → pas de lancement, journal + relais, exit 1 : décision humaine) ;
+#  2) corpus_diverse_v1/train.bin = sha du manifest ; 3) gate_ok.sh = OUI (lourds < 4, aucune attente prioritaire) ;
+#  4) MemAvailable >= max(6 Go, 2 × pic RSS bg1c) (corpus 530 Mo ≈ 1,8 × bg1c 291 Mo) ; 5) pas de sortie cm_div1c_m8.npz.
+# ulimit -v = 1,8 × ULIMIT_KB de bg1c (heuristique taille corpus, plafonné à MemAvailable − 1 Go) : MemoryError propre plutôt qu'OOM-kill.
+export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1 TZ=America/Toronto
+S=/workspace/spira_shared; D=$S/checker/scale; PY=/workspace/SPIRA-venv/bin/python; cd $D
+OUT=$D/cm_div1d_m8.npz; L=$D/logs/train_cm_div1d_m8.log; BG=$D/cm_bg1d_m8.npz; OK=$D/logs/div1d_test.ok
+SHA=$(python3 -c "import json;print(json.load(open('$S/corpus_diverse_v1/manifest.json'))['files']['train.bin']['sha256'])")
+log(){ echo "$(date '+%F %H:%M:%S') $*" >> $D/logs/gate_div1d_m8.log; }
+log "garde div1d démarrée pid $$ (après bg1d)"
+while true; do
+  if [ -f "$OUT" ]; then log "sortie $OUT existe déjà : rien à lancer"; exit 0; fi
+  nb=$(pgrep -f gate_bg1d_m8.sh | wc -l); nt=$(pgrep -f train_cm_m8_bg1d.py | wc -l)
+  if [ "$nb" -gt 0 ] || [ "$nt" -gt 0 ]; then log "attente bg1d (garde $nb, train $nt)"; sleep 60; continue; fi
+  if [ ! -f "$BG" ]; then
+    if [ -f $D/logs/gate_bg1d_m8.log ] && grep -q "FIN entraînement" $D/logs/gate_bg1d_m8.log; then
+      log "bg1d fini SANS cm_bg1d_m8.npz : div1d NON lancé (exit 1)"
+      nice -n 10 $PY $D/relay_append_v3.py "## Checker — {H} — M8 div1d NON lancé : bg1d fini sans sortie (voir logs/gate_bg1d_m8.log). Décision humaine requise."
+      exit 1
+    fi
+    log "attente : bg1d pas encore lancé/terminé (pas de cm_bg1d_m8.npz)"; sleep 60; continue
+  fi
+  g=$(bash $S/checker/gate_ok.sh); mem=$(awk '/MemAvailable/{print $2}' /proc/meminfo)
+  pk=$(python3 -c "import json;print(json.load(open('${BG%.npz}.summary.json')).get('peak_rss_kB',0))" 2>/dev/null || echo 0)
+  need=$(( 2*pk > 6291456 ? 2*pk : 6291456 ))
+  if [ -f "$OK" ] && echo "$g" | grep -q '^OUI' && [ "$mem" -ge "$need" ]; then
+    cur=$(sha256sum $S/corpus_diverse_v1/train.bin | cut -d' ' -f1)
+    if [ "$cur" != "$SHA" ]; then log "sha corpus_diverse_v1 ≠ manifest ($cur) : NON lancé (exit 1)"; exit 1; fi
+    UB=$(awk -F= '/^ULIMIT_KB=/{print $2}' "$OK" 2>/dev/null); [ -n "$UB" ] || UB=4194304
+    UL=$(( UB*18/10 )); cap=$(( mem - 1048576 )); [ "$UL" -gt "$cap" ] && UL=$cap
+    cd $D/logs; ( ulimit -v $UL; exec nice -n 10 $PY -u $D/train_cm_m8_div1d.py concat $OUT ) > $L 2>&1 & TP=$!; cd $D
+    log "LANCÉ entraînement div1d PID $TP ($g ; mem $((mem/1024)) Mo ; besoin $((need/1024)) Mo ; ulimit -v $UL kB)"
+    echo $TP > $D/logs/div1d.pid
+    nice -n 10 $PY $D/relay_append_v3.py "## Checker — {H} — M8 corpus_diverse_v1 (concat) div1d LANCÉ (après bg1d ; 64 tranches stratifiées)
+- PID $TP (nice 10, OMP_NUM_THREADS=1, ulimit -v $UL kB), sortie checker/scale/cm_div1d_m8.npz, journal checker/scale/logs/train_cm_div1d_m8.log ; job div1_m8 (jobs_v4.tsv) ; part par segment : checker/scale/cm_div1d_m8.cov.json."
+    wait $TP; rc=$?
+    log "FIN entraînement PID $TP : code $rc, dernier $(grep -o 'pic_RSS=[0-9.]*Go' $L | tail -1), sortie $( [ -f $OUT ] && echo présente || echo ABSENTE )"
+    exit 0
+  fi
+  log "attente : $g ; mem $((mem/1024)) Mo ; besoin $((need/1024)) Mo"; sleep 60
+done

v3_scale/scale/gate_ok.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+# gate_ok.sh : imprime OUI/NON. OUI ssi lourds Checker < 4 ET MemAvailable (/proc/meminfo) >= 4 Go ET aucune porte gut1 en attente
+# ET garde M8 broad_gutenberg_v1 bas-mémoire (gate_bg1c_m8.sh) pas en attente (priorité au prochain créneau ;
+# remplace la condition bg1b — 22 h 05 Toronto, bg1b mort OOM 21 h 48). Option --bg1c : appel par gate_bg1c_m8.sh elle-même
+# (ignore sa propre attente bg1c ET l'attente gut1 : priorité à bg1c).
+# Checker 23 h 35 (Toronto) : motif étendu à bg1d (garde gate_bg1d_m8.sh, qui appelle --bg1c) ; sauvegarde scale/logs/gate_ok.sh.bak_2335.
+# Checker 23 h 50 (Toronto) : libellé bg1c_en_attente -> bg1d_en_attente (seule garde bg1 vivante = gate_bg1d_m8.sh ; 756204/756275 arrêtés 23 h 34) ; heavy_checker.sh exclut les process au cwd d'un autre bot ; sauvegarde scale/logs/gate_ok.sh.bak_2350.
+# Checker 23 h 58 (Toronto) : plancher MemAvailable relevé de 4 Gio (4194304 kB) à 4,5 Go décimal (4394532 kB) sur demande Relay 23 h 52 ; sauvegarde scale/logs/gate_ok.sh.bak_2358.
+n=$(bash /workspace/spira_shared/checker/scale/heavy_checker.sh); m=$(awk '/MemAvailable/{print $2}' /proc/meminfo)
+g=$(pgrep -f pipeline_gut1 | wc -l); t=$(pgrep -f "cm_gut1_cov_b" | wc -l)
+w=0; [ "$g" -gt 0 ] && [ "$t" -eq 0 ] && w=1
+b=0; [ $(pgrep -f "gate_bg1[cd]_m8(_v2)?\.sh" | wc -l) -gt 0 ] && [ $(pgrep -f "train_cm_m8_bg1[cd]\.py" | wc -l) -eq 0 ] && b=1  # 23 h 35 : + bg1d (gate_bg1d_m8.sh / train_cm_m8_bg1d.py)
+if [ "$1" = "--bg1c" ]; then w=0; b=0; fi
+if [ "$n" -lt 4 ] && [ "$m" -ge 4394532 ] && [ "$w" -eq 0 ] && [ "$b" -eq 0 ]; then r=OUI; else r=NON; fi
+echo "$r lourds=$n memavail_kB=$m gut1_en_attente=$w bg1d_en_attente=$b"

v3_scale/scale/heavy_checker.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# Lourds Checker (corrigé 21 h 30) : TOUT python SPIRA-venv dont la commande OU le cwd touche un dossier Checker
+# (checker/, spira_m8/, corpus_diverse_v1/, corpus_code_multi_v1/), sauf scripts légers (relay_append, verdict*, test_bitident).
+# Option -v : liste les PID. Sauvegarde avant modif 23 h 50 : scale/logs/heavy_checker.sh.bak_2350.
+n=0; L=""
+for p in $(pgrep -f "^/workspace/SPIRA-venv/bin/python"); do
+  a=$(tr '\0' ' ' < /proc/$p/cmdline 2>/dev/null) || continue; c=$(readlink /proc/$p/cwd 2>/dev/null)
+  echo "$a" | grep -qE "relay_append|verdict|test_bitident" && continue
+  # 23 h 50 (Toronto) : règle = max 4 lourds par bot -> un process dont le cwd est dans le dossier d'un AUTRE bot
+  # (spira_shared/<autre>, /tmp) n'est PAS Checker même s'il appelle un script Checker (ex. 827525 Hashing : eval_bench_v4.py, cwd hashing_gen/m7_v41).
+  case "$c" in /workspace/spira_shared/checker*|/workspace/spira_shared/spira_m8*|/workspace/spira_shared/corpus_diverse_v1*|/workspace/spira_shared/corpus_code_multi_v1*) ;;
+    /tmp*|/workspace/spira_shared/?*) continue ;; esac
+  echo "$a $c" | grep -qE "spira_shared/(checker|spira_m8|corpus_diverse_v1|corpus_code_multi_v1)|checker/|spira_m8/" || continue
+  n=$((n+1)); L="$L $p"
+done
+[ "$1" = -v ] && echo "PIDs:$L"; echo $n