fix: 8 shards (was 4) + round-robin axentx projects + 3min cooldown

USER: 'ทำพร้อมกันทั้งหมด ต้องประสานกันได้ ต้องโตให้ไวที่สุด'

DAEMONS RUNNING IN PARALLEL (31 total, coordinated via central state):
- 8 dataset/source: bulk-ingest-parallel (4\u21928 shards now), discoverer,
GitHub crawler, agentic crawler, scrape-continuous, domain-scrape,
refresh-cve, scrape-sre-postmortems
- 7 dev/orchestrate: auto-orchestrate-continuous (4 workers), dev-loop,
research-apply, research-loop, daemon, work-queue, moa-consensus
- 3 index/RAG: self-ingest FTS5, rag-vector-builder, skill-synthesis
- 3 output/sync: push-training-to-hf, synthetic-data, daily-summary
- 4 infra: status-server, discord-bot, redis, ollama
- 4 helpers (called per orchestrate stage)

COORDINATION VIA CENTRAL STATE:
- dedup.db (every writer flows through)
- hf-dataset-frontier.db (discoverer)
- github-frontier.db (GH crawler)
- agentic-frontier.db (web crawler)
- self-ingest.db (FTS5 RAG)
- rag-vectors.db (vector RAG)
- training-pairs.jsonl (single sink)

CHANGES THIS PUSH:
1. bulk-ingest-parallel: 4 \u2192 8 shards (2\u00d7 throughput)
Cooldown 5min \u2192 3min (faster cycles)
Stagger startup 30s \u2192 15s (faster spinup)

2. auto-orchestrate-loop: round-robin projects (was random)
PROBLEM: 6 commits today all in arkashira/surrogate, 0 in 4 other repos
random.shuffle kept biased toward same repo
FIX: persistent cursor at ~/.surrogate/state/orchestrate-project-cursor
Each call: rotate through (Costinel \u2192 vanguard \u2192 arkship \u2192 surrogate \u2192
workio \u2192 hermes-toolbelt \u2192 cycle)
Result: 4 parallel workers \u00d7 round-robin = even coverage of 6 axentx repos

EXPECTED IMPACT:
- Day-1 ingest: 400K/h \u2192 800K/h (2\u00d7 shards)
- axentx commits: spread evenly across 6 repos
- Same hardware, better parallelism + balance

bin/auto-orchestrate-loop.sh

@@ -47,7 +47,21 @@ PROJECTS = [p for p in PROJECTS if (p/'.git').exists()]
 if not PROJECTS:
     print("{}"); exit()
 
-random.shuffle(PROJECTS)
+# ROUND-ROBIN across projects (instead of random.shuffle which kept hitting same repo).
+# Persistent counter at ~/.surrogate/state/orchestrate-project-cursor — increments each run.
+# Result: every 6 runs covers all 6 axentx repos evenly.
+cursor_file = Path.home() / '.surrogate/state/orchestrate-project-cursor'
+cursor_file.parent.mkdir(parents=True, exist_ok=True)
+try:
+    cursor = int(cursor_file.read_text().strip())
+except Exception:
+    cursor = 0
+PROJECTS = sorted(PROJECTS, key=lambda p: p.name)  # stable order
+cursor = cursor % len(PROJECTS)
+# Rotate so current cursor's project is first
+PROJECTS = PROJECTS[cursor:] + PROJECTS[:cursor]
+cursor_file.write_text(str((cursor + 1) % len(PROJECTS)))
+
 for proj in PROJECTS:
     cmd = ['rg', '--no-heading', '-n', '-m', '5',
            '--type', 'py', '--type', 'ts', '--type', 'go', '--type', 'sh',

bin/bulk-ingest-parallel.sh

@@ -10,8 +10,8 @@ set -a; source "$HOME/.hermes/.env" 2>/dev/null; set +a
 LOG="$HOME/.surrogate/logs/bulk-ingest-parallel.log"
 mkdir -p "$(dirname "$LOG")"
 
-NUM_SHARDS="${INGEST_SHARDS:-4}"
-SHARD_COOLDOWN="${SHARD_COOLDOWN:-300}"  # 5 min between shard cycles
+NUM_SHARDS="${INGEST_SHARDS:-8}"
+SHARD_COOLDOWN="${SHARD_COOLDOWN:-180}"  # 3 min between shard cycles (was 5)
 
 echo "[$(date +%H:%M:%S)] bulk-ingest-parallel start (shards=$NUM_SHARDS)" | tee -a "$LOG"
 
@@ -28,9 +28,9 @@ shard_loop() {
     done
 }
 
-# Stagger startup 30s apart to avoid all hitting HF API simultaneously
+# Stagger startup 15s apart (was 30s) to spin up faster
 for i in $(seq 0 $((NUM_SHARDS - 1))); do
     shard_loop "$i" "$NUM_SHARDS" &
-    sleep 30
+    sleep 15
 done
 wait