diff --git a/.env.example b/.env.example
index f46e737..4a19cb0 100644
--- a/.env.example
+++ b/.env.example
@@ -57,17 +57,27 @@ NETX_UME_NOTIFICATION_TOPIC=ALARM
 # Device collectors run inline with the API by default (frontend+backend start is enough).
 # Production split only: NETX_RUN_INLINE_SCHEDULERS=false and run `python -m netx_api.worker`
 # NETX_RUN_INLINE_SCHEDULERS=false
-# NETX_AUDIT_ASYNC=true
-# NETX_AUDIT_SAMPLE_N=1
-# DB pool (QueuePool). Aim: pool_size + max_overflow >= HTTP peak + NETX_CLI_MAX_CONCURRENT + UME WS burst.
-# NETX_DB_POOL_SIZE=20
-# NETX_DB_MAX_OVERFLOW=20
+# --- Production-recommended capacity (defaults in Settings already match these) ---
+# NETX_DB_POOL_SIZE=25
+# NETX_DB_MAX_OVERFLOW=15
 # NETX_DB_POOL_RECYCLE_SEC=1800
 # NETX_DB_POOL_TIMEOUT_SEC=30
-# Global SSH/Netmiko concurrency across discover / collect / config_sync / port_traffic.
-# NETX_CLI_MAX_CONCURRENT=20
-# NETX_CLI_TIMEOUT_POOL_WORKERS=16
-# NETX_PORT_TRAFFIC_DISPATCH_WORKERS=4
-# NETX_AUDIT_QUEUE_MAX=5000
-# NETX_NE_COLLECT_MAX_OUTPUT_BYTES=8388608
-# NETX_WEBCRT_SESSION_LOG_MAX_BYTES=4194304
+# NETX_CLI_MAX_CONCURRENT=12
+# NETX_CLI_FEATURE_HARD_CAP=16
+# NETX_CLI_TIMEOUT_POOL_WORKERS=12
+# NETX_PORT_TRAFFIC_DISPATCH_WORKERS=3
+# NETX_NE_CONNECT_MAX_WORKERS=4
+# NETX_NE_COLLECT_MAX_WORKERS=4
+# NETX_NE_COLLECT_RUN_TIMEOUT_CAP_SEC=480
+# NETX_NE_COLLECT_MAX_OUTPUT_BYTES=4194304
+# NETX_WEBCRT_MAX_SESSIONS=12
+# NETX_WEBCRT_KEEPALIVE_SEC=30
+# NETX_WEBCRT_OUT_QUEUE_MAX=1500
+# NETX_WEBCRT_SESSION_LOG_MAX_BYTES=2097152
+# NETX_WEBCRT_SFTP_MAX_FILE_BYTES=268435456
+# NETX_AUDIT_ASYNC=true
+# NETX_AUDIT_SAMPLE_N=10
+# NETX_AUDIT_QUEUE_MAX=2000
+# NETX_OCLAW_FORWARD_QUEUE_MAX=2000
+# NETX_OCLAW_FORWARD_MAX_RETRIES=3
+# Lab-only raises (optional): higher CLI/DB when bastion and Postgres are sized for it.
diff --git a/netx_api/app_startup.py b/netx_api/app_startup.py
index a8f19f2..56a5ccf 100644
--- a/netx_api/app_startup.py
+++ b/netx_api/app_startup.py
@@ -142,3 +142,9 @@ def run_api_startup() -> None:
         )
 
     start_api_sideband_threads()
+    try:
+        from .runtime_budget import log_runtime_budget
+
+        log_runtime_budget(role="api")
+    except Exception:
+        _log.exception("startup: runtime budget log failed")
diff --git a/netx_api/cli_budget.py b/netx_api/cli_budget.py
index 18c35bf..cabcf8e 100644
--- a/netx_api/cli_budget.py
+++ b/netx_api/cli_budget.py
@@ -17,7 +17,7 @@ _in_use_lock = threading.Lock()
 def _ensure_sem() -> threading.BoundedSemaphore:
     global _sem, _limit
     with _lock:
-        want = max(1, int(getattr(settings, "cli_max_concurrent", 20) or 20))
+        want = max(1, int(getattr(settings, "cli_max_concurrent", 12) or 12))
         if _sem is None or want != _limit:
             _sem = threading.BoundedSemaphore(want)
             _limit = want
@@ -32,10 +32,20 @@ def cli_budget_status() -> dict[str, int]:
     return {"limit": _limit, "in_use": used, "available": max(0, _limit - used)}
 
 
-def clamp_cli_workers(requested: int, *, hard_cap: int) -> int:
+def clamp_cli_workers(requested: int, *, hard_cap: int | None = None) -> int:
     """Clamp a feature concurrency against the global CLI budget and a hard cap."""
-    budget = max(1, int(getattr(settings, "cli_max_concurrent", 20) or 20))
-    return max(1, min(int(hard_cap), int(requested or 1), budget))
+    budget = max(1, int(getattr(settings, "cli_max_concurrent", 12) or 12))
+    feature_cap = int(
+        hard_cap
+        if hard_cap is not None
+        else (getattr(settings, "cli_feature_hard_cap", 16) or 16)
+    )
+    feature_cap = max(1, feature_cap)
+    return max(1, min(int(feature_cap), int(requested or 1), budget))
+
+
+def feature_hard_cap() -> int:
+    return max(1, int(getattr(settings, "cli_feature_hard_cap", 16) or 16))
 
 
 @contextmanager
diff --git a/netx_api/config.py b/netx_api/config.py
index ca4852a..b189b04 100644
--- a/netx_api/config.py
+++ b/netx_api/config.py
@@ -64,13 +64,14 @@ class Settings(BaseSettings):
     ume_sync_alarms_history_every_hours: int = 24
     # Managed NE credentials (Fernet key; generate with cryptography.fernet.Fernet.generate_key())
     credential_secret_key: str = ""
-    ne_connect_max_workers: int = 5
+    # Production-oriented worker caps (raise only when bastion/DB capacity allows).
+    ne_connect_max_workers: int = 4
     ne_connect_timeout_sec: int = 30
-    ne_collect_max_workers: int = 5
+    ne_collect_max_workers: int = 4
     ne_collect_read_timeout_sec: int = 120
     ne_collect_stale_run_sec: int = 900
     ne_collect_pending_stale_sec: int = 180
-    ne_collect_run_timeout_cap_sec: int = 600
+    ne_collect_run_timeout_cap_sec: int = 480
     ne_collection_data_dir: str = "data/ne_collections"
     # Config sync (periodic running-config backup into DB)
     config_sync_scheduler_enabled: bool = True
@@ -91,29 +92,29 @@ class Settings(BaseSettings):
     # Managed NE exec: max CLI commands per request (lab can raise; hard-capped in ne_exec).
     ne_exec_max_commands: int = 5
     # WebCRT interactive terminal sessions
-    webcrt_max_sessions: int = 20
+    webcrt_max_sessions: int = 12
     webcrt_idle_timeout_sec: int = 1800
     webcrt_connect_timeout_sec: int = 90
     webcrt_attach_timeout_sec: int = 60
     # Keep device PTY after WS drop so the UI can re-attach (reconnect / remount).
     webcrt_detach_grace_sec: int = 120
     webcrt_data_dir: str = "data/webcrt"
-    # SSH transport keepalive interval (seconds); 0 disables (default off).
-    webcrt_keepalive_sec: int = 0
+    # SSH transport keepalive (seconds); 30s recommended behind NAT/firewall.
+    webcrt_keepalive_sec: int = 30
     # Device anti-idle CLI nudge (0 = off). Keep off: NEs close idle VTY themselves.
     webcrt_anti_idle_sec: int = 0
     webcrt_anti_idle_payload: str = " "
     # Cap stdout queue depth (drop oldest when full) to protect memory.
-    webcrt_out_queue_max: int = 2000
+    webcrt_out_queue_max: int = 1500
     # Persist per-session transcripts under webcrt_data_dir/sessions/.
     webcrt_session_log_enabled: bool = True
     # Reader: short blocking wait instead of fixed 40ms spin (seconds).
     webcrt_reader_poll_sec: float = 0.01
-    # WebCRT SFTP transfer limits (streamed; default 512 MiB per file).
-    webcrt_sftp_max_file_bytes: int = 512 * 1024 * 1024
+    # WebCRT SFTP transfer limits (streamed; default 256 MiB per file).
+    webcrt_sftp_max_file_bytes: int = 256 * 1024 * 1024
     webcrt_sftp_chunk_bytes: int = 64 * 1024
     # Cap directory listings so huge folders cannot pin the API/UI.
-    webcrt_sftp_list_max_entries: int = 5000
+    webcrt_sftp_list_max_entries: int = 2000
     webcrt_sftp_list_timeout_sec: float = 30.0
     # Local app login / audit (lab defaults; override in production)
     auth_enabled: bool = True
@@ -132,7 +133,7 @@ class Settings(BaseSettings):
     allow_insecure_defaults: bool = False
     # Async audit writer; sample_n>1 keeps 1/N of generic http.* events.
     audit_async: bool = True
-    audit_sample_n: int = 1
+    audit_sample_n: int = 10
     # Prefer Alembic on API start; brownfield patches live in schema_patches + revisions.
     # Auth column ensures still run as a safety net before bootstrap.
     skip_legacy_startup_ddl: bool = True
@@ -145,22 +146,27 @@ class Settings(BaseSettings):
     run_inline_schedulers: bool = True
     # SQLAlchemy QueuePool (API + background workers share one engine).
     # Rule of thumb: pool_size + max_overflow >= HTTP peak + cli_max_concurrent + UME WS burst.
-    db_pool_size: int = 20
-    db_max_overflow: int = 20
+    db_pool_size: int = 25
+    db_max_overflow: int = 15
     db_pool_recycle_sec: int = 1800
     db_pool_timeout_sec: int = 30
     # Global Netmiko/SSH concurrency across discover / collect / config_sync / port_traffic.
-    cli_max_concurrent: int = 20
+    cli_max_concurrent: int = 12
+    # Per-feature concurrency hard ceiling (API body / policy cannot exceed this).
+    cli_feature_hard_cap: int = 16
     # Shared timeout watchdog pool (not per-task executors).
-    cli_timeout_pool_workers: int = 16
+    cli_timeout_pool_workers: int = 12
     # Port-traffic: how many devices may collect in parallel on the scheduler tick.
-    port_traffic_dispatch_workers: int = 4
+    port_traffic_dispatch_workers: int = 3
     # Bound async audit queue; drop oldest when full to protect RSS.
-    audit_queue_max: int = 5000
+    audit_queue_max: int = 2000
     # Cap NE collection output files (bytes); 0 = unlimited (not recommended).
-    ne_collect_max_output_bytes: int = 8 * 1024 * 1024
+    ne_collect_max_output_bytes: int = 4 * 1024 * 1024
     # WebCRT session transcript rotate size (bytes); 0 disables rotate.
-    webcrt_session_log_max_bytes: int = 4 * 1024 * 1024
+    webcrt_session_log_max_bytes: int = 2 * 1024 * 1024
+    # oclaw alarm forwarder: requeue attempts before drop on send failure.
+    oclaw_forward_max_retries: int = 3
+    oclaw_forward_queue_max: int = 2000
 
 
 settings = Settings()
diff --git a/netx_api/config_sync_common.py b/netx_api/config_sync_common.py
index fa5cba9..ef5c2ab 100644
--- a/netx_api/config_sync_common.py
+++ b/netx_api/config_sync_common.py
@@ -84,7 +84,7 @@ def policy_to_out(row: ConfigSyncPolicy) -> ConfigSyncPolicyOut:
     return ConfigSyncPolicyOut(
         enabled=bool(row.enabled),
         interval_days=max(1, int(row.interval_days or 3)),
-        concurrency=max(1, min(30, int(row.concurrency or 5))),
+        concurrency=max(1, min(16, int(row.concurrency or 5))),
         scope_mode=str(row.scope_mode or "all"),
         selected_targets=_targets_from_json(row.selected_targets),
         history_keep=max(0, min(30, int(row.history_keep if row.history_keep is not None else 3))),
diff --git a/netx_api/config_sync_cycles.py b/netx_api/config_sync_cycles.py
index e9a703b..0602a54 100644
--- a/netx_api/config_sync_cycles.py
+++ b/netx_api/config_sync_cycles.py
@@ -50,7 +50,7 @@ def update_policy(db: Session, body: ConfigSyncPolicyUpdate) -> ConfigSyncPolicy
     if "interval_days" in data and data["interval_days"] is not None:
         row.interval_days = int(data["interval_days"])
     if "concurrency" in data and data["concurrency"] is not None:
-        row.concurrency = max(1, min(30, int(data["concurrency"])))
+        row.concurrency = max(1, min(16, int(data["concurrency"])))
     if "scope_mode" in data and data["scope_mode"] is not None:
         row.scope_mode = str(data["scope_mode"])
     if "selected_targets" in data and data["selected_targets"] is not None:
@@ -183,7 +183,7 @@ def create_cycle(db: Session, body: ConfigSyncCycleCreate) -> ConfigSyncCycleOut
     policy = ensure_policy(db)
     mode = str(body.mode or "full").strip().lower()
     trigger = "retry_failed" if mode == "retry_failed" else "manual"
-    concurrency = max(1, min(30, int(policy.concurrency or 5)))
+    concurrency = max(1, min(16, int(policy.concurrency or 5)))
 
     targets: list[dict[str, str]] = []
     if mode == "retry_failed":
diff --git a/netx_api/config_sync_runner.py b/netx_api/config_sync_runner.py
index 3a064dd..aa354d6 100644
--- a/netx_api/config_sync_runner.py
+++ b/netx_api/config_sync_runner.py
@@ -43,7 +43,7 @@ def _pool_for_cycle(cycle_id: str, concurrency: int) -> ThreadPoolExecutor:
     with _pools_lock:
         pool = _pools.get(cycle_id)
         if pool is None:
-            workers = clamp_cli_workers(int(concurrency or 5), hard_cap=30)
+            workers = clamp_cli_workers(int(concurrency or 5))
             pool = ThreadPoolExecutor(max_workers=workers, thread_name_prefix=f"cfg-sync-{cycle_id[:8]}")
             _pools[cycle_id] = pool
         return pool
@@ -462,7 +462,7 @@ def dispatch_cycle(cycle_id: str) -> int:
             .all()
         )
         task_ids = [str(t.id) for t in pending]
-        concurrency = max(1, min(30, int(cycle.concurrency or 5)))
+        concurrency = max(1, min(16, int(cycle.concurrency or 5)))
     finally:
         db.close()
     return schedule_cycle_tasks(cycle_id, task_ids, concurrency)
diff --git a/netx_api/config_sync_scheduler.py b/netx_api/config_sync_scheduler.py
index 4f052b6..3a6c4d1 100644
--- a/netx_api/config_sync_scheduler.py
+++ b/netx_api/config_sync_scheduler.py
@@ -74,7 +74,7 @@ def try_start_scheduled_cycle() -> str | None:
         if not targets:
             _log.info("config_sync schedule skip: no targets")
             return None
-        concurrency = max(1, min(30, int(policy.concurrency or 5)))
+        concurrency = max(1, min(16, int(policy.concurrency or 5)))
         cycle = ConfigSyncCycle(
             id=uuid4().hex,
             trigger_mode="schedule",
diff --git a/netx_api/config_sync_schemas.py b/netx_api/config_sync_schemas.py
index 67418dc..419696c 100644
--- a/netx_api/config_sync_schemas.py
+++ b/netx_api/config_sync_schemas.py
@@ -27,7 +27,7 @@ class ConfigSyncPolicyOut(BaseModel):
 class ConfigSyncPolicyUpdate(BaseModel):
     enabled: bool | None = None
     interval_days: int | None = Field(default=None, ge=1, le=365)
-    concurrency: int | None = Field(default=None, ge=1, le=30)
+    concurrency: int | None = Field(default=None, ge=1, le=16)
     scope_mode: Literal["all", "selected"] | None = None
     selected_targets: list[ConfigSyncTargetRef] | None = None
     history_keep: int | None = Field(default=None, ge=0, le=30)
diff --git a/netx_api/lldp_collect_schemas.py b/netx_api/lldp_collect_schemas.py
index 39e9e36..a7f72ef 100644
--- a/netx_api/lldp_collect_schemas.py
+++ b/netx_api/lldp_collect_schemas.py
@@ -29,7 +29,7 @@ class LldpCollectPolicyUpdate(BaseModel):
     enabled: bool | None = None
     interval_days: int | None = Field(default=None, ge=1, le=365)
     interval_hours: int | None = Field(default=None, ge=1, le=8760)
-    concurrency: int | None = Field(default=None, ge=1, le=32)
+    concurrency: int | None = Field(default=None, ge=1, le=16)
     scope_mode: str | None = None
     selected_targets: list[LldpCollectTargetRef] | None = None
     auto_add_unmatched: bool | None = None
diff --git a/netx_api/lldp_collect_service.py b/netx_api/lldp_collect_service.py
index 99330bd..f35a377 100644
--- a/netx_api/lldp_collect_service.py
+++ b/netx_api/lldp_collect_service.py
@@ -113,7 +113,7 @@ def update_policy(db: Session, body: LldpCollectPolicyUpdate) -> LldpCollectPoli
         row.interval_days = days
         row.interval_hours = days * 24
     if "concurrency" in data and data["concurrency"] is not None:
-        row.concurrency = max(1, min(32, int(data["concurrency"])))
+        row.concurrency = max(1, min(16, int(data["concurrency"])))
     if "scope_mode" in data and data["scope_mode"] is not None:
         mode = str(data["scope_mode"] or "").strip().lower()
         if mode not in {"all", "selected"}:
@@ -208,7 +208,7 @@ def next_due_at(db: Session, policy: LldpCollectPolicy) -> datetime | None:
 
 
 def build_discover_request(policy: LldpCollectPolicy) -> FabricDiscoverRequest:
-    concurrency = max(1, min(32, int(policy.concurrency or 4)))
+    concurrency = max(1, min(16, int(policy.concurrency or 4)))
     auto_add = bool(policy.auto_add_unmatched)
     if str(policy.scope_mode or "") == "selected":
         managed_ids: list[str] = []
diff --git a/netx_api/metrics_router.py b/netx_api/metrics_router.py
index 05e430d..56e4ee1 100644
--- a/netx_api/metrics_router.py
+++ b/netx_api/metrics_router.py
@@ -83,6 +83,9 @@ def _prom_lines(metrics: dict[str, Any]) -> str:
         ("queue_size", "netx_oclaw_forwarder_queue_size"),
         ("published_ok", "netx_oclaw_forwarder_published_ok"),
         ("published_fail", "netx_oclaw_forwarder_published_fail"),
+        ("dropped", "netx_oclaw_forwarder_dropped"),
+        ("requeued", "netx_oclaw_forwarder_requeued"),
+        ("retry_exhausted", "netx_oclaw_forwarder_retry_exhausted"),
     ):
         if key in fwd:
             lines.append(f"{prom} {int(fwd.get(key) or 0)}")
diff --git a/netx_api/ne_collect_runner.py b/netx_api/ne_collect_runner.py
index 6d066bb..bdbf407 100644
--- a/netx_api/ne_collect_runner.py
+++ b/netx_api/ne_collect_runner.py
@@ -29,7 +29,7 @@ _executor: ThreadPoolExecutor | None = None
 def _executor_pool() -> ThreadPoolExecutor:
     global _executor
     if _executor is None:
-        workers = clamp_cli_workers(int(settings.ne_collect_max_workers or 5), hard_cap=32)
+        workers = clamp_cli_workers(int(settings.ne_collect_max_workers or 4))
         _executor = ThreadPoolExecutor(max_workers=workers, thread_name_prefix="ne-collect")
     return _executor
 
diff --git a/netx_api/ne_connect.py b/netx_api/ne_connect.py
index 08287d2..94b7548 100644
--- a/netx_api/ne_connect.py
+++ b/netx_api/ne_connect.py
@@ -32,7 +32,7 @@ def _executor_pool() -> ThreadPoolExecutor:
     if _executor is None:
         from .cli_budget import clamp_cli_workers
 
-        workers = clamp_cli_workers(int(settings.ne_connect_max_workers or 5), hard_cap=32)
+        workers = clamp_cli_workers(int(settings.ne_connect_max_workers or 4))
         _executor = ThreadPoolExecutor(max_workers=workers, thread_name_prefix="ne-connect")
     return _executor
 
diff --git a/netx_api/oclaw_alarm_forwarder.py b/netx_api/oclaw_alarm_forwarder.py
index f9ea655..164806b 100644
--- a/netx_api/oclaw_alarm_forwarder.py
+++ b/netx_api/oclaw_alarm_forwarder.py
@@ -15,7 +15,8 @@ from .config import settings
 
 _log = logging.getLogger("netx.oclaw.alarm_forwarder")
 
-_OUTBOUND_Q: "queue.Queue[dict[str, Any]]" = queue.Queue(maxsize=5000)
+_OUTBOUND_Q: "queue.Queue[dict[str, Any]] | None" = None
+_Q_LOCK = threading.Lock()
 _STOP_EVENT = threading.Event()
 _THREAD: threading.Thread | None = None
 _CONN_LOCK = threading.Lock()
@@ -28,9 +29,21 @@ _STATS: dict[str, int] = {
     "published_ok": 0,
     "published_fail": 0,
     "queued": 0,
+    "dropped": 0,
+    "requeued": 0,
+    "retry_exhausted": 0,
 }
 
 
+def _outbound_q() -> "queue.Queue[dict[str, Any]]":
+    global _OUTBOUND_Q
+    with _Q_LOCK:
+        if _OUTBOUND_Q is None:
+            maxsize = max(100, int(getattr(settings, "oclaw_forward_queue_max", 2000) or 2000))
+            _OUTBOUND_Q = queue.Queue(maxsize=maxsize)
+        return _OUTBOUND_Q
+
+
 def _utc_now_iso() -> str:
     return datetime.now(timezone.utc).isoformat()
 
@@ -95,15 +108,46 @@ def enqueue_alarm_forward(payload: dict[str, Any]) -> bool:
     if not is_forwarder_operational():
         return False
     try:
-        _OUTBOUND_Q.put_nowait(dict(payload))
+        _outbound_q().put_nowait(dict(payload))
         with _STATS_LOCK:
             _STATS["queued"] = int(_STATS.get("queued", 0)) + 1
         return True
     except queue.Full:
+        with _STATS_LOCK:
+            _STATS["dropped"] = int(_STATS.get("dropped", 0)) + 1
         _log.warning("oclaw alarm forward queue full; dropping alarm_key=%s", payload.get("alarm_key"))
         return False
 
 
+def _requeue_or_drop(payload: dict[str, Any], *, reason: str) -> None:
+    max_retries = max(0, int(getattr(settings, "oclaw_forward_max_retries", 3) or 3))
+    item = dict(payload)
+    attempts = int(item.get("_fwd_attempts") or 0) + 1
+    item["_fwd_attempts"] = attempts
+    if attempts > max_retries:
+        with _STATS_LOCK:
+            _STATS["retry_exhausted"] = int(_STATS.get("retry_exhausted", 0)) + 1
+            _STATS["dropped"] = int(_STATS.get("dropped", 0)) + 1
+        _log.warning(
+            "oclaw forward drop after retries alarm_key=%s attempts=%s reason=%s",
+            item.get("alarm_key"),
+            attempts,
+            reason[:80],
+        )
+        return
+    try:
+        _outbound_q().put_nowait(item)
+        with _STATS_LOCK:
+            _STATS["requeued"] = int(_STATS.get("requeued", 0)) + 1
+    except queue.Full:
+        with _STATS_LOCK:
+            _STATS["dropped"] = int(_STATS.get("dropped", 0)) + 1
+        _log.warning(
+            "oclaw forward requeue full; dropping alarm_key=%s",
+            item.get("alarm_key"),
+        )
+
+
 def _send_auth(ws: Any) -> bool:
     token = _bridge_token()
     ws.send(json.dumps({"type": "auth", "token": token}, ensure_ascii=False))
@@ -187,7 +231,7 @@ def _run_loop() -> None:
                     _notify_status("fwd:disabled")
                     raise RuntimeError("forwarder disabled")
                 try:
-                    payload = _OUTBOUND_Q.get(timeout=1.0)
+                    payload = _outbound_q().get(timeout=1.0)
                 except queue.Empty:
                     try:
                         ws.send(json.dumps({"type": "ping", "ts": _utc_now_iso()}, ensure_ascii=False))
@@ -221,15 +265,13 @@ def _run_loop() -> None:
                             payload.get("alarm_key"),
                             str(ack.get("error") or "")[:120],
                         )
+                        # Soft ack failure: do not infinite-requeue; count as fail only.
                 except Exception as exc:
                     _log.warning("oclaw forward failed alarm_key=%s err=%s", payload.get("alarm_key"), str(exc)[:120])
-                    try:
-                        _OUTBOUND_Q.put_nowait(payload)
-                    except queue.Full:
-                        pass
+                    _requeue_or_drop(payload, reason=str(exc)[:120])
                     raise
                 finally:
-                    _OUTBOUND_Q.task_done()
+                    _outbound_q().task_done()
         except Exception as exc:
             _CONNECTED.clear()
             err = str(exc)[:200]
@@ -272,9 +314,13 @@ def forwarder_status() -> dict[str, Any]:
         "operational": bool(enabled and not paused),
         "paused": paused,
         "connected": _CONNECTED.is_set(),
-        "queue_size": int(_OUTBOUND_Q.qsize()),
+        "queue_size": int(_outbound_q().qsize()),
+        "queue_max": max(100, int(getattr(settings, "oclaw_forward_queue_max", 2000) or 2000)),
         "url": _bridge_url(),
         "published_ok": int(stats.get("published_ok", 0)),
         "published_fail": int(stats.get("published_fail", 0)),
         "queued_total": int(stats.get("queued", 0)),
+        "dropped": int(stats.get("dropped", 0)),
+        "requeued": int(stats.get("requeued", 0)),
+        "retry_exhausted": int(stats.get("retry_exhausted", 0)),
     }
diff --git a/netx_api/port_traffic_scheduler.py b/netx_api/port_traffic_scheduler.py
index deaa031..ccc8fe1 100644
--- a/netx_api/port_traffic_scheduler.py
+++ b/netx_api/port_traffic_scheduler.py
@@ -33,8 +33,7 @@ def _dispatch_pool_get() -> ThreadPoolExecutor:
     with _pool_lock:
         if _dispatch_pool is None:
             workers = clamp_cli_workers(
-                int(getattr(settings, "port_traffic_dispatch_workers", 4) or 4),
-                hard_cap=16,
+                int(getattr(settings, "port_traffic_dispatch_workers", 3) or 3),
             )
             _dispatch_pool = ThreadPoolExecutor(
                 max_workers=workers, thread_name_prefix="pt-dispatch"
diff --git a/netx_api/runtime_budget.py b/netx_api/runtime_budget.py
new file mode 100644
index 0000000..3863a58
--- /dev/null
+++ b/netx_api/runtime_budget.py
@@ -0,0 +1,57 @@
+"""Startup capacity checks and operator-facing budget logging."""
+from __future__ import annotations
+
+import logging
+
+from .cli_budget import cli_budget_status, feature_hard_cap
+from .config import settings
+from .db import db_pool_status
+
+_log = logging.getLogger("netx.runtime.budget")
+
+
+def log_runtime_budget(*, role: str = "api") -> None:
+    """Log effective pools/CLI caps and warn when capacity looks undersized."""
+    pool_size = max(1, int(getattr(settings, "db_pool_size", 25) or 25))
+    overflow = max(0, int(getattr(settings, "db_max_overflow", 15) or 15))
+    pool_cap = pool_size + overflow
+    cli = max(1, int(getattr(settings, "cli_max_concurrent", 12) or 12))
+    hard = feature_hard_cap()
+    http_reserve = 8  # rough floor for request handlers + UME WS bursts
+    recommended_pool = cli + http_reserve
+
+    _log.info(
+        "runtime budget role=%s db_pool=%s+%s=%s cli_max=%s feature_hard_cap=%s "
+        "timeout_pool=%s pt_dispatch=%s webcrt_sessions=%s audit_sample_n=%s inline_schedulers=%s",
+        role,
+        pool_size,
+        overflow,
+        pool_cap,
+        cli,
+        hard,
+        int(getattr(settings, "cli_timeout_pool_workers", 12) or 12),
+        int(getattr(settings, "port_traffic_dispatch_workers", 3) or 3),
+        int(getattr(settings, "webcrt_max_sessions", 12) or 12),
+        int(getattr(settings, "audit_sample_n", 10) or 10),
+        bool(getattr(settings, "run_inline_schedulers", True)),
+    )
+    _log.info("runtime budget snapshot pool=%s cli=%s", db_pool_status(), cli_budget_status())
+
+    if pool_cap < recommended_pool:
+        _log.warning(
+            "DB pool capacity %s < recommended %s (cli_max=%s + http_reserve=%s). "
+            "Raise NETX_DB_POOL_SIZE / NETX_DB_MAX_OVERFLOW or lower NETX_CLI_MAX_CONCURRENT.",
+            pool_cap,
+            recommended_pool,
+            cli,
+            http_reserve,
+        )
+    host = str(getattr(settings, "host", "") or "").strip().lower()
+    if host not in {"127.0.0.1", "localhost", "::1"} and bool(
+        getattr(settings, "run_inline_schedulers", True)
+    ):
+        _log.warning(
+            "Non-loopback bind (%s) with inline schedulers — for production prefer "
+            "NETX_RUN_INLINE_SCHEDULERS=false and `python -m netx_api.worker`.",
+            host,
+        )
diff --git a/netx_api/topology_discover_jobs.py b/netx_api/topology_discover_jobs.py
index 873929e..5c0f037 100644
--- a/netx_api/topology_discover_jobs.py
+++ b/netx_api/topology_discover_jobs.py
@@ -59,7 +59,7 @@ def _run_discover_job(job_id: str, body: FabricDiscoverRequest) -> None:
 
         from .cli_budget import clamp_cli_workers
 
-        concurrency = clamp_cli_workers(int(body.concurrency or 4), hard_cap=32)
+        concurrency = clamp_cli_workers(int(body.concurrency or 4))
         added = 0
         updated = 0
         stale = 0
diff --git a/netx_api/topology_schemas.py b/netx_api/topology_schemas.py
index a564b3a..d4f7598 100644
--- a/netx_api/topology_schemas.py
+++ b/netx_api/topology_schemas.py
@@ -85,7 +85,7 @@ class FabricDiscoverRequest(BaseModel):
         default=True,
         description="Create SSH placeholder ManagedNEs for LLDP neighbors not in inventory",
     )
-    concurrency: int = Field(default=4, ge=1, le=32)
+    concurrency: int = Field(default=4, ge=1, le=16)
     trigger_mode: str = Field(default="manual", description="manual | schedule | topology")
 
 
diff --git a/netx_api/worker.py b/netx_api/worker.py
index f2c97dc..fc6c66a 100644
--- a/netx_api/worker.py
+++ b/netx_api/worker.py
@@ -39,6 +39,12 @@ def main() -> None:
     from .ume_runtime import start_device_schedulers
 
     start_device_schedulers()
+    try:
+        from .runtime_budget import log_runtime_budget
+
+        log_runtime_budget(role="worker")
+    except Exception:
+        _log.exception("worker runtime budget log failed")
     _log.info("netx worker schedulers started (config_sync, lldp_collect, port_traffic)")
 
     while not stop.is_set():
diff --git a/tests/test_stability_hardening.py b/tests/test_stability_hardening.py
index bba42c3..a0731fb 100644
--- a/tests/test_stability_hardening.py
+++ b/tests/test_stability_hardening.py
@@ -1,25 +1,47 @@
-"""Smoke tests for DB pool / CLI budget / metrics wiring."""
+"""Smoke tests for DB pool / CLI budget / metrics / production defaults."""
 
 from __future__ import annotations
 
 import unittest
 
-from netx_api.cli_budget import acquire_cli_slot, clamp_cli_workers, cli_budget_status
-from netx_api.config import settings
+from netx_api.cli_budget import acquire_cli_slot, clamp_cli_workers, cli_budget_status, feature_hard_cap
+from netx_api.config import Settings, settings
 from netx_api.db import db_pool_status
 from netx_api.metrics_router import _prom_lines, collect_runtime_metrics
+from netx_api.runtime_budget import log_runtime_budget
 
 
 class StabilityHardeningTests(unittest.TestCase):
+    def test_production_defaults(self) -> None:
+        s = Settings(_env_file=None)
+        self.assertEqual(s.db_pool_size, 25)
+        self.assertEqual(s.db_max_overflow, 15)
+        self.assertEqual(s.cli_max_concurrent, 12)
+        self.assertEqual(s.cli_feature_hard_cap, 16)
+        self.assertEqual(s.cli_timeout_pool_workers, 12)
+        self.assertEqual(s.port_traffic_dispatch_workers, 3)
+        self.assertEqual(s.ne_connect_max_workers, 4)
+        self.assertEqual(s.ne_collect_max_workers, 4)
+        self.assertEqual(s.webcrt_max_sessions, 12)
+        self.assertEqual(s.webcrt_keepalive_sec, 30)
+        self.assertEqual(s.audit_sample_n, 10)
+        self.assertEqual(s.audit_queue_max, 2000)
+        self.assertEqual(s.oclaw_forward_max_retries, 3)
+        self.assertEqual(s.oclaw_forward_queue_max, 2000)
+        self.assertEqual(s.ne_collect_max_output_bytes, 4 * 1024 * 1024)
+        self.assertEqual(s.webcrt_session_log_max_bytes, 2 * 1024 * 1024)
+        # Pool should cover CLI + HTTP reserve under defaults.
+        self.assertGreaterEqual(s.db_pool_size + s.db_max_overflow, s.cli_max_concurrent + 8)
+
     def test_settings_have_pool_and_budget_knobs(self) -> None:
         self.assertGreaterEqual(int(settings.db_pool_size), 1)
         self.assertGreaterEqual(int(settings.cli_max_concurrent), 1)
         self.assertGreaterEqual(int(settings.audit_queue_max), 100)
 
     def test_clamp_cli_workers(self) -> None:
-        n = clamp_cli_workers(999, hard_cap=32)
+        n = clamp_cli_workers(999)
         self.assertLessEqual(n, int(settings.cli_max_concurrent))
-        self.assertLessEqual(n, 32)
+        self.assertLessEqual(n, feature_hard_cap())
         self.assertGreaterEqual(n, 1)
 
     def test_cli_budget_acquire_release(self) -> None:
@@ -39,6 +61,10 @@ class StabilityHardeningTests(unittest.TestCase):
         self.assertIn("netx_uptime_seconds", body)
         self.assertIn("netx_thread_count", body)
         self.assertIn("netx_cli_budget_limit", body)
+        self.assertIn("netx_oclaw_forwarder_dropped", body)
+
+    def test_log_runtime_budget_does_not_raise(self) -> None:
+        log_runtime_budget(role="test")
 
 
 if __name__ == "__main__":
diff --git a/tests/test_webcrt.py b/tests/test_webcrt.py
index cb70437..fff29a8 100644
--- a/tests/test_webcrt.py
+++ b/tests/test_webcrt.py
@@ -304,8 +304,8 @@ class WebcrtServiceTests(unittest.TestCase):
         self.assertTrue(called_creds["hop_enabled"])
         self.assertEqual(called_creds["hop_vendor"], "bastion")
         self.assertIn("session_log", mock_open.call_args.kwargs)
-        self.assertEqual(mock_open.call_args.kwargs.get("keepalive"), 0)
-        self.assertEqual(out.get("keepalive_sec"), 0)
+        self.assertEqual(mock_open.call_args.kwargs.get("keepalive"), 30)
+        self.assertEqual(out.get("keepalive_sec"), 30)
         fake.remote_conn.resize_pty.assert_called()
         self.assertEqual(out["ne_id"], "ne-hop")
         self.assertFalse(out.get("cli_hop"))  # bastion hop is not vendor CLI hop guard
@@ -867,8 +867,8 @@ class WebcrtServiceTests(unittest.TestCase):
         self.assertIn("filename=", dispo)
         self.assertIn("filename*=UTF-8''", dispo)
         self.assertNotIn("\n", dispo)
-        self.assertEqual(int(cfg.settings.webcrt_sftp_max_file_bytes), 512 * 1024 * 1024)
-        self.assertEqual(int(cfg.settings.webcrt_sftp_list_max_entries), 5000)
+        self.assertEqual(int(cfg.settings.webcrt_sftp_max_file_bytes), 256 * 1024 * 1024)
+        self.assertEqual(int(cfg.settings.webcrt_sftp_list_max_entries), 2000)
 
     def test_sftp_rename_rejects_bad_paths(self) -> None:
         from unittest.mock import MagicMock