feat(issue-5): persistent MQTT publisher prototype and benchmark harness

Splits publish_event.main into three pieces with no behaviour change:

  prepare_event   load the job, consume the seq, build the signed payload,
                  decide retain. The seq is burned here, before any transport
                  runs, so a failed publish can never leave a number a replayed
                  packet could reuse.
  record_outcome  audit record, registry event log and status sync -- all three
                  best effort, all three regardless of the publish result. That
                  is exactly the contract G-1/G-2/G-4 pin down, now in one place.
  main            parse, prepare, publish with retry, record.

publisher.py holds the two transports behind one interface. OneShotPublisher is
the current behaviour. PersistentPublisher keeps a single connection and lets
paho reconnect underneath, raising on an unacknowledged publish so the caller's
retry policy stays the only thing deciding attempt counts. It refuses an event
for a different broker rather than publishing to the wrong place.

Nothing in production uses the persistent path: the CLI and shell helpers exit
after one publish. The consumer is the phase 2a engine, which pays the CONNACK
cost on every event today -- this job's own started event took ~12.5s because
of one CONNACK timeout plus reconnect.

bench.py measures both arms. The fake transport is deterministic and is what
the tests assert on; the real-broker arm is informational, since the PoC
default is a public broker whose latency is not ours to control.

Verified: N publishes cost N connections one-shot and 1 persistent; the two
transports leave field-for-field identical registry and audit records including
the HMAC signature and retain flag; an unacknowledged publish retries on the
same connection and records published=false; a second broker is refused. No
test opens a socket.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd

.agents/skills/lib_py/delegation/bench.py

@@ -0,0 +1,178 @@
+"""Benchmark harness for the persistent-connection prototype.
+
+Two transports, deliberately separated:
+
+  --transport fake    a stub client with a synthetic connect latency. Fully
+                      deterministic, touches no network, and is what CI asserts
+                      on: N publishes cost N connections one-shot and exactly 1
+                      persistent.
+  --transport broker  the real broker named by the job record. Informational
+                      only: the PoC default is a public broker whose CONNACK
+                      latency is not ours to control, so no gate depends on it.
+
+Usage:
+  python -m lib_py.delegation.bench --n 50 --transport fake [--connect-latency-ms 50]
+                                    [--job <id>] [--registry-dir DIR] [--out FILE]
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+import time
+from typing import List, Optional
+
+from lib_py.delegation import mqtt_common, publish_event
+from lib_py.delegation.publisher import OneShotPublisher, PersistentPublisher
+
+
+class FakeInfo(object):
+    def __init__(self, published=True):
+        self._published = published
+
+    def wait_for_publish(self, timeout=None):
+        return None
+
+    def is_published(self):
+        return self._published
+
+
+class FakeClient(object):
+    """Minimal paho stand-in: counts connects and charges a synthetic latency."""
+
+    def __init__(self, connect_latency=0.0, counters=None):
+        self.connect_latency = connect_latency
+        self.counters = counters if counters is not None else {}
+        self.on_connect = None
+
+    def connect(self, host, port, keepalive=60):
+        if self.connect_latency:
+            time.sleep(self.connect_latency)
+        self.counters["connect"] = self.counters.get("connect", 0) + 1
+        if self.on_connect:
+            self.on_connect(self, None, None, 0, None)
+
+    def loop_start(self):
+        return None
+
+    def loop_stop(self):
+        return None
+
+    def reconnect_delay_set(self, min_delay=1, max_delay=16):
+        return None
+
+    def publish(self, topic, payload=None, qos=0, retain=False):
+        self.counters["publish"] = self.counters.get("publish", 0) + 1
+        return FakeInfo()
+
+    def disconnect(self):
+        return None
+
+
+def fake_factory(connect_latency, counters):
+    def factory(_role, _config):
+        return FakeClient(connect_latency=connect_latency, counters=counters)
+    return factory
+
+
+def _percentile(values: List[float], pct: float) -> float:
+    if not values:
+        return 0.0
+    ordered = sorted(values)
+    idx = int(round((pct / 100.0) * (len(ordered) - 1)))
+    return ordered[idx]
+
+
+def _run(publisher, prepared, n: int) -> dict:
+    latencies = []
+    failures = 0
+    for _ in range(n):
+        start = time.monotonic()
+        ok, _err = publisher.publish_with_retry(prepared, attempts=1)
+        latencies.append((time.monotonic() - start) * 1000.0)
+        if not ok:
+            failures += 1
+    return {
+        "publishes": n,
+        "connects": getattr(publisher, "connect_count", None),
+        "failures": failures,
+        "latency_ms": {
+            "p50": round(_percentile(latencies, 50), 3),
+            "p95": round(_percentile(latencies, 95), 3),
+            "max": round(max(latencies), 3) if latencies else 0.0,
+        },
+    }
+
+
+def run_benchmark(n=50, transport="fake", connect_latency_ms=50.0,
+                  job_id=None, registry_dir=None) -> dict:
+    registry_dir = registry_dir or mqtt_common.DEFAULT_REGISTRY_DIR
+
+    if transport == "fake":
+        prepared = publish_event.PreparedEvent(
+            job_id="benchjob", event="progress", detail="bench", seq=0,
+            topic="python/mqtt/jobs/benchjob/events", retain=False,
+            payload={"seq": 0}, body=b'{"seq": 0}',
+            config=mqtt_common.broker_config_from_env(),
+        )
+        latency = connect_latency_ms / 1000.0
+        one_counters, persist_counters = {}, {}
+        one = OneShotPublisher(client_factory=fake_factory(latency, one_counters))
+        persist = PersistentPublisher(client_factory=fake_factory(latency, persist_counters))
+        try:
+            result = {
+                "transport": "fake",
+                "connect_latency_ms": connect_latency_ms,
+                "one_shot": _run(one, prepared, n),
+                "persistent": _run(persist, prepared, n),
+            }
+        finally:
+            persist.close()
+        result["one_shot"]["client_connects"] = one_counters.get("connect", 0)
+        result["persistent"]["client_connects"] = persist_counters.get("connect", 0)
+        return result
+
+    if not job_id:
+        raise ValueError("--transport broker requires --job")
+    prepared = publish_event.prepare_event(
+        job_id=job_id, event="progress", detail="bench", registry_dir=registry_dir,
+    )
+    one = OneShotPublisher()
+    persist = PersistentPublisher()
+    try:
+        return {
+            "transport": "broker",
+            "broker": "{}:{}".format(prepared.config.host, prepared.config.port),
+            "one_shot": _run(one, prepared, n),
+            "persistent": _run(persist, prepared, n),
+        }
+    finally:
+        persist.close()
+
+
+def main(argv: Optional[List[str]] = None) -> int:
+    mqtt_common.ensure_env_loaded()
+    parser = argparse.ArgumentParser(prog="lib_py.delegation.bench")
+    parser.add_argument("--n", type=int, default=50)
+    parser.add_argument("--transport", default="fake", choices=["fake", "broker"])
+    parser.add_argument("--connect-latency-ms", type=float, default=50.0)
+    parser.add_argument("--job", default=None)
+    parser.add_argument("--registry-dir", default=None)
+    parser.add_argument("--out", default=None)
+    args = parser.parse_args(argv)
+
+    result = run_benchmark(
+        n=args.n, transport=args.transport,
+        connect_latency_ms=args.connect_latency_ms,
+        job_id=args.job, registry_dir=args.registry_dir,
+    )
+    text = json.dumps(result, indent=2, sort_keys=True)
+    if args.out:
+        with open(args.out, "w", encoding="utf-8") as fh:
+            fh.write(text + "\n")
+    print(text)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

.agents/skills/lib_py/delegation/publish_event.py

@@ -124,6 +124,85 @@ def _publish_once(config, topic: str, body: bytes, retain: bool) -> None:
             pass
 
 
+class PreparedEvent(object):
+    """Everything decided before the wire: payload, topic, broker, retain flag.
+
+    Splitting this out lets a different transport (one-shot vs persistent) be
+    swapped in without touching the parts that must not vary: the seq is already
+    consumed, the HMAC is already computed over the final payload, and the audit
+    record is written from these exact fields.
+    """
+
+    __slots__ = ("job_id", "event", "detail", "seq", "topic", "retain",
+                 "payload", "body", "config")
+
+    def __init__(self, job_id, event, detail, seq, topic, retain, payload, body, config):
+        self.job_id = job_id
+        self.event = event
+        self.detail = detail
+        self.seq = seq
+        self.topic = topic
+        self.retain = retain
+        self.payload = payload
+        self.body = body
+        self.config = config
+
+
+def prepare_event(job_id, event, detail="", data=None,
+                  registry_dir=DEFAULT_REGISTRY_DIR, retained=False) -> PreparedEvent:
+    """Load the job, consume the next seq, and build the signed payload.
+
+    Consuming the seq here (before any transport runs) is deliberate: a failed
+    publish must still burn the number, so a replayed packet can never carry a
+    seq the subscriber would accept.
+    """
+    job = load_job(job_id, registry_dir)
+    config = broker_config_from_job(job)
+    topic = job.get("topic_prefix")
+    topic = f"{topic}/events" if topic else events_topic_for(job_id)
+    seq = next_seq(job_id, registry_dir)
+    payload = build_payload(
+        job_id=job_id,
+        seq=seq,
+        event=event,
+        detail=detail,
+        data=data,
+        auth_token=job.get("auth_token"),
+    )
+    body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
+    retain = retained or event in TERMINAL_EVENTS
+    return PreparedEvent(job_id, event, detail, seq, topic, retain, payload, body, config)
+
+
+def record_outcome(prepared: PreparedEvent, registry_dir: str,
+                   publish_ok: bool, publish_error: Optional[str]) -> None:
+    """Persist the audit record, the registry event log and the status sync.
+
+    All three are best effort and all three happen whether or not the publish
+    succeeded — that is the contract G-1/G-2/G-4 pin down.
+    """
+    mqtt_common.append_event(prepared.job_id, {
+        "event": "published",
+        "source_event": prepared.event,
+        "seq": prepared.seq,
+        "topic": prepared.topic,
+        "retain": prepared.retain,
+        "timestamp": prepared.payload["timestamp"],
+        "detail": prepared.detail,
+        "payload": prepared.payload,
+        "published": publish_ok,
+        "publish_error": publish_error,
+    })
+
+    registry.append_event(prepared.job_id, registry_dir, prepared.payload)
+    new_status = EVENT_TO_STATUS.get(prepared.event)
+    if new_status:
+        try:
+            mqtt_common.update_job_status(prepared.job_id, registry_dir, status=new_status)
+        except Exception as exc:  # pragma: no cover - best effort
+            logger.warning("status sync failed: %s", exc)
+
+
 def _resolve_job_id(args) -> Optional[str]:
     if args.pick_pending:
         return registry.pick_pending(args.agent_session, args.registry_dir)
@@ -168,26 +247,20 @@ def main(argv=None) -> int:
         return 1
 
     try:
-        job = load_job(job_id, args.registry_dir)
+        prepared = prepare_event(
+            job_id=job_id,
+            event=args.event,
+            detail=args.detail,
+            data=data,
+            registry_dir=args.registry_dir,
+            retained=args.retained,
+        )
     except FileNotFoundError as exc:
         logger.error("%s", exc)
         return 1
 
-    config = broker_config_from_job(job)
-    topic = job.get("topic_prefix")
-    topic = f"{topic}/events" if topic else events_topic_for(job_id)
-    seq = next_seq(job_id, args.registry_dir)
-    payload = build_payload(
-        job_id=job_id,
-        seq=seq,
-        event=args.event,
-        detail=args.detail,
-        data=data,
-        auth_token=job.get("auth_token"),
-    )
-    body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
-    retain = args.retained or args.event in TERMINAL_EVENTS
-
+    # Policy Note: seq consumption on failure is intentional (see prepare_event)
+    # and the failure is documented explicitly in the audit record.
     publish = with_retry(
         _publish_once,
         attempts=args.attempts,
@@ -196,43 +269,19 @@ def main(argv=None) -> int:
     publish_ok = True
     publish_error: Optional[str] = None
     try:
-        publish(config, topic, body, retain)
+        publish(prepared.config, prepared.topic, prepared.body, prepared.retain)
     except Exception as exc:
         publish_ok = False
         publish_error = str(exc)
         logger.error("publish failed after %d attempts: %s", args.attempts, exc)
 
-    # Persistent audit log: record the exact payload we put on the wire (or intended to).
-    # Best-effort (isolated inside append_event) — never fails the publish.
-    # Policy Note: Seq consumption on failure is intentionally maintained for monotonic replay
-    # defense (> highest accepted seq), with failure documented explicitly in the audit record.
-    mqtt_common.append_event(job_id, {
-        "event": "published",
-        "source_event": args.event,
-        "seq": seq,
-        "topic": topic,
-        "retain": retain,
-        "timestamp": payload["timestamp"],
-        "detail": args.detail,
-        "payload": payload,
-        "published": publish_ok,
-        "publish_error": publish_error,
-    })
-
-    # Best-effort side effects: registry status sync + (debug) event log. Never
-    # fail the publish on these.
-    registry.append_event(job_id, args.registry_dir, payload)
-    new_status = EVENT_TO_STATUS.get(args.event)
-    if new_status:
-        try:
-            mqtt_common.update_job_status(job_id, args.registry_dir, status=new_status)
-        except Exception as exc:  # pragma: no cover - best effort
-            logger.warning("status sync failed: %s", exc)
+    record_outcome(prepared, args.registry_dir, publish_ok, publish_error)
 
     if not publish_ok:
         return 2
 
-    logger.info("published %s seq=%d job=%s retain=%s", args.event, seq, job_id, retain)
+    logger.info("published %s seq=%d job=%s retain=%s",
+                args.event, prepared.seq, job_id, prepared.retain)
     return 0
 
 

.agents/skills/lib_py/delegation/publisher.py

@@ -0,0 +1,169 @@
+"""Publish transports: one connection per event, or one connection reused.
+
+Phase 1 ships the persistent transport as a prototype only. Nothing in
+production uses it: the CLI and the shell helpers exit after a single publish,
+so reusing a connection buys them nothing. The consumer is the phase 2a engine,
+which is long-running and pays the CONNACK cost on every event today.
+
+Both transports raise on failure so the caller's `with_retry` policy stays the
+single place that decides how many attempts an event gets.
+"""
+from __future__ import annotations
+
+import logging
+import time
+from typing import Optional, Tuple
+
+from lib_py.delegation import mqtt_common
+
+logger = logging.getLogger("delegate_job.publisher")
+
+CONNECT_ACK_TIMEOUT = 10  # seconds to wait for CONNACK
+PUBLISH_ACK_TIMEOUT = 5   # seconds to wait for QoS-1 PUBACK
+
+
+class _BasePublisher(object):
+    """Shared retry wrapper so both transports report outcomes identically."""
+
+    def publish(self, prepared):  # pragma: no cover - interface
+        raise NotImplementedError
+
+    def publish_with_retry(self, prepared, attempts=3) -> Tuple[bool, Optional[str]]:
+        runner = mqtt_common.with_retry(
+            self.publish,
+            attempts=attempts,
+            exceptions=(OSError, TimeoutError, ConnectionError, ValueError),
+        )
+        try:
+            runner(prepared)
+        except Exception as exc:
+            logger.error("publish failed after %d attempts: %s", attempts, exc)
+            return False, str(exc)
+        return True, None
+
+    def open(self):
+        return self
+
+    def close(self):
+        return None
+
+
+class OneShotPublisher(_BasePublisher):
+    """Connect, publish, wait for the ACK, disconnect. The current behaviour."""
+
+    def __init__(self, client_factory=None):
+        self.client_factory = client_factory or mqtt_common.make_client
+        self.connect_count = 0
+
+    def publish(self, prepared):
+        client = self.client_factory("publisher", prepared.config)
+        self.connect_count += 1
+        connected = {"rc": None}
+
+        def on_connect(_c, _u, _flags, reason_code, _props):
+            connected["rc"] = reason_code
+
+        client.on_connect = on_connect
+        client.connect(prepared.config.host, prepared.config.port, prepared.config.keepalive)
+        client.loop_start()
+        try:
+            deadline = time.monotonic() + CONNECT_ACK_TIMEOUT
+            while connected["rc"] is None and time.monotonic() < deadline:
+                time.sleep(0.05)
+            if connected["rc"] is None:
+                raise TimeoutError("no CONNACK from broker")
+            if mqtt_common.reason_code_value(connected["rc"]) != 0:
+                raise ConnectionError("broker refused connection: rc={}".format(connected["rc"]))
+            info = client.publish(prepared.topic, payload=prepared.body,
+                                  qos=1, retain=prepared.retain)
+            info.wait_for_publish(timeout=PUBLISH_ACK_TIMEOUT)
+            if not info.is_published():
+                raise TimeoutError("publish not acknowledged within timeout")
+        finally:
+            client.loop_stop()
+            try:
+                client.disconnect()
+            except Exception:  # pragma: no cover - disconnect best effort
+                pass
+
+
+class PersistentPublisher(_BasePublisher):
+    """One connection reused across events; paho handles reconnect underneath.
+
+    Bound to a single broker: the connection is established from the first
+    event's config, so an event for a job on a different broker is refused
+    rather than silently published to the wrong place.
+    """
+
+    def __init__(self, client_factory=None):
+        self.client_factory = client_factory or mqtt_common.make_client
+        self.connect_count = 0
+        self._client = None
+        self._config = None
+        self._connected = {"rc": None}
+
+    def __enter__(self):
+        return self.open()
+
+    def __exit__(self, exc_type, exc, tb):
+        self.close()
+        return False
+
+    def open(self):
+        return self
+
+    def _ensure_client(self, config):
+        if self._client is not None:
+            if (config.host, config.port) != (self._config.host, self._config.port):
+                raise ValueError(
+                    "PersistentPublisher is bound to {}:{}; refusing an event for {}:{}".format(
+                        self._config.host, self._config.port, config.host, config.port)
+                )
+            return self._client
+
+        client = self.client_factory("publisher", config)
+        self.connect_count += 1
+
+        def on_connect(_c, _u, _flags, reason_code, _props):
+            self._connected["rc"] = reason_code
+
+        client.on_connect = on_connect
+        client.connect(config.host, config.port, config.keepalive)
+        client.loop_start()
+        reconnect_delay_set = getattr(client, "reconnect_delay_set", None)
+        if reconnect_delay_set is not None:
+            reconnect_delay_set(min_delay=1, max_delay=16)
+        deadline = time.monotonic() + CONNECT_ACK_TIMEOUT
+        while self._connected["rc"] is None and time.monotonic() < deadline:
+            time.sleep(0.05)
+        if self._connected["rc"] is None:
+            client.loop_stop()
+            raise TimeoutError("no CONNACK from broker")
+        if mqtt_common.reason_code_value(self._connected["rc"]) != 0:
+            client.loop_stop()
+            raise ConnectionError("broker refused connection: rc={}".format(self._connected["rc"]))
+        self._client = client
+        self._config = config
+        return client
+
+    def publish(self, prepared):
+        client = self._ensure_client(prepared.config)
+        info = client.publish(prepared.topic, payload=prepared.body,
+                              qos=1, retain=prepared.retain)
+        info.wait_for_publish(timeout=PUBLISH_ACK_TIMEOUT)
+        if not info.is_published():
+            # Surfaced to the caller's with_retry; paho reconnects underneath.
+            raise TimeoutError("publish not acknowledged within timeout")
+
+    def close(self):
+        if self._client is None:
+            return
+        try:
+            self._client.loop_stop()
+            self._client.disconnect()
+        except Exception:  # pragma: no cover - best effort
+            pass
+        finally:
+            self._client = None
+            self._config = None
+            self._connected = {"rc": None}

tests/test_issue5_phase1_delegation.py

@@ -720,3 +720,163 @@ def test_publish_failure_semantics_match_across_entry_points(mam_sandbox, entry)
     published = [e for e in events if e.get("event") == "published"]
     assert published and published[-1]["published"] is False
     assert published[-1]["publish_error"]
+
+
+# ==============================================================================
+# 6. P1-03 — transports, parity and the benchmark harness (no network at all)
+# ==============================================================================
+
+def _prepared(topic="python/mqtt/jobs/j/events", host="broker.example", port=1883):
+    from lib_py.delegation import mqtt_common, publish_event
+
+    config = mqtt_common.broker_config_from_env({"host": host, "port": port})
+    return publish_event.PreparedEvent(
+        job_id="j", event="progress", detail="d", seq=1, topic=topic, retain=False,
+        payload={"seq": 1}, body=b'{"seq": 1}', config=config,
+    )
+
+
+def _fake_factory(counters, publishes_ok=True, connect_rc=0):
+    from lib_py.delegation.bench import FakeClient, FakeInfo
+
+    def factory(_role, _config):
+        client = FakeClient(connect_latency=0.0, counters=counters)
+        if not publishes_ok:
+            def unacked(*_a, **_k):
+                counters["publish"] = counters.get("publish", 0) + 1
+                return FakeInfo(published=False)
+            client.publish = unacked
+        if connect_rc:
+            original = client.connect
+
+            def connect(host, port, keepalive=60):
+                original(host, port, keepalive)
+                client.on_connect(client, None, None, connect_rc, None)
+            client.connect = connect
+        return client
+    return factory
+
+
+def test_persistent_publisher_reuses_one_connection():
+    """The whole point of the prototype: N publishes, 1 connection."""
+    from lib_py.delegation.publisher import OneShotPublisher, PersistentPublisher
+
+    prepared = _prepared()
+    one_counters, persist_counters = {}, {}
+    one = OneShotPublisher(client_factory=_fake_factory(one_counters))
+    persist = PersistentPublisher(client_factory=_fake_factory(persist_counters))
+    try:
+        for _ in range(5):
+            assert one.publish_with_retry(prepared, attempts=1) == (True, None)
+            assert persist.publish_with_retry(prepared, attempts=1) == (True, None)
+    finally:
+        persist.close()
+    assert one_counters["connect"] == 5
+    assert persist_counters["connect"] == 1
+    assert one_counters["publish"] == persist_counters["publish"] == 5
+
+
+def test_transport_parity_of_registry_and_audit_records(mam_sandbox, monkeypatch):
+    """Same input, two transports: registry and audit records must match field for field."""
+    from lib_py.delegation import mqtt_common, publish_event, registry
+    from lib_py.delegation.publisher import OneShotPublisher, PersistentPublisher
+
+    monkeypatch.chdir(mam_sandbox)
+    reg_dir = str(Path(mam_sandbox) / ".mam" / "jobs")
+    os.makedirs(reg_dir, exist_ok=True)
+
+    results = {}
+    for label, publisher in (
+        ("one_shot", OneShotPublisher(client_factory=_fake_factory({}))),
+        ("persistent", PersistentPublisher(client_factory=_fake_factory({}))),
+    ):
+        job_id = registry.register_job("parity " + label, registry_dir=reg_dir)
+        prepared = publish_event.prepare_event(
+            job_id=job_id, event="completed", detail="done", registry_dir=reg_dir,
+        )
+        ok, err = publisher.publish_with_retry(prepared, attempts=1)
+        publish_event.record_outcome(prepared, reg_dir, ok, err)
+        if hasattr(publisher, "close"):
+            publisher.close()
+        record = mqtt_common.load_job(job_id, reg_dir)
+        events = list(mqtt_common.iter_logged_events(job_id, mqtt_common.get_logs_dir()))
+        audit = [e for e in events if e.get("event") == "published"][-1]
+        results[label] = {
+            "ok": ok,
+            "status": record["status"],
+            "last_seq": record.get("last_seq"),
+            "retain": audit["retain"],
+            "published": audit["published"],
+            "seq": audit["seq"],
+            "has_sig": "hmac_sig" in audit["payload"].get("data", {}),
+            "source_event": audit["source_event"],
+        }
+
+    assert results["one_shot"] == results["persistent"], results
+    assert results["one_shot"]["retain"] is True, "terminal events must retain"
+    assert results["one_shot"]["status"] == "completed"
+
+
+def test_persistent_publisher_unacked_publish_retries_then_records_failure(mam_sandbox, monkeypatch):
+    """No PUBACK -> exception -> with_retry -> record_outcome(published=False)."""
+    from lib_py.delegation import mqtt_common, publish_event, registry
+    from lib_py.delegation.publisher import PersistentPublisher
+
+    monkeypatch.chdir(mam_sandbox)
+    reg_dir = str(Path(mam_sandbox) / ".mam" / "jobs")
+    os.makedirs(reg_dir, exist_ok=True)
+    job_id = registry.register_job("unacked", registry_dir=reg_dir)
+    prepared = publish_event.prepare_event(
+        job_id=job_id, event="completed", detail="d", registry_dir=reg_dir,
+    )
+    counters = {}
+    publisher = PersistentPublisher(client_factory=_fake_factory(counters, publishes_ok=False))
+    try:
+        ok, err = publisher.publish_with_retry(prepared, attempts=3)
+    finally:
+        publisher.close()
+    assert ok is False and "not acknowledged" in err
+    assert counters["publish"] == 3, "every attempt must reach the transport"
+    assert counters["connect"] == 1, "retries reuse the connection"
+
+    publish_event.record_outcome(prepared, reg_dir, ok, err)
+    record = mqtt_common.load_job(job_id, reg_dir)
+    assert record["status"] == "completed", "status syncs even when the publish failed"
+    assert record.get("last_seq") == prepared.seq
+    audit = [e for e in mqtt_common.iter_logged_events(job_id, mqtt_common.get_logs_dir())
+             if e.get("event") == "published"][-1]
+    assert audit["published"] is False and audit["publish_error"]
+
+
+def test_persistent_publisher_refuses_a_second_broker():
+    from lib_py.delegation.publisher import PersistentPublisher
+
+    publisher = PersistentPublisher(client_factory=_fake_factory({}))
+    try:
+        assert publisher.publish_with_retry(_prepared(host="a.example"), attempts=1)[0] is True
+        ok, err = publisher.publish_with_retry(_prepared(host="b.example"), attempts=1)
+        assert ok is False
+        assert "bound to a.example" in err, err
+    finally:
+        publisher.close()
+
+
+def test_bench_fake_transport_schema_and_counts():
+    from lib_py.delegation import bench
+
+    result = bench.run_benchmark(n=7, transport="fake", connect_latency_ms=1.0)
+    assert result["transport"] == "fake"
+    for arm in ("one_shot", "persistent"):
+        assert result[arm]["publishes"] == 7
+        assert result[arm]["failures"] == 0
+        assert set(result[arm]["latency_ms"]) == {"p50", "p95", "max"}
+    assert result["one_shot"]["client_connects"] == 7
+    assert result["persistent"]["client_connects"] == 1
+    assert result["persistent"]["latency_ms"]["p50"] < result["one_shot"]["latency_ms"]["p50"]
+
+
+def test_bench_broker_transport_requires_a_job():
+    from lib_py.delegation import bench
+
+    with pytest.raises(ValueError):
+        bench.run_benchmark(n=1, transport="broker")