feat(issue-5): auto-completion gate and durable pending notification

The gate does not re-decide anything. Phase 2a already decides, fail-closed and
against contract-pinned vectors; this supplies real evidence to that decision and
commits only if three independent gates agree: the global switch is on, the
backend has observed evidence that it reaches a quiescent state, and the job
itself opted in. Default is off.

The backend allowlist is evidence-based rather than optimistic. Inferring
"finished" from stillness requires stillness to have been observed, and only
three of six backends have that. An operator may narrow the list; a name that
was never observed is dropped rather than trusted, so it cannot be widened by
configuration.

Everything else holds. An unreadable manifest, a report rewritten after
finalization, missing evidence for a declared side effect, a policy that raises:
all refuse. "Could not decide" is never "done". A late explicit terminal wins,
because the boundary refuses the inferred one -- that is the boundary working,
not an error in the gate.

Notification intent is committed in the same atomic write as the state, so a
crash between committing and publishing leaves a record that still says the
world has not been told. Recovery re-sends; duplicates are already dropped by
sequence. A failed send deliberately leaves the intent pending, because an
unsent notification that looks sent is worse than one that is retried.

Two fixes this surfaced. Settling used a request that terminal records refuse,
which made an owed notification impossible to ever settle; settling is now a
status-free bookkeeping request, valid in any state. And the phase 2a guard
against shell references was matching hashlib.sha256 as a substring, so it now
requires a word boundary.

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

tests/test_issue5_phase2b_autocomplete.py

@@ -271,3 +271,249 @@ def test_publish_event_reports_a_bad_side_effect_instead_of_publishing(mam_sandb
         "--finalize-report", str(path), "--side-effect", "malformed",
     ])
     assert rc == 1
+
+
+# ==============================================================================
+# 3. autocomplete — the gate, and everything that must stop it
+# ==============================================================================
+
+def _ready_job(tmp_path, mam_sandbox, *, backend="claude-code", opt_in=True,
+               artifacts=None, body="report\n"):
+    """A job in exactly the state that should auto-complete, plus its evidence."""
+    from lib_py.loop import finalize, store
+
+    reg = _reg(tmp_path)
+    job_id = _make_job(reg, agent=backend, expected_artifacts=artifacts or [])
+    from lib_py.delegation import engine
+    engine.transition(job_id, "claim", registry_dir=reg)
+    if opt_in:
+        engine.transition(job_id, "progress", registry_dir=reg,
+                          fields={"auto_complete": True})
+
+    scenario = _scenario()
+    dispatch = scenario["dispatch_context"]
+    dispatch["job_id"] = job_id
+    for item in scenario["timeline"]:
+        env = item.get("collector_attribution")
+        if env:
+            env["job_id"] = job_id
+        # A non-herdr payload is a legitimate identifier source, so it has to
+        # agree too. Leaving the fixture's original id here is a contradiction,
+        # and the policy correctly refuses it.
+        payload = item.get("payload") or {}
+        if item.get("source_kind") != "herdr" and "job_id" in payload:
+            payload["job_id"] = job_id
+    store.record_dispatch(job_id, dispatch["attempt_id"], dispatch["agent_session_uuid"],
+                          dispatch["pane_process_generation"], workspace=str(mam_sandbox))
+    store.mark_execution_anchor(job_id, workspace=str(mam_sandbox))
+
+    report = _report(tmp_path, body)
+    finalize.finalize_report(report, job_id, dispatch["attempt_id"],
+                             side_effects=finalize.parse_side_effects(
+                                 ["{}=ref".format(a) for a in (artifacts or [])]))
+    return job_id, reg, scenario, report.parent
+
+
+def test_gate_commits_when_every_condition_holds(mam_sandbox, tmp_path, monkeypatch):
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox)
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed, result.reason
+    record = _load(reg, job_id)
+    assert record["status"] == "completed"
+    assert record["completed_by"] == "auto_complete", "an inferred terminal must be marked"
+    assert record["pending_notification"]["event"] == "completed"
+
+
+def test_g4_default_off_commits_nothing_and_leaves_the_registry_untouched(
+    mam_sandbox, tmp_path, monkeypatch
+):
+    """G4: with the flag unset, the registry must not change at all."""
+    monkeypatch.chdir(mam_sandbox)
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox)
+    before = _fingerprint(Path(reg))
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False
+    assert "disabled" in result.reason
+    assert _fingerprint(Path(reg)) == before, "the default must write nothing"
+
+
+@pytest.mark.parametrize("backend", ["agy", "hermes", "opencode", "unknown-agent"])
+def test_g9_a_backend_without_quiescence_evidence_never_auto_completes(
+    mam_sandbox, tmp_path, monkeypatch, backend
+):
+    """Inferring 'finished' from stillness needs stillness to have been observed."""
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox, backend=backend)
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False and "quiescent-state evidence" in result.reason
+    assert _load(reg, job_id)["status"] == "running"
+
+
+def test_the_allowlist_can_be_narrowed_but_never_widened(monkeypatch):
+    from lib_py.loop import autocomplete
+
+    monkeypatch.setenv("MAM_AUTO_COMPLETE_BACKENDS", "claude")
+    assert autocomplete.allowed_backends() == ("claude",)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE_BACKENDS", "claude,hermes,opencode")
+    assert autocomplete.allowed_backends() == ("claude",), "unobserved backends are dropped"
+
+
+def test_a_job_that_did_not_opt_in_is_not_auto_completed(mam_sandbox, tmp_path, monkeypatch):
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox, opt_in=False)
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False and "opt in" in result.reason
+
+
+def test_g5_a_rewritten_report_stops_the_gate(mam_sandbox, tmp_path, monkeypatch):
+    """G5: the manifest is checked against the file, so a late edit disqualifies."""
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox)
+    (report_dir / "report-final.md").write_text("rewritten after finalization\n", encoding="utf-8")
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False and "changed after finalization" in result.reason
+    assert _load(reg, job_id)["status"] == "running"
+
+
+def test_g5_missing_side_effect_evidence_stops_the_gate(mam_sandbox, tmp_path, monkeypatch):
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete, finalize
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox,
+                                                   artifacts=["git_push"])
+    # re-declare without the evidence the job said it would produce
+    finalize.finalize_report(report_dir / "report-final.md", job_id, 1, side_effects=[])
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False and "git_push" in result.reason
+
+
+def test_d7_a_policy_that_cannot_decide_holds(mam_sandbox, tmp_path, monkeypatch):
+    """'Could not decide' is never 'done'."""
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.loop import autocomplete, policy
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox)
+
+    def explode(_scenario):
+        raise RuntimeError("evidence is contradictory")
+
+    monkeypatch.setattr(policy, "decide", explode)
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False and "policy error" in result.reason
+    assert _load(reg, job_id)["status"] == "running"
+
+
+def test_a_late_explicit_terminal_beats_the_gate(mam_sandbox, tmp_path, monkeypatch):
+    """If the agent already reported, the boundary refuses and that is correct."""
+    monkeypatch.chdir(mam_sandbox)
+    monkeypatch.setenv("MAM_AUTO_COMPLETE", "1")
+    from lib_py.delegation import mqtt_common
+    from lib_py.loop import autocomplete
+
+    job_id, reg, scenario, report_dir = _ready_job(tmp_path, mam_sandbox)
+    mqtt_common.update_job_status(job_id, reg, _source="publish_event", status="error")
+    result = autocomplete.evaluate(job_id, scenario, registry_dir=reg,
+                                   workspace=str(mam_sandbox), report_dir=report_dir)
+    assert result.committed is False
+    record = _load(reg, job_id)
+    assert record["status"] == "error", "the explicit report wins"
+    assert "completed_by" not in record
+
+
+# ==============================================================================
+# 4. notifier — the intent is durable, and recovery re-sends it
+# ==============================================================================
+
+def test_the_notification_intent_is_committed_with_the_state(mam_sandbox, tmp_path, monkeypatch):
+    monkeypatch.chdir(mam_sandbox)
+    from lib_py.delegation import engine, notifier
+
+    reg = _reg(tmp_path)
+    job_id = _make_job(reg)
+    engine.transition(job_id, "claim", registry_dir=reg)
+    engine.transition(job_id, "completed", registry_dir=reg, notify="completed")
+    record = _load(reg, job_id)
+    assert record["status"] == "completed"
+    intent = notifier.pending(job_id, reg)
+    assert intent and intent["event"] == "completed" and intent["seq"] == record["transition_seq"]
+
+
+def test_flush_publishes_then_settles(mam_sandbox, tmp_path, monkeypatch):
+    monkeypatch.chdir(mam_sandbox)
+    from lib_py.delegation import engine, notifier
+
+    reg = _reg(tmp_path)
+    job_id = _make_job(reg)
+    engine.transition(job_id, "claim", registry_dir=reg)
+    engine.transition(job_id, "completed", registry_dir=reg, notify="completed")
+
+    sent = []
+    assert notifier.flush(job_id, reg, publish=lambda j, i, r: sent.append(i) or 0)
+    assert len(sent) == 1 and sent[0]["event"] == "completed"
+    assert notifier.pending(job_id, reg) is None
+    # settling is idempotent: a second flush has nothing to do
+    assert notifier.flush(job_id, reg, publish=lambda *a: 1 / 0)
+
+
+def test_a_failed_send_leaves_the_intent_pending(mam_sandbox, tmp_path, monkeypatch):
+    """An unsent notification that looks sent is worse than one that is retried."""
+    monkeypatch.chdir(mam_sandbox)
+    from lib_py.delegation import engine, notifier
+
+    reg = _reg(tmp_path)
+    job_id = _make_job(reg)
+    engine.transition(job_id, "claim", registry_dir=reg)
+    engine.transition(job_id, "completed", registry_dir=reg, notify="completed")
+
+    assert notifier.flush(job_id, reg, publish=lambda *a: 2) is False
+    assert notifier.pending(job_id, reg) is not None
+    assert notifier.flush(job_id, reg, publish=lambda *a: (_ for _ in ()).throw(OSError())) is False
+    assert notifier.pending(job_id, reg) is not None
+    assert _load(reg, job_id)["status"] == "completed", "the state stays committed regardless"
+
+
+def test_g7_restart_recovery_resends_every_owed_notification(mam_sandbox, tmp_path, monkeypatch):
+    """G7: the crash window between committing and publishing."""
+    monkeypatch.chdir(mam_sandbox)
+    from lib_py.delegation import engine, notifier
+
+    reg = _reg(tmp_path)
+    owed = []
+    for _ in range(3):
+        job_id = _make_job(reg)
+        engine.transition(job_id, "claim", registry_dir=reg)
+        engine.transition(job_id, "completed", registry_dir=reg, notify="completed")
+        owed.append(job_id)
+    settled = _make_job(reg)
+    engine.transition(settled, "claim", registry_dir=reg)
+
+    assert sorted(notifier.jobs_with_pending(reg)) == sorted(owed)
+
+    sent = []
+    results = notifier.recover(reg, publish=lambda j, i, r: sent.append(j) or 0)
+    assert all(results.values()) and sorted(sent) == sorted(owed)
+    assert notifier.jobs_with_pending(reg) == [], "nothing is owed after recovery"