@team-agent/installer 0.2.10 → 0.3.0

package/crates/team-agent/src/messaging/tests/runtime.rs

@@ -0,0 +1,1422 @@
+use super::*;
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP E — _fail_leader_delivery: bug-52 fallback-log semantics. ok=True but
+// status=FallbackLog (NOT a real submit). leader.py:394-436.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn fail_leader_delivery_returns_fallback_log_ok_true_not_submitted() {
+    let ws = tmp_ws("faillead");
+    let payload = json(serde_json::json!({
+        "to": "leader", "content": "hi", "sender": "coordinator"
+    }));
+    let out = fail_leader_delivery(
+        &ws,
+        &payload,
+        DeliveryRefusal::LeaderNotAttached,
+        Some("No direct leader tmux pane is attached. Run team-agent attach-leader."),
+    )
+    .unwrap();
+    // leader.py:423-431 — ok True, status fallback_log, channel fallback_inbox.
+    assert!(out.ok);
+    assert_eq!(out.status, DeliveryStatus::FallbackLog);
+    assert_eq!(out.reason, Some(DeliveryRefusal::LeaderNotAttached));
+    // The audit must be distinguishable from a real submit (Delivered).
+    assert_ne!(out.status, DeliveryStatus::Delivered);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP F — session_drift_refusal: None-vs-refused fallthrough chain.
+// session_drift.py:69-91.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn session_drift_refusal_none_for_no_target_leader_or_broadcast() {
+    let ws = tmp_ws("driftnone");
+    let log = EventLog::new(&ws);
+    let state = json(serde_json::json!({"agents": {}}));
+    // target == leader_id → None (no refusal).
+    assert!(
+        session_drift_refusal(&state, "leader", "leader", "s", None, &log)
+            .unwrap()
+            .is_none()
+    );
+    // target == "*" (broadcast) → None.
+    assert!(session_drift_refusal(&state, "*", "leader", "s", None, &log)
+        .unwrap()
+        .is_none());
+}
+
+#[test]
+fn session_drift_refusal_none_when_status_not_drift() {
+    let ws = tmp_ws("driftok");
+    let log = EventLog::new(&ws);
+    let state = json(serde_json::json!({"agents": {"w1": {"status": "idle"}}}));
+    assert!(session_drift_refusal(&state, "w1", "leader", "s", None, &log)
+        .unwrap()
+        .is_none());
+}
+
+#[test]
+fn session_drift_refusal_refuses_when_agent_in_drift() {
+    // session_drift.py:84-91 → ok False, reason session_drift, action reset-agent.
+    let ws = tmp_ws("driftrefuse");
+    let log = EventLog::new(&ws);
+    let state = json(serde_json::json!({
+        "agents": {"w1": {"status": "session_drift",
+            "session_drift": {"stored_session_id": "S", "actual_thread_id": "A"}}}
+    }));
+    let out = session_drift_refusal(&state, "w1", "leader", "leader", None, &log)
+        .unwrap()
+        .expect("drift agent must be refused");
+    assert!(!out.ok);
+    assert_eq!(out.status, DeliveryStatus::Refused);
+    assert_eq!(out.reason, Some(DeliveryRefusal::SessionDrift));
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP G — classify_agent_activity: every branch incl. the uncertain
+// fallthrough iron law. activity_detector.py:90-146 (golden probed).
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn classify_pane_in_mode_is_uncertain_high_confidence() {
+    let state = json(serde_json::json!({}));
+    let a = classify_agent_activity(&state, "", true, None, None);
+    assert_eq!(a.status, ActivityStatus::Uncertain);
+    assert_eq!(a.confidence, 0.9);
+}
+
+#[test]
+fn classify_idle_prompt_is_idle() {
+    // "❯ \n" matches the Claude idle prompt → idle 0.9.
+    let state = json(serde_json::json!({}));
+    let a = classify_agent_activity(&state, "❯ \n", false, None, None);
+    assert_eq!(a.status, ActivityStatus::Idle);
+    assert_eq!(a.confidence, 0.9);
+}
+
+#[test]
+fn classify_working_indicator_is_working() {
+    let state = json(serde_json::json!({}));
+    let a = classify_agent_activity(&state, "Working (5s)", false, None, None);
+    assert_eq!(a.status, ActivityStatus::Working);
+    assert_eq!(a.confidence, 0.9);
+}
+
+#[test]
+fn classify_stale_working_is_stuck() {
+    // elapsed >= stuck_timeout (300) → stuck 0.85.
+    let state = json(serde_json::json!({}));
+    let a = classify_agent_activity(&state, "Working (400s)", false, None, None);
+    assert_eq!(a.status, ActivityStatus::Stuck);
+    assert_eq!(a.confidence, 0.85);
+}
+
+#[test]
+fn classify_no_signal_is_uncertain_never_idle() {
+    // THE IRON LAW: no decisive prompt/working signal → uncertain 0.5, NOT idle.
+    let state = json(serde_json::json!({}));
+    let a = classify_agent_activity(&state, "random prose nothing", false, None, None);
+    assert_eq!(a.status, ActivityStatus::Uncertain);
+    assert_eq!(a.confidence, 0.5);
+    assert_ne!(a.status, ActivityStatus::Idle);
+}
+
+#[test]
+fn classify_recent_provider_output_is_working_low_confidence() {
+    // age <= 120 with provider/no command → working 0.7.
+    let state = json(serde_json::json!({}));
+    let now = chrono::Utc::now();
+    let recent = (now - chrono::Duration::seconds(30)).to_rfc3339();
+    let a = classify_agent_activity(&state, "prose", false, None, Some(&recent));
+    assert_eq!(a.status, ActivityStatus::Working);
+    assert_eq!(a.confidence, 0.7);
+}
+
+// STAGE-B REGRESSION RED (dispatch-to-just-launched-agent → deferred_busy never closes the round-trip).
+// golden activity_detector.py (classify_agent_activity): the provider IDLE PROMPT is checked FIRST as a
+// scrollback-position signal (C14, "provider idle prompt is the latest scrollback signal" → idle 0.9),
+// BEFORE the `age<=120 → working 0.7` recent-output branch (:56). Rust (activity.rs:192) fires
+// `recent_rfc3339(last_output_at,120) → Working` BEFORE `latest_prompt_signal` (:200), so a just-launched
+// agent (startup banner = recent output, but pane shows the idle prompt awaiting input) mis-classifies
+// WORKING → sync_agent_health writes agent_health=WORKING → recipient_is_busy → send.deferred_busy.
+// Golden evidence (probe): classify(idle-prompt scrollback, recent last_output_at) = idle 0.9 regardless
+// of active_task. FIX = reorder: latest_prompt_signal (idle/working scrollback-position) BEFORE the
+// last_output_at age block.
+#[test]
+fn classify_idle_prompt_beats_recent_output_for_just_launched_agent() {
+    let state = json(serde_json::json!({}));
+    let recent = chrono::Utc::now().to_rfc3339();
+    let a = classify_agent_activity(&state, "codex ready\n❯ \n", false, Some("codex"), Some(&recent));
+    assert_eq!(
+        a.status,
+        ActivityStatus::Idle,
+        "just-launched agent showing the idle prompt must classify IDLE (golden idle-prompt-position is the \
+         latest scrollback signal, checked before the age<=120 recent-output branch), not WORKING because \
+         the startup banner is recent (activity.rs:192 recent-output fires before latest_prompt_signal:200) \
+         — the Stage B dispatch deferred_busy regression. got {a:?}"
+    );
+    assert_eq!(a.confidence, 0.9, "golden idle-prompt confidence is 0.9; got {a:?}");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP H — attempt_trust_auto_answer: own-vs-foreign realpath + fail-safe
+// pane-width + opt-in gate + reason byte-locks. leader_panes.py:383-470.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn trust_auto_answer_pane_id_missing_reason() {
+    // leader_panes.py:417-424 — pane_id None → pane_id_missing (after opt-in).
+    let ws = tmp_ws("trustpane");
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+    let out = attempt_trust_auto_answer(
+        &ws,
+        &t,
+        None,
+        "some prompt",
+        &PaneWidthQuery::Ok { pane_width: 120 },
+        &log,
+    )
+    .unwrap();
+    assert!(!out.ok);
+    assert!(!out.answered);
+    assert_eq!(out.reason, "pane_id_missing");
+}
+
+#[test]
+fn trust_auto_answer_foreign_workspace_refused() {
+    // leader_panes.py:430-444 — prompt names a FOREIGN dir → workspace_dir_mismatch,
+    // action prompt_leader. (own-vs-foreign realpath gate.)
+    let ws = tmp_ws("trustforeign");
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+    let pane = PaneId::new("%7");
+    let foreign_tail = "Allow Codex to access /some/other/foreign/dir ?";
+    let out = attempt_trust_auto_answer(
+        &ws,
+        &t,
+        Some(&pane),
+        foreign_tail,
+        &PaneWidthQuery::Ok { pane_width: 120 },
+        &log,
+    )
+    .unwrap();
+    assert!(!out.answered);
+    assert_eq!(out.reason, "workspace_dir_mismatch");
+    assert_eq!(out.action.as_deref(), Some("prompt_leader"));
+}
+
+#[test]
+fn trust_auto_answer_own_workspace_realpath_equal_answers() {
+    // Exact canonical equality of the prompt path with the workspace → answered.
+    let ws = tmp_ws("trustown");
+    let canonical = std::fs::canonicalize(&ws).unwrap();
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+    let pane = PaneId::new("%7");
+    let tail = format!("Allow Codex to write to {} ?", canonical.display());
+    let out = attempt_trust_auto_answer(
+        &canonical,
+        &t,
+        Some(&pane),
+        &tail,
+        &PaneWidthQuery::Ok { pane_width: 240 },
+        &log,
+    )
+    .unwrap();
+    assert!(out.answered, "own-workspace realpath-equal prompt must auto-answer");
+    assert_eq!(out.reason, "trust_auto_answered");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP I — PaneWidthQuery fail-safe (bug-064/082): Failed NEVER carries a
+// default width; tmux_pane_width returns Failed on any query failure.
+// delivery.py:20-51.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn pane_width_failed_forces_exact_match_never_default() {
+    // GROUP-I fail-safe (bug-064/082, folded from the old structural placeholder):
+    // calls the REAL attempt_trust_auto_answer. A FOREIGN path that is merely a
+    // truncated PREFIX of this workspace would only match with a width signal that
+    // proves right-edge truncation (leader_panes.py:_token_reaches_right_edge). With
+    // PaneWidthQuery::Failed there is NO width signal and NO default width to leak,
+    // so the matcher MUST fall back to exact canonical equality → the prefix does
+    // NOT match → workspace_dir_mismatch / prompt_leader. Probed golden:
+    // leader_panes.py:430-444 with pane_width=None (Failed) → workspace_dir_mismatch.
+    let ws = tmp_ws("panewidthfailsafe");
+    let canonical = std::fs::canonicalize(&ws).unwrap();
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+    let pane = PaneId::new("%7");
+    // A right-edge-truncated prefix of the real workspace path (drop the last char):
+    // would auto-answer IF a width signal proved truncation — but Failed forbids that.
+    let canon_str = canonical.to_string_lossy();
+    let truncated_prefix = &canon_str[..canon_str.len().saturating_sub(1)];
+    let tail = format!("Allow Codex to write to {truncated_prefix}");
+    let out = attempt_trust_auto_answer(
+        &canonical,
+        &t,
+        Some(&pane),
+        &tail,
+        &PaneWidthQuery::Failed {
+            error: "tmux_query_failed:CalledProcessError".to_string(),
+        },
+        &log,
+    )
+    .unwrap();
+    // fail-safe: no width → exact-equality only → truncated prefix refused.
+    assert!(!out.answered, "Failed pane-width must NOT enable prefix/truncation matching");
+    assert_eq!(out.reason, "workspace_dir_mismatch");
+    assert_eq!(out.action.as_deref(), Some("prompt_leader"));
+}
+
+#[test]
+fn tmux_pane_width_failure_yields_failed_not_default() {
+    // delivery.py:37-50 — any failure path returns Failed (never a guessed width).
+    let t = NoopTransport;
+    let target = Target::Pane(PaneId::new("%nonexistent"));
+    let q = tmux_pane_width(&t, &target);
+    assert!(
+        matches!(q, PaneWidthQuery::Failed { .. }),
+        "query failure must be fail-safe Failed, never a default width"
+    );
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP J — trust retry status machine: bounded attempt → exhausted terminal.
+// delivery.py:221-319 (_handle_trust_retry_needed).
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn handle_trust_retry_below_max_schedules_retry() {
+    // attempt 1 (< 4) → next_attempt 2 scheduled, status retry_scheduled,
+    // stage trust_auto_answer_dismissal_wait. NOT marked terminal-failed.
+    let ws = tmp_ws("trustretry1");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let payload = TrustRetryPayload {
+        message_id: "m1".to_string(),
+        attempt: 1,
+        max_attempts: TRUST_RETRY_MAX_ATTEMPTS,
+        first_target: PaneId::new("%7"),
+    };
+    let out = handle_trust_retry_needed(&store, &payload, &log).unwrap();
+    assert_eq!(out.status, DeliveryStatus::RetryScheduled);
+    assert_eq!(out.stage, Some(DeliveryStage::TrustAutoAnswerDismissalWait));
+    assert!(!out.ok);
+}
+
+#[test]
+fn handle_trust_retry_at_max_is_exhausted_terminal() {
+    // attempt == 4 (== MAX) → next_attempt 5 > MAX → terminal exhausted, marks
+    // message failed, emits trust_auto_answer_exhausted. delivery.py:246-266.
+    let ws = tmp_ws("trustretry4");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let payload = TrustRetryPayload {
+        message_id: "m1".to_string(),
+        attempt: TRUST_RETRY_MAX_ATTEMPTS,
+        max_attempts: TRUST_RETRY_MAX_ATTEMPTS,
+        first_target: PaneId::new("%7"),
+    };
+    let out = handle_trust_retry_needed(&store, &payload, &log).unwrap();
+    // delivery.py:257-259 — terminal exhausted: ok False, status the dedicated
+    // trust_auto_answer_exhausted (a bounded-loop termination guarantee, NOT a
+    // refusal reason — `reason` stays None at the typed boundary).
+    assert_eq!(out.status, DeliveryStatus::TrustAutoAnswerExhausted);
+    assert!(!out.ok);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP K — send_message target resolution / fallback chain (send.py:36-372).
+// RED via unimplemented!(); golden status/reason encoded in assertions.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn send_message_target_not_in_team_is_refused() {
+    // send.py:259-261 — non-leader, non-team target → refused/target_not_in_team.
+    let ws = tmp_ws("sendrefuse");
+    let opts = SendOptions::default();
+    let out = send_message(
+        &ws,
+        &MessageTarget::Single("ghost".to_string()),
+        "hi",
+        &opts,
+    )
+    .unwrap();
+    assert_eq!(out.status, DeliveryStatus::Refused);
+    assert_eq!(out.reason, Some(DeliveryRefusal::TargetNotInTeam));
+}
+
+#[test]
+fn send_message_broadcast_empty_team_skips_no_recipients() {
+    // send.py:391-393 — "*" with no team recipients →
+    //   {"ok": False, "status": "failed", "reason": "no_team_recipients", "to": "*"}.
+    // Post-#230 N31/N32 funnel implementation (cr-approved): broadcast now expands the
+    // recipient set via `broadcast_recipients(state, sender, team)` and routes each
+    // recipient through the SAME primitives as a single send (leader → primitive,
+    // peer → send_message). The assertions stay the same: with no agents seeded and
+    // sender="leader" (default opts.sender), `broadcast_recipients` returns an empty
+    // list — outcome is Failed/no-recipients with channel="*". The "*" channel label
+    // is preserved through the new `fanout_send(..., channel_label="*")` parameter so
+    // legacy consumers can still tell broadcast (`*`) apart from explicit fanout list.
+    let ws = tmp_ws("sendbcast");
+    let opts = SendOptions::default();
+    let out = send_message(&ws, &MessageTarget::Broadcast, "hi", &opts).unwrap();
+    assert!(!out.ok);
+    assert_eq!(out.status, DeliveryStatus::Failed);
+    assert_eq!(
+        out.reason, None,
+        "no_team_recipients is a failed-status terminal, not a typed refusal reason"
+    );
+    assert_eq!(
+        out.channel.as_deref(),
+        Some("*"),
+        "broadcast outcome must carry the '*' channel (send.py to='*'); fanout_send(channel_label=\"*\") preserves this"
+    );
+}
+
+#[test]
+fn send_message_fanout_empty_recipients_fails() {
+    // send.py:456-457 — fanout with no usable recipients → ok False,
+    // no_fanout_recipients. (Dedup-then-deliver happy path needs team fixtures.)
+    let ws = tmp_ws("sendfanout");
+    let opts = SendOptions::default();
+    let out = send_message(&ws, &MessageTarget::Fanout(vec![]), "hi", &opts).unwrap();
+    assert!(!out.ok);
+    assert_eq!(out.status, DeliveryStatus::Failed);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP L — apply_worker_sender_bypass: owner-gate first-door bypass event.
+// owner_bypass.py:9-26.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn worker_sender_bypass_false_for_leader_sender() {
+    // owner_bypass.py — leader sender never bypasses (worker_sender_bypasses=None).
+    let ws = tmp_ws("bypassleader");
+    let log = EventLog::new(&ws);
+    let state = json(serde_json::json!({"agents": {"w1": {}}}));
+    let bypassed = apply_worker_sender_bypass(
+        &state,
+        Some("leader"),
+        &MessageTarget::Single("w1".to_string()),
+        None,
+        &log,
+    )
+    .unwrap();
+    assert!(!bypassed);
+}
+
+#[test]
+#[serial_test::serial(env)]
+fn worker_sender_bypass_true_for_known_worker_sender() {
+    // owner_bypass.py:18-26 — worker in agents bypasses, writes
+    // send.bypassed_owner_gate_worker_sender.
+    // Isolate from ambient TEAM_AGENT_ID: the env identity gate only activates when
+    // TEAM_AGENT_ID is SET (and != sender → deny, see p2_owner_bypass_denies_*). Unset
+    // here so the agents-membership bypass is tested deterministically regardless of the
+    // process env (workers run with TEAM_AGENT_ID set; the leader does not).
+    let _g = ENV_LOCK_MSG.lock().unwrap_or_else(|p| p.into_inner());
+    let _e = EnvGuardMsg::set("TEAM_AGENT_ID", None);
+    let ws = tmp_ws("bypassworker");
+    let log = EventLog::new(&ws);
+    let state = json(serde_json::json!({"agents": {"w1": {}}}));
+    let bypassed = apply_worker_sender_bypass(
+        &state,
+        Some("w1"),
+        &MessageTarget::Single("w2".to_string()),
+        None,
+        &log,
+    )
+    .unwrap();
+    assert!(bypassed);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP M — report_result intake (results.py:191-227): validate envelope,
+// queue leader notify (channel coordinator), return ok shape.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn report_result_valid_envelope_returns_ok_with_result_id() {
+    let ws = tmp_ws("report");
+    let envelope = json(serde_json::json!({
+        "schema_version": "result_envelope_v1",
+        "task_id": "t1", "agent_id": "alice", "status": "success",
+        "summary": "done", "changes": [], "tests": [], "risks": [],
+        "artifacts": [], "next_actions": []
+    }));
+    let out = report_result(&ws, &envelope).unwrap();
+    // results.py:216-227 — ok True with result_id/task_id/agent_id echoed.
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(true));
+    assert_eq!(
+        out.get("task_id").and_then(|v| v.as_str()),
+        Some("t1")
+    );
+    assert_eq!(
+        out.get("agent_id").and_then(|v| v.as_str()),
+        Some("alice")
+    );
+    assert!(out.get("result_id").and_then(|v| v.as_str()).is_some());
+}
+
+#[test]
+fn report_result_funnels_into_leader_delivery_primitive_not_queued_scheduled_event() {
+    // #230 N31/N32 funnel (cr verdict §3 I-3 + MUST-8):
+    //
+    // [OLD assertion] report_result inserted a parallel `scheduled_events(kind='send',
+    // target='leader', status='pending')` row + returned `notification_status="queued"`,
+    // and the worker-facing tool body claimed success while leader had not yet seen the
+    // result. That queued-only path was MUST-8 / I-3 violating — `notification_status=
+    // queued` was returned as success but the leader pane never actually got the text.
+    //
+    // [NEW assertion] report_result now synchronously funnels through the shared leader-
+    // delivery primitive (`send_to_leader_receiver`), creating a `messages` row that
+    // `deliver_pending_messages` picks up on the same tick (NO `scheduled_events` row).
+    // Without a bound leader pane (this fixture has no `leader_receiver.pane_id`), the
+    // primitive returns I-4 `rebind_required` (Blocked / ok=false) — the row is persisted
+    // as `failed` for audit and the tool body's `notification_status` is `rebind_required`,
+    // NOT a misleading `queued` success. The contract grep that bans `queue_report_result_
+    // notification` / `notification_status="queued[_only]"` literals in `results.rs` is the
+    // direct mechanical counterpart of this assertion.
+    let ws = tmp_ws("reportnotify");
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({
+            "session_name": null,
+            "leader": {"id": "leader"},
+            "agents": {"worker": {"status": "running"}},
+            "tasks": [{"id": "task_1", "status": "running", "assignee": "worker"}]
+        }),
+    )
+    .unwrap();
+    let store = store_for(&ws);
+    let envelope = json(serde_json::json!({
+        "schema_version": "result_envelope_v1",
+        "task_id": "task_1",
+        "agent_id": "worker",
+        "status": "success",
+        "summary": "done",
+        "changes": [],
+        "tests": [{"command": "cargo test", "status": "passed"}],
+        "risks": [],
+        "artifacts": [],
+        "next_actions": []
+    }));
+
+    let out = report_result(&ws, &envelope).unwrap();
+    let result_id = out
+        .get("result_id")
+        .and_then(|v| v.as_str())
+        .expect("report_result returns generated result_id");
+    assert!(
+        result_id.starts_with("res_"),
+        "MessageStore.add_result generates res_* ids; got {result_id}"
+    );
+
+    // No `scheduled_events` rows: the queued parallel path is gone.
+    let conn = seed_conn(&store);
+    let scheduled_count: i64 = conn
+        .query_row("select count(*) from scheduled_events", [], |row| row.get(0))
+        .unwrap();
+    assert_eq!(
+        scheduled_count, 0,
+        "N31/N32 funnel: report_result must NOT insert a parallel scheduled_events 'send' row; the leader-delivery primitive is the single funnel"
+    );
+
+    // Tool body: no `queued`/`queued_only` notification_status. Without a bound leader
+    // pane this fixture surfaces I-4 `rebind_required` (ok=false on the leader delivery,
+    // but the result row + audit trail are durable for rebind replay).
+    assert_eq!(
+        out.get("notification_status").and_then(|v| v.as_str()),
+        Some("rebind_required"),
+        "I-4: unbound leader pane → rebind_required, never queued/queued_only success"
+    );
+    assert_eq!(
+        out.get("leader_notified").and_then(|v| v.as_bool()),
+        Some(false),
+        "I-4: leader_notified=false when no leader pane is bound"
+    );
+    assert!(
+        out.get("notification_event_id").is_some_and(|v| v.is_null()),
+        "no scheduled_events row → notification_event_id is null"
+    );
+
+    // Audit events: the funnel emits leader_receiver.delivery_blocked (I-4 rebind),
+    // and the legacy mcp.report_result_notify_queued audit is gone.
+    let events_path = ws.join(".team").join("logs").join("events.jsonl");
+    let event_lines = std::fs::read_to_string(events_path)
+        .expect("report_result writes events.jsonl");
+    assert!(
+        event_lines.contains("\"leader_receiver.delivery_blocked\""),
+        "I-4 rebind path must emit leader_receiver.delivery_blocked audit; got {event_lines}",
+    );
+    assert!(
+        !event_lines.contains("mcp.report_result_notify_queued"),
+        "legacy queued-notification audit must be gone; got {event_lines}",
+    );
+    assert!(
+        event_lines.contains("\"mcp.report_result\""),
+        "report_result still emits its own audit event; got {event_lines}",
+    );
+}
+
+#[test]
+fn report_result_invalid_envelope_errors_validation() {
+    // validate_result_envelope raises ValidationError → MessagingError::Validation.
+    let ws = tmp_ws("reportbad");
+    let envelope = json(serde_json::json!({"schema_version": "result_envelope_v1"}));
+    let err = report_result(&ws, &envelope).unwrap_err();
+    assert!(
+        matches!(err, MessagingError::Validation(_)),
+        "missing required fields must surface as Validation, got {err:?}"
+    );
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP N — notify_result_watchers dedupe (exactly-once, Gap 32/38).
+// result_delivery.py:38-132. superseded for duplicate watchers same result.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn notify_result_watchers_no_match_returns_empty() {
+    // result_delivery.py:51-52 — no candidate watcher matches → empty list.
+    let ws = tmp_ws("notifyempty");
+    let log = EventLog::new(&ws);
+    let result = json(serde_json::json!({
+        "result_id": "r1", "task_id": "t1", "agent_id": "alice"
+    }));
+    let watchers = vec![json(serde_json::json!({
+        "watcher_id": "w-x", "task_id": "OTHER", "agent_id": "alice"
+    }))];
+    let notices = notify_result_watchers(
+        &ws,
+        &result,
+        &log,
+        Some(&watchers),
+        None,
+    )
+    .unwrap();
+    assert!(notices.is_empty());
+}
+
+#[test]
+fn notify_result_watchers_supersedes_duplicate_watchers() {
+    // result_delivery.py:53-78 — two watchers same (task,agent,result): earliest is
+    // primary, the other gets superseded (ok False, notice records superseded).
+    let ws = tmp_ws("notifysup");
+    let log = EventLog::new(&ws);
+    let result = json(serde_json::json!({
+        "result_id": "r1", "task_id": "t1", "agent_id": "alice"
+    }));
+    let watchers = vec![
+        json(serde_json::json!({
+            "watcher_id": "w-early", "task_id": "t1", "agent_id": "alice",
+            "created_at": "2026-06-02T10:00:00+00:00"
+        })),
+        json(serde_json::json!({
+            "watcher_id": "w-late", "task_id": "t1", "agent_id": "alice",
+            "created_at": "2026-06-02T11:00:00+00:00"
+        })),
+    ];
+    let notices =
+        notify_result_watchers(&ws, &result, &log, Some(&watchers), None).unwrap();
+    // The late watcher must appear as a superseded (not-ok) notice — exactly-once.
+    let superseded = notices
+        .iter()
+        .find(|n| n.watcher_id == "w-late")
+        .expect("late watcher must be reported");
+    assert!(!superseded.ok, "duplicate watcher must be superseded, not re-delivered");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP O — requeue_after_claim_leader: notified_message_id must SURVIVE (Gap
+// 32) — already-notified watchers are NOT requeued. result_delivery.py:428-506.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn requeue_after_claim_leader_skips_already_notified() {
+    // Gap 32 (result_delivery.py:467-471) — SEEDED dedupe gate: two same-team
+    // watchers, one already notified (notified_message_id set), one un-notified.
+    // requeue must return ONLY the un-notified watcher; the notified one is NOT
+    // requeued and its notified_message_id SURVIVES (clearing it would cause a
+    // second injection). Probed golden: requeued == [w_un] (result_id null,
+    // prior_state "pending"); notified watcher keeps notified_message_id.
+    let ws = tmp_ws("requeue");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let team = TeamKey::new("team-a");
+    let pane = PaneId::new("%new-leader");
+
+    let w_un = seed_watcher(&store, "w-unnotified", "team-a", "t1", "alice", "pending", None, None);
+    let w_notified = seed_watcher(
+        &store, "w-notified", "team-a", "t2", "bob", "pending", None, Some("msg_already"),
+    );
+
+    let requeued =
+        requeue_after_claim_leader(&ws, &store, &log, &team, &pane, None).unwrap();
+
+    // ONLY the un-notified watcher requeues (the notified one is the dedupe gate).
+    let ids: Vec<&str> = requeued.iter().map(|n| n.watcher_id.as_str()).collect();
+    assert_eq!(ids, vec![w_un.as_str()], "exactly the un-notified watcher requeues");
+    assert!(
+        !requeued.iter().any(|n| n.watcher_id == w_notified),
+        "already-notified watcher must NOT be requeued (Gap 32)"
+    );
+
+    // Gap 32 survival: notified_message_id is preserved on the skipped watcher.
+    let (_status, notified) = watcher_state(&store, &w_notified);
+    assert_eq!(
+        notified.as_deref(),
+        Some("msg_already"),
+        "notified_message_id MUST survive requeue — clearing it re-injects (Gap 32)"
+    );
+}
+
+#[test]
+fn requeue_delivery_exhausted_watchers_reopens_only_exhausted() {
+    let ws = tmp_ws("requeueexhausted");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let team = TeamKey::new("team-a");
+    let pane = PaneId::new("%leader");
+
+    let rid = seed_result(&store, "res_exhausted", "t1", "alice", "success");
+    let exhausted = seed_watcher(
+        &store,
+        "w-exhausted",
+        "team-a",
+        "t1",
+        "alice",
+        "delivery_exhausted",
+        Some(&rid),
+        None,
+    );
+    let notified = seed_watcher(
+        &store,
+        "w-exhausted-notified",
+        "team-a",
+        "t2",
+        "bob",
+        "delivery_exhausted",
+        Some("res_skip"),
+        Some("msg_done"),
+    );
+    let failed = seed_watcher(
+        &store,
+        "w-failed",
+        "team-a",
+        "t3",
+        "carol",
+        "notify_failed",
+        Some("res_failed"),
+        None,
+    );
+
+    let requeued =
+        requeue_delivery_exhausted_watchers(&ws, &store, &log, &team, &pane).unwrap();
+
+    assert_eq!(requeued.len(), 1, "only delivery_exhausted unnotified watchers requeue");
+    let notice = &requeued[0];
+    assert_eq!(notice.watcher_id, exhausted);
+    assert_eq!(notice.result_id.as_deref(), Some(rid.as_str()));
+    assert_eq!(notice.prior_state.as_deref(), Some("delivery_exhausted"));
+    // R8 (golden result_watchers.py:95): attach requeue flips delivery_exhausted -> notify_failed (NOT pending).
+    assert_eq!(notice.status.as_deref(), Some("notify_failed"));
+
+    let (status, _notified_id) = watcher_state(&store, &exhausted);
+    // R8 (golden): attach requeue leaves the watcher at notify_failed and DEFERS retry to the coordinator
+    // tick — it does NOT immediately re-deliver (only the claim path retries). So the persisted status is
+    // notify_failed, not 'notified'.
+    assert_eq!(status, "notify_failed", "attach requeue flips to notify_failed and defers retry (golden)");
+    let (status, notified_id) = watcher_state(&store, &notified);
+    assert_eq!(status, "delivery_exhausted");
+    assert_eq!(notified_id.as_deref(), Some("msg_done"));
+    let (status, _notified_id) = watcher_state(&store, &failed);
+    assert_eq!(status, "notify_failed", "non-exhausted watcher is not selected");
+
+    // R8 (golden leader/__init__.py:46-50): result_watcher.requeued is the ATTACH form
+    // {watcher_id, trigger:"attach_leader", new_pane_id} — NOT the claim-style {prior_state,claimed_pane_id,team_id}.
+    let events = log.tail(0).unwrap();
+    let ev = events.iter().rev()
+        .find(|event| event.get("event").and_then(|v| v.as_str()) == Some("result_watcher.requeued"))
+        .expect("result_watcher.requeued event");
+    let keys: std::collections::BTreeSet<&str> = ev.as_object().unwrap().keys()
+        .map(String::as_str).filter(|k| *k != "ts" && *k != "event").collect();
+    let expected: std::collections::BTreeSet<&str> = ["watcher_id", "trigger", "new_pane_id"].into_iter().collect();
+    assert_eq!(keys, expected, "result_watcher.requeued must be golden attach form {{watcher_id, trigger, new_pane_id}}");
+    assert_eq!(ev.get("watcher_id").and_then(|v| v.as_str()), Some("w-exhausted"));
+    assert_eq!(ev.get("trigger").and_then(|v| v.as_str()), Some("attach_leader"));
+    assert_eq!(ev.get("new_pane_id").and_then(|v| v.as_str()), Some("%leader"));
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP P — stuck_cancel owner-gate + invalid alert type refusal.
+// scheduler.py:262-294.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn stuck_cancel_none_alert_type_expands_to_all() {
+    // alert_type None == Python "all" → sorted(_ALERT_TYPES) expansion.
+    let ws = tmp_ws("stuckcancel");
+    let out = stuck_cancel(&ws, "w1", None, "leader").unwrap();
+    // The suppression result must enumerate all three alert types.
+    let types = out
+        .get("alert_types")
+        .and_then(|v| v.as_array())
+        .map(|a| a.iter().filter_map(|x| x.as_str().map(str::to_string)).collect::<Vec<_>>());
+    assert_eq!(
+        types,
+        Some(vec![
+            "cross_worker_deadlock".to_string(),
+            "idle_fallback".to_string(),
+            "stuck".to_string()
+        ])
+    );
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP Q — collect intake (results.py:45-167): valid result advances task,
+// returns collected_results + delivered_messages + results counts shape.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn collect_without_spec_surfaces_validation_error() {
+    // results.py:46-48 — collect() reads spec_path then load_spec(); against a
+    // workspace with NO team.spec.yaml, load_spec RAISES ValidationError
+    // ("Cannot read <path>: ...", spec.py:18-20) BEFORE any collection. The
+    // previous `.unwrap()` (expecting an Ok dict with present-only keys) was wrong:
+    // the real Python collect on a bare workspace does not return a dict, it raises.
+    // At the typed boundary that surfaces as MessagingError::Validation.
+    //
+    // The full collected_results-count golden (seed an uncollected result, assert it
+    // collects and the task advances) is DEFERRED: it requires a valid on-disk
+    // team.spec.yaml + runtime state, whose formats are owned by the spec/state lanes
+    // (this file may not edit them). seed_result() exists for the retry path that
+    // does NOT need a spec; the collect happy-path needs an integration fixture.
+    let ws = tmp_ws("collect");
+    let err = collect(&ws, None, false).unwrap_err();
+    assert!(
+        matches!(err, MessagingError::Validation(_)),
+        "collect without a team spec must surface Validation, got {err:?}"
+    );
+}
+
+#[test]
+fn collect_accepts_message_scoped_result_for_matching_recipient() {
+    let ws = tmp_ws("collectmsgok");
+    std::fs::write(ws.join("team.spec.yaml"), "version: 1\n").unwrap();
+    let store = store_for(&ws);
+    let message_id = store
+        .create_message(None, "leader", "w1", "please reply", None, false, None)
+        .unwrap();
+    seed_result(&store, "res_msg_ok", &message_id, "w1", "success");
+
+    let out = collect(&ws, None, false).unwrap();
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(true));
+    let collected = out
+        .get("collected_results")
+        .and_then(|v| v.as_array())
+        .expect("collected_results");
+    assert_eq!(collected.len(), 1);
+    assert_eq!(collected[0].get("task_id").and_then(|v| v.as_str()), Some(message_id.as_str()));
+    assert_eq!(collected[0].get("agent_id").and_then(|v| v.as_str()), Some("w1"));
+    assert_eq!(collected[0].get("scope").and_then(|v| v.as_str()), Some("message"));
+    // D3 (leader-adjudicated): golden collected_results entry is EXACTLY the 8-key summary for BOTH
+    // scopes; golden's task_status feeds only the `collect.result` EVENT, never the entry. So a
+    // message-scope entry carries NO task_status key (the prior `Some("message_scoped")` lock encoded a
+    // port divergence — dropped per ruling).
+    assert!(
+        collected[0].get("task_status").is_none(),
+        "collected_results entry must NOT carry task_status (golden 8-key summary; event-only); got {:?}",
+        collected[0]
+    );
+    let keys: Vec<&str> = collected[0].as_object().expect("entry is an object").keys().map(String::as_str).collect();
+    assert_eq!(
+        keys,
+        vec!["result_id", "task_id", "agent_id", "status", "summary", "tests", "created_at", "scope"],
+        "message-scope collected_results entry must be EXACTLY the golden 8 keys in order; got {keys:?}"
+    );
+}
+
+#[test]
+fn collect_rejects_message_scoped_result_without_matching_recipient() {
+    let ws = tmp_ws("collectmsgbad");
+    std::fs::write(ws.join("team.spec.yaml"), "version: 1\n").unwrap();
+    let store = store_for(&ws);
+    let message_id = store
+        .create_message(None, "leader", "w1", "please reply", None, false, None)
+        .unwrap();
+    seed_result(&store, "res_msg_bad", &message_id, "w2", "success");
+
+    let out = collect(&ws, None, false).unwrap();
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(false));
+    assert!(
+        out.get("collected_results")
+            .and_then(|v| v.as_array())
+            .is_some_and(Vec::is_empty),
+        "recipient mismatch must not collect as message-scoped"
+    );
+    let invalid = out
+        .get("invalid_results")
+        .and_then(|v| v.as_array())
+        .expect("invalid_results");
+    assert_eq!(invalid.len(), 1);
+    assert_eq!(invalid[0].get("task_id").and_then(|v| v.as_str()), Some(message_id.as_str()));
+    assert_eq!(
+        invalid[0].get("error").and_then(|v| v.as_str()),
+        Some(format!("unknown task id: {message_id}").as_str())
+    );
+}
+
+#[test]
+fn allow_peer_talk_records_bidirectional_allowlist_and_event() {
+    let ws = tmp_ws("allowpeer");
+    let out = allow_peer_talk(&ws, "alice", "bob").unwrap();
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(true));
+    assert_eq!(out.get("a").and_then(|v| v.as_str()), Some("alice"));
+    assert_eq!(out.get("b").and_then(|v| v.as_str()), Some("bob"));
+    assert_eq!(out.get("status").and_then(|v| v.as_str()), Some("compat_noop"));
+    assert_eq!(
+        out.get("reason").and_then(|v| v.as_str()),
+        Some("team_scoped_peer_messages_enabled")
+    );
+
+    let store = store_for(&ws);
+    let conn = seed_conn(&store);
+    let rows = conn
+        .prepare("select a, b from peer_allowlist order by a, b")
+        .unwrap()
+        .query_map([], |row| Ok((row.get::<_, String>(0)?, row.get::<_, String>(1)?)))
+        .unwrap()
+        .collect::<Result<Vec<_>, _>>()
+        .unwrap();
+    assert_eq!(
+        rows,
+        vec![
+            ("alice".to_string(), "bob".to_string()),
+            ("bob".to_string(), "alice".to_string()),
+        ]
+    );
+
+    let events = EventLog::new(&ws).tail(10).unwrap();
+    let event = events
+        .iter()
+        .find(|event| event.get("event").and_then(|v| v.as_str()) == Some("communication.peer_allowed"))
+        .expect("communication.peer_allowed event");
+    assert_eq!(event.get("a").and_then(|v| v.as_str()), Some("alice"));
+    assert_eq!(event.get("b").and_then(|v| v.as_str()), Some("bob"));
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP R — run_comms_selftest: §84 / MUST-NOT-13 zero-provider-SDK gate.
+// diagnose/comms.py:21-47. The whole point: assert zero provider client calls.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn run_comms_selftest_zero_provider_sdk_passes_and_locks_scope() {
+    let ws = tmp_ws("selftestok");
+    let driver = ZeroSdkDriver {
+        run_id: Some("fixedrunid01".to_string()),
+        calls: ProviderSdkCalls::default(),
+    };
+    let report = run_comms_selftest(&ws, None, &driver).unwrap();
+    // diagnose/comms.py:44 scope == "binding_consistency".
+    assert_eq!(report.scope, "binding_consistency");
+    assert_eq!(report.run_id, "fixedrunid01");
+    // The mechanical gate: provider_sdk_calls check is a Pass with all-zero evidence.
+    assert_eq!(report.provider_sdk_calls.status, CheckStatus::Pass);
+    match &report.provider_sdk_calls.evidence {
+        CheckEvidence::ProviderSdkCalls(calls) => assert!(calls.is_zero()),
+        other => panic!("expected ProviderSdkCalls evidence, got {other:?}"),
+    }
+    assert!(report.ok);
+}
+
+#[test]
+fn run_comms_selftest_nonzero_provider_sdk_fails_gate() {
+    // Any non-zero SDK call count → provider_sdk_calls check FAILS, report not ok.
+    let ws = tmp_ws("selftestbad");
+    let driver = ZeroSdkDriver {
+        run_id: Some("r2".to_string()),
+        calls: ProviderSdkCalls {
+            anthropic: 1,
+            openai: 0,
+            httpx: 0,
+        },
+    };
+    let report = run_comms_selftest(&ws, None, &driver).unwrap();
+    assert_eq!(report.provider_sdk_calls.status, CheckStatus::Fail);
+    assert!(!report.ok);
+}
+
+#[test]
+fn run_comms_selftest_contract_suite_is_deferred() {
+    // diagnose/comms.py:132-139 — contract_suite is always deferred (test files
+    // not shipped) and counts as a pass for the overall gate.
+    let ws = tmp_ws("selftestdefer");
+    let driver = ZeroSdkDriver {
+        run_id: Some("r3".to_string()),
+        calls: ProviderSdkCalls::default(),
+    };
+    let report = run_comms_selftest(&ws, None, &driver).unwrap();
+    assert_eq!(report.contract_suite.status, CheckStatus::Deferred);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP S — evaluate_idle_behavior: claimed_status normalization
+// (IDLE/WORKING/RUNNING → not_challenged). diagnose/comms.py:50-94.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn evaluate_idle_behavior_recognized_status_is_not_challenged() {
+    // diagnose/comms.py:86-94 — claimed_status in {IDLE,WORKING,RUNNING} (case-
+    // insensitive) and no driver result → status not_challenged, ok True.
+    let ws = tmp_ws("idleeval");
+    let driver = ZeroSdkDriver {
+        run_id: None,
+        calls: ProviderSdkCalls::default(),
+    };
+    let out = evaluate_idle_behavior(&ws, "w1", "IDLE", None, &driver).unwrap();
+    assert_eq!(out.status, CheckStatus::NotChallenged);
+    assert!(out.ok);
+    assert_eq!(out.agent_id, "w1");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP T — deliver_pending_message claim atomicity + status machine.
+// delivery.py:63-218. missing message / unknown recipient / already-claimed.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn deliver_pending_message_missing_message_fails() {
+    // delivery.py:73-75 — no such message row → ok False, status failed,
+    // reason message_missing.
+    let ws = tmp_ws("delivermissing");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+    let out = deliver_pending_message(&ws, &store, &t, "nope", &log, &serde_json::json!({})).unwrap();
+    assert!(!out.ok);
+    assert_eq!(out.status, DeliveryStatus::Failed);
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP U — fire_due_scheduled_events: exhaustive ScheduledKind dispatch +
+// send dedupe. scheduler.py:41-121. Returns fired event-id list.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn fire_due_scheduled_events_fires_each_scheduled_kind() {
+    // SEEDED exhaustive-dispatch contract (scheduler.py:41-121): seed one due row
+    // of EACH ScheduledKind (send / health_ping / trust_retry). The dispatch loop
+    // must fire all three (one match arm per kind, no runtime fallback) and return
+    // each fired event id. Probed golden: a due health_ping fires → marked 'done'
+    // with {"ok":true,"status":"logged"} and its id appears in the fired list;
+    // every due row's id is appended regardless of kind (scheduler.py:118).
+    let ws = tmp_ws("scheduler");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let t = NoopTransport;
+
+    let send_id = seed_scheduled_event(
+        &store,
+        ScheduledKind::Send,
+        "%w1",
+        &serde_json::json!({"content": "ping", "attempt": 1, "max_attempts": 1}),
+    );
+    let ping_id =
+        seed_scheduled_event(&store, ScheduledKind::HealthPing, "%w1", &serde_json::json!({}));
+    let trust_id = seed_scheduled_event(
+        &store,
+        ScheduledKind::TrustRetry,
+        "%w1",
+        &serde_json::json!({"message_id": "m1", "attempt": 1, "max_attempts": 4}),
+    );
+
+    let fired = fire_due_scheduled_events(&ws, &store, &t, &log).unwrap();
+
+    // Every seeded due kind must be dispatched and its id returned (exhaustive,
+    // no kind silently dropped via a fallthrough).
+    for id in [send_id, ping_id, trust_id] {
+        assert!(
+            fired.contains(&id),
+            "scheduled event id {id} (each ScheduledKind) must fire; got {fired:?}"
+        );
+    }
+    assert_eq!(fired.len(), 3, "exactly the three seeded due events fire, no extras");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP V — retry_result_deliveries: re-route notify_failed watchers with
+// dedupe_reason rebind_retry. result_delivery.py:19-35.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn retry_result_deliveries_retries_notify_failed_watcher() {
+    // SEEDED contract (result_delivery.py:18-34): retry_result_deliveries scans
+    // retryable_result_watchers (status in pending/notify_failed), resolves each
+    // watcher's result via result_by_id, and re-routes through notify_result_watchers
+    // with dedupe_reason="rebind_retry". Seed a notify_failed watcher + its matching
+    // result row → the watcher IS retried and a WatcherNotice for it is returned.
+    // Probed golden: notices == [{watcher_id, result_id, ok, ...}] for the seeded
+    // watcher (delivery ok depends on full team state; the retry-was-attempted
+    // contract is the invariant — an empty store would NOT exercise it).
+    let ws = tmp_ws("retrydeliv");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+
+    let rid = seed_result(&store, "res_r1", "t1", "alice", "success");
+    let w = seed_watcher(
+        &store, "w-failed", "team-a", "t1", "alice", "notify_failed", Some(&rid), None,
+    );
+
+    let notices = retry_result_deliveries(&ws, &log).unwrap();
+
+    assert_eq!(notices.len(), 1, "the single notify_failed watcher must be retried");
+    let notice = &notices[0];
+    assert_eq!(notice.watcher_id, w, "the retried notice names the seeded watcher");
+    assert_eq!(
+        notice.result_id.as_deref(),
+        Some(rid.as_str()),
+        "retry resolves and carries the watcher's result_id (rebind_retry path)"
+    );
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP W — collect_results_and_notify_watchers orchestration shape.
+// results.py:430-447.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn collect_results_and_notify_watchers_returns_concrete_ok_shape() {
+    // SEEDED contract (results.py:430-447): with NO uncollected results, collect() is
+    // skipped (the `if store.results(uncollected_only=True)` guard is false), so the
+    // result stays {ok:true, collected_results:[]}; a seeded notify_failed watcher
+    // whose result_id has no matching results row is resolved to None by
+    // retry_result_deliveries → skipped → notified stays []. Probed golden (against
+    // exactly this fixture): {"ok": true, "collected": 0, "notified": []}.
+    // (The previous test asserted only out["ok"].is_some(), trivially passed by
+    // {"ok": false}.)
+    let ws = tmp_ws("collectnotify");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+
+    seed_watcher(
+        &store, "w-orphan", "team-a", "t1", "alice", "notify_failed", Some("res_missing"), None,
+    );
+
+    let out = collect_results_and_notify_watchers(&ws, &log).unwrap();
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(true), "ok==true");
+    assert_eq!(
+        out.get("collected").and_then(|v| v.as_i64()),
+        Some(0),
+        "no uncollected results → collected==0"
+    );
+    assert_eq!(
+        out.get("notified").and_then(|v| v.as_array()).map(|a| a.len()),
+        Some(0),
+        "orphan watcher (missing result) is skipped → notified empty"
+    );
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// GROUP X — delivered_result_message content-level dedupe lookup +
+// result_id_from_text dual (scheduler send dedupe path). result_delivery.py:394.
+// ════════════════════════════════════════════════════════════════════════
+
+#[test]
+fn delivered_result_message_none_in_fresh_store() {
+    let ws = tmp_ws("delivdedupe");
+    let store = store_for(&ws);
+    let found = delivered_result_message(&store, "r1", None, None).unwrap();
+    assert!(found.is_none());
+}
+
+#[test]
+fn delivered_result_message_empty_result_id_is_none() {
+    // result_delivery.py:401-402 — empty result_id short-circuits to None.
+    let ws = tmp_ws("delivdedupe2");
+    let store = store_for(&ws);
+    let found = delivered_result_message(&store, "", None, None).unwrap();
+    assert!(found.is_none());
+}
+
+// ═══════════════════════════════════════════════════════════════════════════
+// collect #223 — task-scoped collect + send --task validation (RED).
+// ═══════════════════════════════════════════════════════════════════════════
+
+// (c) a result whose task_id ∈ state.tasks collects as scope:"task"; the task row advances to
+// "done" (success → done, runtime.py:1066); results.collected ≥ 1. Proves the collect-READ works
+// once state.tasks is seeded — so the #223 fix target is the upstream seeding, not collect.
+#[test]
+fn collect_task_scoped_result_collects_and_marks_task_done() {
+    let ws = tmp_ws("collecttask223");
+    std::fs::write(ws.join("team.spec.yaml"), "version: 1\n").unwrap();
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({
+            "session_name": "team-x",
+            "agents": { "w1": { "provider": "codex" } },
+            "tasks": [ { "id": "t2", "assignee": "w1", "title": "t2", "status": "pending" } ]
+        }),
+    ).unwrap();
+    let store = store_for(&ws);
+    seed_result(&store, "res_t2", "t2", "w1", "success");
+
+    let out = collect(&ws, None, false).unwrap();
+    assert_eq!(out.get("ok").and_then(|v| v.as_bool()), Some(true), "no invalid → ok:true");
+    let cr = out.get("collected_results").and_then(|v| v.as_array()).expect("collected_results");
+    assert_eq!(cr.len(), 1, "the seeded t2 result must collect");
+    assert_eq!(cr[0].get("scope").and_then(|v| v.as_str()), Some("task"), "t2 ∈ state.tasks → scope:task");
+    assert_eq!(cr[0].get("task_id").and_then(|v| v.as_str()), Some("t2"));
+    assert_eq!(cr[0].get("agent_id").and_then(|v| v.as_str()), Some("w1"));
+    assert!(
+        out.get("results").and_then(|r| r.get("collected")).and_then(|v| v.as_i64()).unwrap_or(0) >= 1,
+        "results.collected must be ≥ 1"
+    );
+    let st = crate::state::persist::load_runtime_state(&ws).unwrap();
+    let t2_status = st.get("tasks").and_then(|v| v.as_array())
+        .and_then(|ts| ts.iter().find(|t| t.get("id").and_then(|v| v.as_str()) == Some("t2")))
+        .and_then(|t| t.get("status")).and_then(|v| v.as_str());
+    assert_eq!(t2_status, Some("done"), "success result → task row status 'done' (runtime.py:1066)");
+}
+
+// (c-C1) collect OUTPUT shape: collected_results entries are the 8-KEY SUMMARY (NO inlined
+// envelope; carry summary+tests) and the full envelopes live in a SEPARATE top-level `collected`
+// list (golden results.py:86/131). Rust inlines `envelope`+`owner_team_id` and emits no
+// `collected` list → RED.
+#[test]
+fn collect_output_matches_golden_collected_shape() {
+    let ws = tmp_ws("collectshape223");
+    std::fs::write(ws.join("team.spec.yaml"), "version: 1\n").unwrap();
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({
+            "session_name": "team-x",
+            "agents": { "w1": { "provider": "codex" } },
+            "tasks": [ { "id": "t2", "assignee": "w1", "title": "t2", "status": "pending" } ]
+        }),
+    ).unwrap();
+    let store = store_for(&ws);
+    seed_result(&store, "res_t2s", "t2", "w1", "success");
+
+    let out = collect(&ws, None, false).unwrap();
+    let cr = out.get("collected_results").and_then(|v| v.as_array()).expect("collected_results");
+    let e = &cr[0];
+    // C1: collected_results entry is the 8-key SUMMARY — NO envelope inlined; carries summary+tests.
+    assert!(e.get("envelope").is_none(),
+        "collected_results entry must NOT inline `envelope` (golden 8-key summary); the full envelope belongs in `collected`. got {e:?}");
+    assert!(e.get("summary").is_some() && e.get("tests").is_some(),
+        "collected_results summary entry must carry `summary`+`tests` (golden results.py:131)");
+    // C1: the full envelopes live in a separate top-level `collected` list.
+    let collected = out.get("collected").and_then(|v| v.as_array())
+        .expect("golden collect returns a top-level `collected` list of full envelopes");
+    assert!(
+        collected.first().and_then(|env| env.get("schema_version")).and_then(|v| v.as_str())
+            == Some("result_envelope_v1"),
+        "collected[0] must be the full result_envelope_v1 envelope; got {collected:?}"
+    );
+
+    // ── STRENGTHENED (option-B byte-parity, leader-adjudicated 0700cff review) ──
+    // D3 — task-scope collected_results entry must be EXACTLY the golden 8 keys, in order, NO task_status.
+    let keys: Vec<&str> = e.as_object().expect("entry is an object").keys().map(String::as_str).collect();
+    assert_eq!(
+        keys,
+        vec!["result_id", "task_id", "agent_id", "status", "summary", "tests", "created_at", "scope"],
+        "collected_results entry must be EXACTLY the golden 8 keys in order (results.py:131; no task_status/envelope/owner_team_id); got {keys:?}"
+    );
+    // D1+D2 — collect RETURN top-level key order must match golden EXACTLY: delivered_messages BEFORE
+    // invalid_results, AND a `coordinator` key (mirroring golden _ensure_coordinator_after_collect).
+    let top: Vec<&str> = out.as_object().expect("collect result is an object").keys().map(String::as_str).collect();
+    assert_eq!(
+        top,
+        vec!["ok", "collected", "collected_results", "delivered_messages", "invalid_results", "results", "state_file", "coordinator"],
+        "collect return top-level key order must match golden return shape; got {top:?}"
+    );
+}
+
+// (d) send --task <unknown id> must RAISE golden "unknown task id" (runtime.py:1032 _find_task),
+// not silently create a message. Rust send_message attaches task_id without validating → Ok. RED.
+// block_until_delivered=false isolates the task-validation from any delivery side-effect.
+#[test]
+fn send_with_unknown_task_id_raises_unknown_task() {
+    let ws = tmp_ws("sendunknowntask223");
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({
+            "session_name": "team-x",
+            "agents": { "w1": { "provider": "codex" } },
+            "tasks": []
+        }),
+    ).unwrap();
+    let _ = store_for(&ws);
+    let opts = SendOptions {
+        task_id: Some(crate::model::ids::TaskId::new("t2-unknown")),
+        block_until_delivered: false,
+        ..SendOptions::default()
+    };
+    let out = send_message(&ws, &MessageTarget::Single("w1".to_string()), "go", &opts);
+    match out {
+        Err(e) => {
+            // SURFACED error = the CLI `error` field = CliError::from(MessagingError).to_string()
+            // (to_payload uses self.to_string(), types.rs:59). Must EQUAL golden's bare message —
+            // NO "validation:" variant prefix (golden runtime.py:1032 surfaces str(exc)).
+            let surfaced = crate::cli::CliError::from(e).to_string();
+            assert_eq!(
+                surfaced, "unknown task id: t2-unknown",
+                "surfaced CLI error must EQUAL golden's message with NO variant prefix; got {surfaced:?}"
+            );
+        }
+        Ok(o) => panic!(
+            "send --task <unknown id> must RAISE 'unknown task id' (golden runtime.py:1032 _find_task), \
+             not silently create a message; got Ok({o:?})"
+        ),
+    }
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// P0 REGRESSION (0700cff "send 0 bytes, nothing queued" / coordinator never delivers).
+// golden gates the unknown-task RAISE on route_task_id (send.py:204 `if task_id and route_task_id`);
+// delivery/fanout/internal sends pass route_task_id=False (internal_delivery.py:44, send.py:412/481)
+// → the task is a label, NOT validated. 0700cff's UNCONDITIONAL task_exists gate broke every
+// task-tagged delivery/internal send at CREATION time. The gate the OfflineTransport tests missed.
+// ════════════════════════════════════════════════════════════════════════
+
+// (a) [REGRESSION GATE] route_task_id=false + task_id NOT in state.tasks → send SUCCEEDS and the
+// message is QUEUED (real create path; no transport). Must NOT raise "unknown task id".
+#[test]
+fn send_route_task_id_false_skips_task_validation_and_queues() {
+    let ws = tmp_ws("sendroutefalse");
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({ "session_name": "team-x", "agents": { "w1": { "provider": "codex" } }, "tasks": [] }),
+    ).unwrap();
+    let _ = store_for(&ws);
+    let opts = SendOptions {
+        task_id: Some(crate::model::ids::TaskId::new("t-not-seeded")),
+        route_task_id: false,
+        block_until_delivered: false,
+        ..SendOptions::default()
+    };
+    let out = send_message(&ws, &MessageTarget::Single("w1".to_string()), "deliver me", &opts)
+        .expect("route_task_id=false must NOT validate the task — golden delivery/internal path queues regardless of state.tasks");
+    assert!(
+        out.message_id.is_some(),
+        "the message must be CREATED (message_id present) on the route_task_id=false path; got {out:?}"
+    );
+    // real queue verification (not an Ok shell): the message landed in w1's inbox.
+    let inbox = store_for(&ws).inbox("w1", 10, None).expect("inbox");
+    assert!(
+        !inbox.is_empty(),
+        "the task-tagged message must be QUEUED for w1 on the delivery/internal path; inbox empty (0 bytes queued = the P0)"
+    );
+}
+
+// (c) [LOCK] route_task_id=true + task_id IN state.tasks → send SUCCEEDS (routing happy-path).
+#[test]
+fn send_route_task_id_true_known_task_succeeds() {
+    let ws = tmp_ws("sendrouteknown");
+    crate::state::persist::save_runtime_state(
+        &ws,
+        &serde_json::json!({
+            "session_name": "team-x",
+            "agents": { "w1": { "provider": "codex" } },
+            "tasks": [ { "id": "t-known", "assignee": "w1", "title": "t", "status": "pending" } ]
+        }),
+    ).unwrap();
+    let _ = store_for(&ws);
+    let opts = SendOptions {
+        task_id: Some(crate::model::ids::TaskId::new("t-known")),
+        route_task_id: true,
+        block_until_delivered: false,
+        ..SendOptions::default()
+    };
+    let out = send_message(&ws, &MessageTarget::Single("w1".to_string()), "go", &opts)
+        .expect("route_task_id=true with a KNOWN task must succeed");
+    assert!(out.message_id.is_some(), "known-task routing send must create the message; got {out:?}");
+}
+
+// ════════════════════════════════════════════════════════════════════════
+// R8 byte-parity (leader attach requeue, advisor-ruled + e3eac28-reconciled):
+// drive a watcher to delivery_exhausted via notify_result_watchers (attempts>=MAX) — proving the
+// requeue input is REAL (non-空过) — then attach-requeue and assert the golden observable contract:
+//   D2 status: delivery_exhausted -> notify_failed (golden result_watchers.py:95), NOT 'pending'.
+//   D1 ✦ team-scoped + unnotified SELECTION (anti cross-team pollution / CP-1) — KEEP.
+//   D3 result_watcher.requeued payload == golden attach form {watcher_id, trigger:"attach_leader", new_pane_id}.
+// (D4 leader_receiver.requeued_exhausted_watchers + D6 string return are the attach-wrapper/CLI layer —
+//  lease.rs:140 + cli/mod.rs:1088 — flagged for the porter; D5 event-layer is internal/optional.)
+// ════════════════════════════════════════════════════════════════════════
+#[test]
+fn r8_attach_requeue_exhausted_to_notify_failed_golden_attach_event() {
+    let ws = tmp_ws("r8requeue");
+    let store = store_for(&ws);
+    let log = EventLog::new(&ws);
+    let team = TeamKey::new("team-a");
+    let pane = PaneId::new("%leader-new");
+
+    // --- Sub-A: DRIVE w-r8 (team-a) to delivery_exhausted via notify_result_watchers (attempts>=MAX) ---
+    let rid = seed_result(&store, "res_r8", "t1", "alice", "success");
+    seed_watcher(&store, "w-r8", "team-a", "t1", "alice", "pending", Some(&rid), None);
+    // attempts are EVENT-counted (result_watcher.notify_failed/retry_notified) — seed MAX prior failures.
+    for n in 0..u64::from(RESULT_DELIVERY_MAX_ATTEMPTS) {
+        log.write(
+            "result_watcher.notify_failed",
+            json(serde_json::json!({"watcher_id": "w-r8", "result_id": rid.as_str(), "status": "notify_failed", "error": "x", "n": n})),
+        ).unwrap();
+    }
+    let result_env = json(serde_json::json!({"result_id": rid.as_str(), "task_id": "t1", "agent_id": "alice"}));
+    let watcher_view = json(serde_json::json!({
+        "watcher_id": "w-r8", "task_id": "t1", "agent_id": "alice",
+        "created_at": "2026-01-01T00:00:00Z", "owner_team_id": "team-a",
+        "leader_id": "leader", "result_id": rid.as_str()
+    }));
+    notify_result_watchers(&ws, &result_env, &log, Some(&[watcher_view]), None).unwrap();
+    let (driven, _) = watcher_state(&store, "w-r8");
+    assert_eq!(driven, "delivery_exhausted",
+        "PRECONDITION: notify_result_watchers at attempts>=MAX must persist delivery_exhausted (watchers.rs:161-168) — \
+         proves the attach-requeue input is real, not 空过");
+
+    // selection-lock fixtures: cross-team exhausted + notified exhausted (Gap-32) + pending.
+    let team_b = seed_watcher(&store, "w-teamb", "team-b", "t2", "bob", "delivery_exhausted", Some("res_b"), None);
+    let notif = seed_watcher(&store, "w-notified", "team-a", "t3", "carol", "delivery_exhausted", Some("res_c"), Some("msg_done"));
+    seed_watcher(&store, "w-pending", "team-a", "t4", "dave", "pending", Some("res_d"), None);
+
+    // --- Sub-B: attach requeue (golden contract) ---
+    let requeued = requeue_delivery_exhausted_watchers(&ws, &store, &log, &team, &pane).unwrap();
+
+    // D2: team-a exhausted -> notify_failed (NOT pending).
+    let (st_a, _) = watcher_state(&store, "w-r8");
+    assert_eq!(st_a, "notify_failed",
+        "D2: attach requeue must flip delivery_exhausted -> 'notify_failed' (golden result_watchers.py:95), not 'pending'");
+    // D1 ✦ team-scoped: cross-team exhausted watcher must NOT requeue onto team-a's pane.
+    let (st_b, _) = watcher_state(&store, &team_b);
+    assert_eq!(st_b, "delivery_exhausted",
+        "D1 ✦: team-scoped selection — a team-b exhausted watcher must NOT be requeued by a team-a attach (anti cross-team pollution / CP-1)");
+    // Gap-32: a notified watcher is never requeued; its notified_message_id survives.
+    let (st_n, nid) = watcher_state(&store, &notif);
+    assert_eq!(st_n, "delivery_exhausted", "Gap-32: notified watcher not requeued");
+    assert_eq!(nid.as_deref(), Some("msg_done"), "Gap-32: notified_message_id preserved");
+    // only the team-a unnotified exhausted watcher requeues.
+    let ids: Vec<&str> = requeued.iter().map(|n| n.watcher_id.as_str()).collect();
+    assert_eq!(ids, vec!["w-r8"], "only team-a unnotified delivery_exhausted watcher requeues");
+
+    // D3: result_watcher.requeued payload == golden ATTACH form {watcher_id, trigger, new_pane_id}.
+    let events = log.tail(0).unwrap();
+    let ev = events.iter().rev()
+        .find(|e| e.get("event").and_then(|v| v.as_str()) == Some("result_watcher.requeued"))
+        .expect("result_watcher.requeued event");
+    let keys: std::collections::BTreeSet<&str> = ev.as_object().unwrap().keys()
+        .map(String::as_str).filter(|k| *k != "ts" && *k != "event").collect();
+    let expected: std::collections::BTreeSet<&str> = ["watcher_id", "trigger", "new_pane_id"].into_iter().collect();
+    assert_eq!(keys, expected,
+        "D3: result_watcher.requeued must be golden ATTACH form {{watcher_id, trigger, new_pane_id}} (leader/__init__.py:46-50), not claim-style; got {keys:?}");
+    assert_eq!(ev.get("trigger").and_then(|v| v.as_str()), Some("attach_leader"));
+    assert_eq!(ev.get("new_pane_id").and_then(|v| v.as_str()), Some("%leader-new"));
+}