wip(k1): checkpoint connection recovery rewrite

Capture the current unreleased K1 connection, recovery, lifecycle, viewer, and test work as a single known-bad baseline for subsequent fixes.

src/k1link/artifacts.py

@@ -14,7 +14,12 @@ def utc_now_iso() -> str:
 
 
 def write_json_atomic(path: Path, payload: Any) -> None:
-    """Write JSON without exposing a partially written artifact."""
+    """Write JSON without exposing or acknowledging a partial commit.
+
+    Flushing the temporary file protects its contents, but a crash can still
+    lose the directory entry created by ``replace``. The parent directory is
+    therefore flushed after the atomic rename as the second durability edge.
+    """
     path = path.expanduser()
     path.parent.mkdir(parents=True, exist_ok=True)
     serialized = json.dumps(payload, ensure_ascii=False, indent=2) + "\n"
@@ -34,6 +39,13 @@ def write_json_atomic(path: Path, payload: Any) -> None:
             stream.flush()
             os.fsync(stream.fileno())
         Path(temp_name).replace(path)
+        directory_flags = os.O_RDONLY | getattr(os, "O_CLOEXEC", 0)
+        directory_flags |= getattr(os, "O_DIRECTORY", 0)
+        directory_descriptor = os.open(path.parent, directory_flags)
+        try:
+            os.fsync(directory_descriptor)
+        finally:
+            os.close(directory_descriptor)
     finally:
         if temp_name is not None:
             Path(temp_name).unlink(missing_ok=True)

src/k1link/compute/l34a_assisted_yolox_error_audit.py

@@ -142,7 +142,10 @@ def build_l34a_assisted_yolox_error_audit(
         },
         "limitations": [
             "the review was seeded from the same frozen candidate and is not independent truth",
-            "precision, recall and F1 are assisted diagnostic alignment metrics, not acceptance metrics",
+            (
+                "precision, recall and F1 are assisted diagnostic alignment metrics, "
+                "not acceptance metrics"
+            ),
             "custom labels are retained as proposed ontology terms and were not adjudicated",
             "the result is source-scoped to 32 RAVNOVES00 right-camera frames",
         ],

src/k1link/compute/live_perception.py

@@ -10,7 +10,7 @@ import threading
 import time
 import zlib
 from collections import deque
-from collections.abc import Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence
 from dataclasses import dataclass
 from hashlib import sha256
 from typing import Any, Final, Literal
@@ -33,7 +33,7 @@ LiveIngressModality = Literal[
 
 LIVE_INGRESS_SCHEMA: Final = "missioncore.live-perception-ingress/v1"
 LIVE_INGRESS_WIRE_SCHEMA: Final = "missioncore.live-perception-wire/v1"
-LIVE_RESULT_WIRE_SCHEMA: Final = "missioncore.live-perception-result-wire/v1"
+LIVE_RESULT_WIRE_SCHEMA: Final = "missioncore.live-perception-result-wire/v2"
 LIVE_RESULT_MAGIC: Final = b"MCPR"
 LIVE_RESULT_MAX_HEADER_BYTES: Final = 256 * 1024
 LIVE_RESULT_MAX_PAYLOAD_BYTES: Final = 2 * 1024 * 1024
@@ -41,6 +41,8 @@ LIVE_RESULT_MAX_PAYLOAD_BYTES: Final = 2 * 1024 * 1024
 
 @dataclass(frozen=True, slots=True)
 class LivePerceptionResultFrame:
+    session_id: str
+    session_generation: int
     frame_index: int
     source_frame_index: int
     session_seconds: float
@@ -53,6 +55,8 @@ class LivePerceptionResultFrame:
 
 def encode_live_perception_result(
     *,
+    session_id: str,
+    session_generation: int,
     frame_index: int,
     source_frame_index: int,
     session_seconds: float,
@@ -65,7 +69,11 @@ def encode_live_perception_result(
     """Encode one bounded, non-authoritative worker-to-viewer result frame."""
 
     if (
-        frame_index < 0
+        not session_id
+        or len(session_id) > 160
+        or isinstance(session_generation, bool)
+        or session_generation < 1
+        or frame_index < 0
         or source_frame_index < 0
         or captured_at_epoch_ns < 0
         or not math.isfinite(session_seconds)
@@ -91,6 +99,8 @@ def encode_live_perception_result(
         raise ValueError("live perception result payload exceeds the bound")
     header = {
         "schema_version": LIVE_RESULT_WIRE_SCHEMA,
+        "session_id": session_id,
+        "session_generation": session_generation,
         "frame_index": frame_index,
         "source_frame_index": source_frame_index,
         "session_seconds": session_seconds,
@@ -192,11 +202,19 @@ def decode_live_perception_result(encoded: bytes) -> LivePerceptionResultFrame:
         mask = np.frombuffer(raw_mask, dtype=np.uint8).reshape((600, 800)).copy()
     normalized_objects = tuple(_normalize_live_result_object(value) for value in objects)
     frame_index = header.get("frame_index")
+    session_id = header.get("session_id")
+    session_generation = header.get("session_generation")
     source_frame_index = header.get("source_frame_index")
     session_seconds = header.get("session_seconds")
     captured_at_epoch_ns = header.get("captured_at_epoch_ns")
     if (
-        not isinstance(frame_index, int)
+        not isinstance(session_id, str)
+        or not session_id
+        or len(session_id) > 160
+        or not isinstance(session_generation, int)
+        or isinstance(session_generation, bool)
+        or session_generation < 1
+        or not isinstance(frame_index, int)
         or isinstance(frame_index, bool)
         or frame_index < 0
         or not isinstance(source_frame_index, int)
@@ -212,6 +230,8 @@ def decode_live_perception_result(encoded: bytes) -> LivePerceptionResultFrame:
     ):
         raise ValueError("live perception result time identity is invalid")
     return LivePerceptionResultFrame(
+        session_id=session_id,
+        session_generation=session_generation,
         frame_index=frame_index,
         source_frame_index=source_frame_index,
         session_seconds=float(session_seconds),
@@ -297,6 +317,7 @@ class LiveIngressEvent:
 
     ingress_sequence: int
     session_id: str
+    session_generation: int
     modality: LiveIngressModality
     source_id: str
     source_sequence: int
@@ -310,6 +331,7 @@ class LiveIngressEvent:
                 "schema_version": LIVE_INGRESS_WIRE_SCHEMA,
                 "ingress_sequence": self.ingress_sequence,
                 "session_id": self.session_id,
+                "session_generation": self.session_generation,
                 "modality": self.modality,
                 "source_id": self.source_id,
                 "source_sequence": self.source_sequence,
@@ -382,9 +404,13 @@ class LivePerceptionIngress:
         }
         self._ingress_sequence = 0
         self._session_id: str | None = None
+        self._session_generation = 0
         self._active = False
         self._closed = False
         self._consumer_id: str | None = None
+        self._results_accepted = 0
+        self._results_rejected_stale = 0
+        self._results_rejected_receiver = 0
 
     def begin_session(self, session_id: str) -> None:
         if not session_id or len(session_id) > 160:
@@ -402,6 +428,7 @@ class LivePerceptionIngress:
             for queue in self._queues.values():
                 queue.items.clear()
             self._session_id = session_id
+            self._session_generation += 1
             self._active = True
             self._publish_locked(
                 modality="control",
@@ -470,6 +497,31 @@ class LivePerceptionIngress:
                 self._consumer_id = None
                 self._condition.notify_all()
 
+    def admit_result(
+        self,
+        *,
+        session_id: str,
+        session_generation: int,
+        receiver: Callable[[], bool],
+    ) -> bool:
+        """Atomically reject a late worker result before viewer admission."""
+
+        with self._condition:
+            if (
+                self._closed
+                or not self._active
+                or self._session_id != session_id
+                or self._session_generation != session_generation
+            ):
+                self._results_rejected_stale += 1
+                return False
+            accepted = receiver()
+            if accepted:
+                self._results_accepted += 1
+            else:
+                self._results_rejected_receiver += 1
+            return accepted
+
     def take_next(
         self,
         consumer_id: str,
@@ -509,7 +561,11 @@ class LivePerceptionIngress:
                 "mode": "shadow-diagnostic-only",
                 "active": self._active,
                 "session_id": self._session_id,
+                "session_generation": self._session_generation,
                 "consumer_connected": self._consumer_id is not None,
+                "results_accepted": self._results_accepted,
+                "results_rejected_stale": self._results_rejected_stale,
+                "results_rejected_receiver": self._results_rejected_receiver,
                 "commands_enabled": False,
                 "navigation_or_safety_accepted": False,
                 "closed": self._closed,
@@ -548,6 +604,7 @@ class LivePerceptionIngress:
         event = LiveIngressEvent(
             ingress_sequence=self._ingress_sequence,
             session_id=session_id,
+            session_generation=self._session_generation,
             modality=modality,
             source_id=source_id,
             source_sequence=source_sequence,

src/k1link/device_plugins/xgrids_k1/application_control_process_lease.py

@@ -0,0 +1,265 @@
+from __future__ import annotations
+
+import fcntl
+import os
+import stat
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Literal
+
+from k1link.sessions.store import resolve_missioncore_data_dir
+
+APPLICATION_CONTROL_LOCK_FILENAME = ".application-control.lock"
+
+
+class ApplicationControlProcessLeaseError(RuntimeError):
+    """The process-wide K1 control fence cannot be trusted."""
+
+    reason_code = "application-control-process-lease-error"
+
+
+class ApplicationControlProcessLeaseUnavailable(ApplicationControlProcessLeaseError):
+    """Another Mission Core process owns the canonical K1 control dialogue."""
+
+    reason_code = "application-control-process-lease-unavailable"
+
+
+ApplicationControlProcessLeaseReleaseState = Literal["owned", "released", "ambiguous"]
+ApplicationControlProcessLeaseReleaseDisposition = Literal[
+    "released",
+    "already-released",
+]
+
+
+class ApplicationControlProcessLeaseReleaseAmbiguous(ApplicationControlProcessLeaseError):
+    """Neither strict unlock nor close proved the OS fence disposition."""
+
+    reason_code = "application-control-process-lease-release-ambiguous"
+
+
+@dataclass(frozen=True, slots=True)
+class ApplicationControlProcessLeaseReleaseOutcome:
+    """Terminal release result, including non-retryable syscall diagnostics."""
+
+    disposition: ApplicationControlProcessLeaseReleaseDisposition
+    unlock_error_code: str | None = None
+    close_error_code: str | None = None
+
+
+@dataclass(slots=True)
+class ApplicationControlProcessLease:
+    """Non-persistent ownership of the one canonical K1 control dialogue.
+
+    The lock file is deliberately stable while ownership lives only in the OS
+    lock attached to ``_descriptor``.  A process crash normally releases
+    control admission without manufacturing any K1 command or durable recovery
+    claim.  A target-owning child may deliberately inherit a duplicate of that
+    exact description; after a parent crash the kernel then preserves the fence
+    until the child itself exits.
+    """
+
+    path: Path
+    _descriptor: int
+    _identity: tuple[int, int]
+    _released: bool = False
+    _release_ambiguous: bool = False
+
+    @property
+    def release_state(self) -> ApplicationControlProcessLeaseReleaseState:
+        if self._release_ambiguous:
+            return "ambiguous"
+        return "released" if self._released else "owned"
+
+    @classmethod
+    def acquire(cls, repository_root: Path) -> ApplicationControlProcessLease:
+        data_dir = resolve_missioncore_data_dir(repository_root)
+        lock_dir = data_dir / "xgrids-k1"
+        _ensure_private_directory(data_dir, parents=True)
+        _ensure_private_directory(lock_dir, parents=False)
+        path = lock_dir / APPLICATION_CONTROL_LOCK_FILENAME
+
+        flags = os.O_RDWR | os.O_CREAT | getattr(os, "O_CLOEXEC", 0)
+        flags |= getattr(os, "O_NOFOLLOW", 0)
+        try:
+            descriptor = os.open(path, flags, 0o600)
+        except OSError as exc:
+            raise ApplicationControlProcessLeaseError(
+                "application control process lock cannot be opened safely"
+            ) from exc
+
+        locked = False
+        try:
+            opened = os.fstat(descriptor)
+            _validate_private_lock_file(opened)
+            try:
+                current = path.lstat()
+            except OSError as exc:
+                raise ApplicationControlProcessLeaseError(
+                    "application control process lock identity is unavailable"
+                ) from exc
+            _validate_private_lock_file(current)
+            identity = (opened.st_dev, opened.st_ino)
+            if identity != (current.st_dev, current.st_ino):
+                raise ApplicationControlProcessLeaseError(
+                    "application control process lock changed while opening"
+                )
+            try:
+                fcntl.flock(descriptor, fcntl.LOCK_EX | fcntl.LOCK_NB)
+                locked = True
+            except BlockingIOError as exc:
+                raise ApplicationControlProcessLeaseUnavailable(
+                    "another Mission Core process owns K1 application control"
+                ) from exc
+            except OSError as exc:
+                raise ApplicationControlProcessLeaseError(
+                    "application control process lock cannot be acquired safely"
+                ) from exc
+            try:
+                locked_path = path.lstat()
+            except OSError as exc:
+                raise ApplicationControlProcessLeaseError(
+                    "application control process lock disappeared after acquisition"
+                ) from exc
+            _validate_private_lock_file(locked_path)
+            if identity != (locked_path.st_dev, locked_path.st_ino):
+                raise ApplicationControlProcessLeaseError(
+                    "application control process lock changed during acquisition"
+                )
+            return cls(path=path, _descriptor=descriptor, _identity=identity)
+        except BaseException:
+            if locked:
+                _unlock_descriptor(descriptor)
+            os.close(descriptor)
+            raise
+
+    def release(self) -> ApplicationControlProcessLeaseReleaseOutcome:
+        if self._released:
+            return ApplicationControlProcessLeaseReleaseOutcome(
+                disposition="already-released"
+            )
+        if self._release_ambiguous:
+            raise ApplicationControlProcessLeaseReleaseAmbiguous(
+                "application control process lock release remains ambiguous"
+            )
+        unlock_error: OSError | None = None
+        close_error: OSError | None = None
+        try:
+            fcntl.flock(self._descriptor, fcntl.LOCK_UN)
+        except OSError as exc:
+            unlock_error = exc
+        try:
+            os.close(self._descriptor)
+        except OSError as exc:
+            close_error = exc
+        if unlock_error is not None and close_error is not None:
+            # There is no portable ownership answer after two failing syscalls.
+            # Never retry either syscall on this descriptor: quarantine the
+            # process-local object and require process restart.
+            self._release_ambiguous = True
+            error = ApplicationControlProcessLeaseReleaseAmbiguous(
+                "application control process lock release outcome is ambiguous"
+            )
+            error.add_note(
+                "explicit unlock failed: "
+                f"{type(unlock_error).__name__}: {unlock_error}"
+            )
+            error.add_note(
+                "descriptor close failed: "
+                f"{type(close_error).__name__}: {close_error}"
+            )
+            raise error from close_error
+        # Either strict LOCK_UN or close proved that the flock can no longer
+        # be reused through this object.  A diagnostic on the other syscall is
+        # terminal, not a retry instruction.
+        self._released = True
+        return ApplicationControlProcessLeaseReleaseOutcome(
+            disposition="released",
+            unlock_error_code=_os_error_code(unlock_error),
+            close_error_code=_os_error_code(close_error),
+        )
+
+    def duplicate_descriptor_for_child(self) -> int:
+        """Duplicate the live flock description for one target-owning child.
+
+        ``flock`` ownership follows the open file description across ``dup``
+        and ``exec``.  A camera adapter that inherits this duplicate therefore
+        keeps the K1 lifecycle fence after an abrupt parent-process exit.  The
+        caller must pass the returned descriptor through ``pass_fds`` and
+        close its parent-side duplicate immediately after spawning.
+
+        Normal shutdown must still terminate and reap the child before calling
+        :meth:`release`: an explicit ``LOCK_UN`` on any duplicate unlocks the
+        shared description for every process.
+        """
+
+        if self._released or self._release_ambiguous:
+            raise ApplicationControlProcessLeaseError(
+                "released or quarantined application control lease cannot be inherited"
+            )
+        try:
+            descriptor = os.dup(self._descriptor)
+        except OSError as exc:
+            raise ApplicationControlProcessLeaseError(
+                "application control process lock cannot be duplicated safely"
+            ) from exc
+        try:
+            metadata = os.fstat(descriptor)
+            _validate_private_lock_file(metadata)
+            if (metadata.st_dev, metadata.st_ino) != self._identity:
+                raise ApplicationControlProcessLeaseError(
+                    "duplicated application control process lock changed identity"
+                )
+            return descriptor
+        except BaseException:
+            os.close(descriptor)
+            raise
+
+    def __enter__(self) -> ApplicationControlProcessLease:
+        return self
+
+    def __exit__(self, *_: object) -> None:
+        self.release()
+
+
+def _ensure_private_directory(path: Path, *, parents: bool) -> None:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        try:
+            path.mkdir(mode=0o700, parents=parents, exist_ok=False)
+        except FileExistsError:
+            metadata = path.lstat()
+        else:
+            metadata = path.lstat()
+    except OSError as exc:
+        raise ApplicationControlProcessLeaseError(
+            "application control process lock directory is unavailable"
+        ) from exc
+    if not stat.S_ISDIR(metadata.st_mode) or stat.S_IMODE(metadata.st_mode) != 0o700:
+        raise ApplicationControlProcessLeaseError(
+            "application control process lock directory is not private"
+        )
+
+
+def _validate_private_lock_file(metadata: os.stat_result) -> None:
+    if (
+        not stat.S_ISREG(metadata.st_mode)
+        or stat.S_IMODE(metadata.st_mode) != 0o600
+        or metadata.st_nlink != 1
+    ):
+        raise ApplicationControlProcessLeaseError(
+            "application control process lock is not a private regular file"
+        )
+
+
+def _os_error_code(error: OSError | None) -> str | None:
+    if error is None:
+        return None
+    return f"{type(error).__name__}:{error.errno if error.errno is not None else 'unknown'}"
+
+
+def _unlock_descriptor(descriptor: int) -> None:
+    try:
+        fcntl.flock(descriptor, fcntl.LOCK_UN)
+    except OSError:
+        return

src/k1link/device_plugins/xgrids_k1/ble/ap_activation.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 import asyncio
-from collections.abc import AsyncIterator
+from collections.abc import AsyncIterator, Callable
 from contextlib import AsyncExitStack, asynccontextmanager
 from importlib.metadata import version
 from time import monotonic
@@ -11,7 +11,19 @@ from bleak import BleakClient, BleakScanner
 from bleak.exc import BleakDeviceNotFoundError, BleakError
 
 from k1link.artifacts import utc_now_iso
-from k1link.device_plugins.xgrids_k1.ble.scanner import discovered_device_selection
+from k1link.device_plugins.xgrids_k1.ble.runtime_arbiter import (
+    BleOperationProgress,
+    run_ble_operation_session,
+)
+from k1link.device_plugins.xgrids_k1.ble.scanner import (
+    CapturedDiscoveredDevice,
+    captured_device_handle,
+    connected_device_capture,
+    demote_connected_device_handle_after_gatt_failure,
+    discovered_device_selection,
+    mark_captured_device_gatt_validated,
+    retrieve_connected_device_capture,
+)
 from k1link.device_plugins.xgrids_k1.ble.wifi_provisioning import (
     AP_FALLBACK_IPV4,
     SERVICE_UUID,
@@ -23,6 +35,7 @@ from k1link.device_plugins.xgrids_k1.ble.wifi_provisioning import (
     WifiStatus,
     WriteMode,
     _annotate_ble_operation_error,
+    _optional_int_attribute,
     parse_wifi_status,
 )
 
@@ -30,6 +43,7 @@ PROFILE_ID = "xgrids-k1-fw3-quick-connect-ap-v1"
 FRAME_LENGTH = 100
 COMMAND_OFFSET = 99
 ENABLE_AP_COMMAND = 1
+BLE_AP_ENABLE_HARD_TIMEOUT_GRACE_SECONDS = 25.0
 
 ApActivationOutcome = Literal[
     "already_active",
@@ -56,7 +70,7 @@ class ApActivationResult(TypedDict):
     write_performed: bool
     write_mode: ResolvedWriteMode | None
     write_without_response_advertised: bool
-    max_write_without_response_size: int
+    max_write_without_response_size: int | None
     frame_length: int
     baseline_status: WifiStatus
     observations: list[StatusObservation]
@@ -100,11 +114,16 @@ def _outcome(
 
 
 @asynccontextmanager
-async def device_ap_activation_session(
+async def _device_ap_activation_session_impl(
     device_macos_uuid: str,
     timeout_seconds: float = 15.0,
     poll_interval_seconds: float = 0.5,
     write_mode: WriteMode = "auto",
+    *,
+    captured_device: CapturedDiscoveredDevice | None,
+    recovery_device_session_id: str | None,
+    on_write_dispatch: Callable[[WifiStatus, ResolvedWriteMode], None] | None,
+    progress: BleOperationProgress,
 ) -> AsyncIterator[ApActivationResult]:
     """Keep BLE connected around one reviewed Quick Connect AP-enable write.
 
@@ -130,6 +149,12 @@ async def device_ap_activation_session(
     operation_stage: BleOperationStage = "resolution"
     device_write_attempted = False
     device_write_confirmed = False
+    resolved_write_mode_for_error: ResolvedWriteMode | None = None
+    write_characteristic_properties: tuple[str, ...] | None = None
+    max_without_response: int | None = None
+    active_captured_device: CapturedDiscoveredDevice | None = None
+    gatt_baseline_validated = False
+    progress.operation_stage = operation_stage
 
     try:
         try:
@@ -137,9 +162,44 @@ async def device_ap_activation_session(
                 # Keep the explicit scan and AP activation in one CoreBluetooth
                 # lifecycle. Re-looking up the UUID here lost a physically present
                 # K1 during acceptance, while the retained BLEDevice connected.
-                selection = discovered_device_selection(device_macos_uuid)
-                device = selection.device
-                if device is None and not selection.from_fresh_scan:
+                if captured_device is not None:
+                    device = (
+                        captured_device_handle(captured_device)
+                        if captured_device.macos_uuid.casefold()
+                        == device_macos_uuid.casefold()
+                        else None
+                    )
+                    if device is not None:
+                        active_captured_device = captured_device
+                    selection_from_fresh_scan = False
+                elif recovery_device_session_id is not None:
+                    # Retrieval is part of the same serialized ap-enable
+                    # lease as connect, baseline read and the one command.
+                    active_captured_device = connected_device_capture(
+                        device_macos_uuid,
+                        device_session_id=recovery_device_session_id,
+                    )
+                    if active_captured_device is None:
+                        active_captured_device = await retrieve_connected_device_capture(
+                            device_macos_uuid,
+                            device_session_id=recovery_device_session_id,
+                        )
+                    device = (
+                        captured_device_handle(active_captured_device)
+                        if active_captured_device is not None
+                        else None
+                    )
+                    selection_from_fresh_scan = False
+                else:
+                    selection = discovered_device_selection(device_macos_uuid)
+                    device = selection.device
+                    selection_from_fresh_scan = selection.from_fresh_scan
+                if (
+                    device is None
+                    and captured_device is None
+                    and recovery_device_session_id is None
+                    and not selection_from_fresh_scan
+                ):
                     # Preserve a bounded fallback for non-UI callers that did not
                     # establish a fresh explicit scan lease.
                     device = await BleakScanner.find_device_by_address(
@@ -157,21 +217,34 @@ async def device_ap_activation_session(
                 operation_stage=operation_stage,
                 device_write_attempted=device_write_attempted,
                 device_write_confirmed=device_write_confirmed,
+                resolved_write_mode=resolved_write_mode_for_error,
+                write_characteristic_properties=write_characteristic_properties,
+                max_write_without_response_size=max_without_response,
+                frame_length=len(frame),
             )
             raise
 
         async with AsyncExitStack() as client_stack:
             operation_stage = "connect"
+            progress.operation_stage = operation_stage
             try:
                 client = await client_stack.enter_async_context(
                     BleakClient(device, timeout=timeout_seconds, pair=False)
                 )
             except Exception as exc:
+                if active_captured_device is not None:
+                    demote_connected_device_handle_after_gatt_failure(
+                        active_captured_device
+                    )
                 _annotate_ble_operation_error(
                     exc,
                     operation_stage=operation_stage,
                     device_write_attempted=device_write_attempted,
                     device_write_confirmed=device_write_confirmed,
+                    resolved_write_mode=resolved_write_mode_for_error,
+                    write_characteristic_properties=write_characteristic_properties,
+                    max_write_without_response_size=max_without_response,
+                    frame_length=len(frame),
                 )
                 raise
 
@@ -179,6 +252,7 @@ async def device_ap_activation_session(
                 async with asyncio.timeout(timeout_seconds + 10.0):
                     device_name = client.name
                     operation_stage = "gatt-contract"
+                    progress.operation_stage = operation_stage
                     service = client.services.get_service(SERVICE_UUID)
                     write_characteristic = client.services.get_characteristic(
                         WRITE_CHARACTERISTIC_UUID
@@ -210,7 +284,11 @@ async def device_ap_activation_session(
                         raise ValueError("Reviewed K1 status characteristic is not readable")
 
                     properties = set(write_characteristic.properties)
-                    max_without_response = write_characteristic.max_write_without_response_size
+                    write_characteristic_properties = tuple(sorted(properties))
+                    max_without_response = _optional_int_attribute(
+                        write_characteristic,
+                        "max_write_without_response_size",
+                    )
                     resolved_write_mode: ResolvedWriteMode
                     if write_mode == "auto":
                         if "write-without-response" in properties:
@@ -226,17 +304,32 @@ async def device_ap_activation_session(
                             )
                         resolved_write_mode = "with_response"
                     else:
+                        resolved_write_mode = "without_response"
+
+                    resolved_write_mode_for_error = resolved_write_mode
+                    if resolved_write_mode == "without_response":
+                        if max_without_response is None:
+                            raise ValueError(
+                                "Negotiated write-without-response size is unavailable"
+                            )
                         if len(frame) > max_without_response:
                             raise ValueError(
                                 "AP activation frame exceeds the negotiated "
                                 "write-without-response size"
                             )
-                        resolved_write_mode = "without_response"
 
                     operation_stage = "baseline-read"
+                    progress.operation_stage = operation_stage
                     baseline = parse_wifi_status(
                         bytes(await client.read_gatt_char(status_characteristic))
                     )
+                    if active_captured_device is not None and not (
+                        mark_captured_device_gatt_validated(active_captured_device)
+                    ):
+                        raise RuntimeError(
+                            "Exact BLE recovery handle changed before AP validation"
+                        )
+                    gatt_baseline_validated = True
                     # WIFI_AP is a control-mode status, not proof that the radio is
                     # still beaconing. A physical run found the exact SSID shortly
                     # after AP-enable, then found no beacon while 7f02 continued to
@@ -246,17 +339,26 @@ async def device_ap_activation_session(
                     # a stale-ready status. There is still no automatic retry.
 
                     operation_stage = "gatt-write"
+                    # Persist the write barrier before handing the frame to
+                    # CoreBluetooth. The callback is deliberately synchronous
+                    # and secret-free; failure here prevents the device write.
+                    if on_write_dispatch is not None:
+                        on_write_dispatch(baseline, resolved_write_mode)
                     device_write_attempted = True
+                    progress.operation_stage = operation_stage
+                    progress.device_write_attempted = True
                     await client.write_gatt_char(
                         write_characteristic,
                         frame,
                         response=resolved_write_mode == "with_response",
                     )
                     device_write_confirmed = resolved_write_mode == "with_response"
+                    progress.device_write_confirmed = device_write_confirmed
                     write_completed = monotonic()
                     deadline = write_completed + timeout_seconds
 
                     operation_stage = "status-poll"
+                    progress.operation_stage = operation_stage
                     while monotonic() < deadline:
                         try:
                             status = parse_wifi_status(
@@ -306,11 +408,19 @@ async def device_ap_activation_session(
                         "outcome": _outcome(baseline, observations, disconnected),
                     }
             except Exception as exc:
+                if active_captured_device is not None and not gatt_baseline_validated:
+                    demote_connected_device_handle_after_gatt_failure(
+                        active_captured_device
+                    )
                 _annotate_ble_operation_error(
                     exc,
                     operation_stage=operation_stage,
                     device_write_attempted=device_write_attempted,
                     device_write_confirmed=device_write_confirmed,
+                    resolved_write_mode=resolved_write_mode_for_error,
+                    write_characteristic_properties=write_characteristic_properties,
+                    max_write_without_response_size=max_without_response,
+                    frame_length=len(frame),
                 )
                 raise
 
@@ -321,14 +431,62 @@ async def device_ap_activation_session(
             # annotated as BLE failures when they are thrown back through yield.
             yield result
     finally:
+        # The runtime arbiter enforces a hard deadline by cancellation.  Even
+        # when that bypasses the normal exception annotator, a captured
+        # recovery object that never passed baseline GATT validation must be
+        # discarded.
+        if active_captured_device is not None and not gatt_baseline_validated:
+            demote_connected_device_handle_after_gatt_failure(active_captured_device)
         frame[:] = b"\x00" * len(frame)
 
 
+@asynccontextmanager
+async def device_ap_activation_session(
+    device_macos_uuid: str,
+    timeout_seconds: float = 15.0,
+    poll_interval_seconds: float = 0.5,
+    write_mode: WriteMode = "auto",
+    *,
+    captured_device: CapturedDiscoveredDevice | None = None,
+    recovery_device_session_id: str | None = None,
+    on_write_dispatch: Callable[[WifiStatus, ResolvedWriteMode], None] | None = None,
+) -> AsyncIterator[ApActivationResult]:
+    """Own one process BLE lease through AP-ready and host Wi-Fi handoff."""
+
+    if captured_device is not None and recovery_device_session_id is not None:
+        raise ValueError(
+            "captured_device and recovery_device_session_id are mutually exclusive"
+        )
+    if recovery_device_session_id == "":
+        raise ValueError("recovery_device_session_id must not be empty")
+    progress = BleOperationProgress(operation_stage="resolution")
+    async with run_ble_operation_session(
+        "ap-enable",
+        hard_setup_timeout_seconds=(timeout_seconds + BLE_AP_ENABLE_HARD_TIMEOUT_GRACE_SECONDS),
+        operation=lambda operation_progress: _device_ap_activation_session_impl(
+            device_macos_uuid,
+            timeout_seconds=timeout_seconds,
+            poll_interval_seconds=poll_interval_seconds,
+            write_mode=write_mode,
+            captured_device=captured_device,
+            recovery_device_session_id=recovery_device_session_id,
+            on_write_dispatch=on_write_dispatch,
+            progress=operation_progress,
+        ),
+        progress=progress,
+    ) as result:
+        yield result
+
+
 async def activate_device_ap_once(
     device_macos_uuid: str,
     timeout_seconds: float = 15.0,
     poll_interval_seconds: float = 0.5,
     write_mode: WriteMode = "auto",
+    *,
+    captured_device: CapturedDiscoveredDevice | None = None,
+    recovery_device_session_id: str | None = None,
+    on_write_dispatch: Callable[[WifiStatus, ResolvedWriteMode], None] | None = None,
 ) -> ApActivationResult:
     """Run one AP activation and release BLE immediately after its result.
 
@@ -341,5 +499,8 @@ async def activate_device_ap_once(
         timeout_seconds=timeout_seconds,
         poll_interval_seconds=poll_interval_seconds,
         write_mode=write_mode,
+        captured_device=captured_device,
+        recovery_device_session_id=recovery_device_session_id,
+        on_write_dispatch=on_write_dispatch,
     ) as result:
         return result

src/k1link/device_plugins/xgrids_k1/ble/gatt.py

@@ -8,6 +8,10 @@ from bleak import BleakClient, BleakScanner
 from bleak.exc import BleakDeviceNotFoundError
 
 from k1link.artifacts import utc_now_iso
+from k1link.device_plugins.xgrids_k1.ble.runtime_arbiter import (
+    BleOperationProgress,
+    run_ble_operation,
+)
 
 
 class DescriptorRecord(TypedDict):
@@ -48,57 +52,68 @@ async def dump_metadata(device_macos_uuid: str, timeout_seconds: float) -> GattD
     if timeout_seconds <= 0:
         raise ValueError("timeout_seconds must be positive")
 
-    started_at = utc_now_iso()
-    async with asyncio.timeout(timeout_seconds):
-        device = await BleakScanner.find_device_by_address(
-            device_macos_uuid,
-            timeout=min(20.0, timeout_seconds),
-        )
-        if device is None:
-            raise BleakDeviceNotFoundError(
+    async def perform(progress: BleOperationProgress) -> GattDumpResult:
+        started_at = utc_now_iso()
+        async with asyncio.timeout(timeout_seconds):
+            progress.operation_stage = "gatt-metadata-discovery"
+            device = await BleakScanner.find_device_by_address(
                 device_macos_uuid,
-                "Device was not rediscovered; keep the K1 powered and nearby.",
+                timeout=min(20.0, timeout_seconds),
             )
+            if device is None:
+                raise BleakDeviceNotFoundError(
+                    device_macos_uuid,
+                    "Device was not rediscovered; keep the K1 powered and nearby.",
+                )
 
-        async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
-            services: list[ServiceRecord] = []
-            for service in client.services:
-                characteristics: list[CharacteristicRecord] = []
-                for characteristic in service.characteristics:
-                    descriptors: list[DescriptorRecord] = []
-                    for descriptor in characteristic.descriptors:
-                        descriptors.append(
+            progress.operation_stage = "gatt-metadata-connect"
+            async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
+                progress.operation_stage = "gatt-metadata-enumeration"
+                services: list[ServiceRecord] = []
+                for service in client.services:
+                    characteristics: list[CharacteristicRecord] = []
+                    for characteristic in service.characteristics:
+                        descriptors: list[DescriptorRecord] = []
+                        for descriptor in characteristic.descriptors:
+                            descriptors.append(
+                                {
+                                    "uuid": descriptor.uuid,
+                                    "handle": descriptor.handle,
+                                    "description": descriptor.description,
+                                }
+                            )
+                        characteristics.append(
                             {
-                                "uuid": descriptor.uuid,
-                                "handle": descriptor.handle,
-                                "description": descriptor.description,
+                                "uuid": characteristic.uuid,
+                                "handle": characteristic.handle,
+                                "description": characteristic.description,
+                                "properties": sorted(characteristic.properties),
+                                "descriptors": descriptors,
                             }
                         )
-                    characteristics.append(
+                    services.append(
                         {
-                            "uuid": characteristic.uuid,
-                            "handle": characteristic.handle,
-                            "description": characteristic.description,
-                            "properties": sorted(characteristic.properties),
-                            "descriptors": descriptors,
+                            "uuid": service.uuid,
+                            "handle": service.handle,
+                            "description": service.description,
+                            "characteristics": characteristics,
                         }
                     )
-                services.append(
-                    {
-                        "uuid": service.uuid,
-                        "handle": service.handle,
-                        "description": service.description,
-                        "characteristics": characteristics,
-                    }
-                )
-            return {
-                "schema_version": 1,
-                "started_at_utc": started_at,
-                "completed_at_utc": utc_now_iso(),
-                "adapter": "CoreBluetooth",
-                "bleak_version": version("bleak"),
-                "device_macos_uuid": device_macos_uuid,
-                "device_name": client.name,
-                "metadata_only": True,
-                "services": services,
-            }
+                progress.operation_stage = "gatt-metadata-complete"
+                return {
+                    "schema_version": 1,
+                    "started_at_utc": started_at,
+                    "completed_at_utc": utc_now_iso(),
+                    "adapter": "CoreBluetooth",
+                    "bleak_version": version("bleak"),
+                    "device_macos_uuid": device_macos_uuid,
+                    "device_name": client.name,
+                    "metadata_only": True,
+                    "services": services,
+                }
+
+    return await run_ble_operation(
+        "status-read",
+        hard_timeout_seconds=timeout_seconds + 5.0,
+        operation=perform,
+    )

src/k1link/device_plugins/xgrids_k1/ble/reader.py

@@ -8,6 +8,10 @@ from bleak import BleakClient, BleakScanner
 from bleak.exc import BleakDeviceNotFoundError
 
 from k1link.artifacts import utc_now_iso
+from k1link.device_plugins.xgrids_k1.ble.runtime_arbiter import (
+    BleOperationProgress,
+    run_ble_operation,
+)
 
 
 class CharacteristicReadResult(TypedDict):
@@ -33,36 +37,47 @@ async def read_characteristic_once(
     if timeout_seconds <= 0:
         raise ValueError("timeout_seconds must be positive")
 
-    started_at = utc_now_iso()
-    async with asyncio.timeout(timeout_seconds):
-        device = await BleakScanner.find_device_by_address(
-            device_macos_uuid,
-            timeout=min(20.0, timeout_seconds),
-        )
-        if device is None:
-            raise BleakDeviceNotFoundError(
+    async def perform(progress: BleOperationProgress) -> CharacteristicReadResult:
+        started_at = utc_now_iso()
+        async with asyncio.timeout(timeout_seconds):
+            progress.operation_stage = "gatt-characteristic-discovery"
+            device = await BleakScanner.find_device_by_address(
                 device_macos_uuid,
-                "Device was not rediscovered; keep the K1 powered and nearby.",
+                timeout=min(20.0, timeout_seconds),
             )
+            if device is None:
+                raise BleakDeviceNotFoundError(
+                    device_macos_uuid,
+                    "Device was not rediscovered; keep the K1 powered and nearby.",
+                )
 
-        async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
-            characteristic = client.services.get_characteristic(characteristic_uuid)
-            if characteristic is None:
-                raise ValueError(f"Characteristic not found: {characteristic_uuid}")
-            if "read" not in characteristic.properties:
-                raise ValueError(f"Characteristic is not readable: {characteristic_uuid}")
-            value = bytes(await client.read_gatt_char(characteristic))
+            progress.operation_stage = "gatt-characteristic-connect"
+            async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
+                characteristic = client.services.get_characteristic(characteristic_uuid)
+                if characteristic is None:
+                    raise ValueError(f"Characteristic not found: {characteristic_uuid}")
+                if "read" not in characteristic.properties:
+                    raise ValueError(f"Characteristic is not readable: {characteristic_uuid}")
+                progress.operation_stage = "gatt-characteristic-read"
+                value = bytes(await client.read_gatt_char(characteristic))
 
-            return {
-                "schema_version": 1,
-                "started_at_utc": started_at,
-                "completed_at_utc": utc_now_iso(),
-                "adapter": "CoreBluetooth",
-                "bleak_version": version("bleak"),
-                "device_macos_uuid": device_macos_uuid,
-                "device_name": client.name,
-                "characteristic_uuid": characteristic.uuid,
-                "operation": "single_gatt_read_no_pair_no_write",
-                "value_length": len(value),
-                "value_hex": value.hex(),
-            }
+                progress.operation_stage = "gatt-characteristic-complete"
+                return {
+                    "schema_version": 1,
+                    "started_at_utc": started_at,
+                    "completed_at_utc": utc_now_iso(),
+                    "adapter": "CoreBluetooth",
+                    "bleak_version": version("bleak"),
+                    "device_macos_uuid": device_macos_uuid,
+                    "device_name": client.name,
+                    "characteristic_uuid": characteristic.uuid,
+                    "operation": "single_gatt_read_no_pair_no_write",
+                    "value_length": len(value),
+                    "value_hex": value.hex(),
+                }
+
+    return await run_ble_operation(
+        "status-read",
+        hard_timeout_seconds=timeout_seconds + 5.0,
+        operation=perform,
+    )

src/k1link/device_plugins/xgrids_k1/ble/runtime_arbiter.py

@@ -0,0 +1,745 @@
+from __future__ import annotations
+
+import asyncio
+import threading
+from collections.abc import AsyncIterator, Callable, Coroutine, Iterator
+from contextlib import AbstractAsyncContextManager, asynccontextmanager, contextmanager
+from contextvars import ContextVar
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Literal, TypedDict
+
+from k1link.device_plugins.xgrids_k1.application_control_process_lease import (
+    APPLICATION_CONTROL_LOCK_FILENAME,
+    ApplicationControlProcessLease,
+)
+from k1link.sessions.store import resolve_missioncore_data_dir
+
+BleOperationKind = Literal["scan", "status-read", "wifi-provision", "ap-enable"]
+BleRuntimeIdleCallbackDisposition = Literal["released", "deferred", "poisoned"]
+
+
+class BleRuntimeBusy(RuntimeError):
+    """Another process-owned CoreBluetooth operation still owns the adapter."""
+
+    def __init__(
+        self,
+        *,
+        active_operation_kind: BleOperationKind,
+        cleanup_pending: bool,
+    ) -> None:
+        message = (
+            "предыдущая BLE-операция ещё завершает безопасную очистку"
+            if cleanup_pending
+            else "другая BLE-операция уже выполняется в этом процессе"
+        )
+        super().__init__(message)
+        self.reason_code = "ble-runtime-cleanup-pending" if cleanup_pending else "ble-runtime-busy"
+        self.active_operation_kind = active_operation_kind
+        self.cleanup_pending = cleanup_pending
+
+
+class BleRuntimeOwnerLoopConflict(RuntimeError):
+    """The process BLE runtime is still owned by another live event loop."""
+
+    def __init__(self) -> None:
+        super().__init__("BLE runtime уже привязан к другому активному event loop")
+        self.reason_code = "ble-runtime-owner-loop-conflict"
+
+
+class BleRuntimePoisoned(RuntimeError):
+    """A closed owner loop abandoned an operation whose cleanup is unproved."""
+
+    def __init__(self) -> None:
+        super().__init__(
+            "BLE runtime требует перезапуска: предыдущая очистка CoreBluetooth не подтверждена"
+        )
+        self.reason_code = "ble-runtime-restart-required"
+
+
+class BleRuntimeProcessLeaseNotConfigured(RuntimeError):
+    """Low-level BLE was entered before its canonical OS lock was configured."""
+
+    def __init__(self) -> None:
+        super().__init__("BLE runtime process lease repository root is not configured")
+        self.reason_code = "ble-runtime-process-lease-not-configured"
+
+
+class BleRuntimeProcessLeaseConfigurationConflict(RuntimeError):
+    """The canonical OS lock target changed while BLE ownership was live."""
+
+    def __init__(self) -> None:
+        super().__init__("BLE runtime process lease configuration cannot change while active")
+        self.reason_code = "ble-runtime-process-lease-configuration-conflict"
+
+
+class BleRuntimeProcessLeaseBorrowInvalid(RuntimeError):
+    """An explicit higher-level OS lease borrow is stale or mismatched."""
+
+    def __init__(self) -> None:
+        super().__init__("BLE runtime process lease borrow is not active for this lock target")
+        self.reason_code = "ble-runtime-process-lease-borrow-invalid"
+
+
+@dataclass(slots=True)
+class BleOperationProgress:
+    """Non-secret facts copied onto a hard-deadline exception."""
+
+    operation_stage: str = "pending"
+    device_write_attempted: bool = False
+    device_write_confirmed: bool = False
+    owner_epoch: int = 0
+
+
+class BleOperationHardTimeout(TimeoutError):
+    """A hard caller deadline elapsed while cleanup continues in the background."""
+
+    def __init__(
+        self,
+        operation_kind: BleOperationKind,
+        progress: BleOperationProgress,
+    ) -> None:
+        super().__init__("BLE-операция не завершилась в отведённое время")
+        self.reason_code = {
+            "scan": "ble-discovery-timeout",
+            "status-read": "ble-status-read-timeout",
+            "wifi-provision": "ble-provisioning-timeout",
+            "ap-enable": "ble-ap-enable-timeout",
+        }[operation_kind]
+        self.operation_kind = operation_kind
+        self.operation_stage = progress.operation_stage
+        self.device_write_attempted = progress.device_write_attempted
+        self.device_write_confirmed = progress.device_write_confirmed
+
+
+class BleRuntimeSnapshot(TypedDict):
+    owner_epoch: int
+    owner_loop_bound: bool
+    active_operation_kind: BleOperationKind | None
+    cleanup_pending: bool
+    poisoned: bool
+
+
+@dataclass(slots=True)
+class BleRuntimeProcessLeaseBorrowToken:
+    """An opaque, context-local proof that a higher layer owns the OS lease."""
+
+    repository_root: Path
+    lock_path: Path
+    configuration_epoch: int
+    _lease: ApplicationControlProcessLease
+    _active: bool = True
+
+
+@dataclass(slots=True)
+class _ActiveLease:
+    token: int
+    operation_kind: BleOperationKind
+    owner_epoch: int
+    cleanup_pending: bool = False
+    task: asyncio.Task[Any] | None = None
+    owned_process_lease: ApplicationControlProcessLease | None = None
+    borrowed_process_lease: ApplicationControlProcessLease | None = None
+
+
+class _ProcessBleRuntimeArbiter:
+    """One fail-fast process lease shared by every supported BLE entrypoint."""
+
+    def __init__(self) -> None:
+        self._lock = threading.Lock()
+        self._owner_loop: asyncio.AbstractEventLoop | None = None
+        self._owner_epoch = 0
+        self._next_token = 0
+        self._active: _ActiveLease | None = None
+        self._poisoned = False
+        self._idle_callbacks: list[Callable[[], None]] = []
+        self._process_lease_repository_root: Path | None = None
+        self._process_lease_lock_path: Path | None = None
+        self._process_lease_configuration_epoch = 0
+
+    def configure_process_lease(self, repository_root: Path) -> int:
+        """Pin every low-level BLE entrypoint to one canonical OS lock file."""
+
+        resolved_root = repository_root.expanduser().resolve()
+        lock_path = _process_lease_path(resolved_root)
+        with self._lock:
+            if (
+                self._process_lease_repository_root == resolved_root
+                and self._process_lease_lock_path == lock_path
+            ):
+                return self._process_lease_configuration_epoch
+            if self._active is not None or self._poisoned or self._idle_callbacks:
+                raise BleRuntimeProcessLeaseConfigurationConflict()
+            self._process_lease_configuration_epoch += 1
+            self._process_lease_repository_root = resolved_root
+            self._process_lease_lock_path = lock_path
+            return self._process_lease_configuration_epoch
+
+    def create_process_lease_borrow(
+        self,
+        lease: ApplicationControlProcessLease,
+    ) -> BleRuntimeProcessLeaseBorrowToken:
+        """Validate an already-held higher-level lease before context borrowing."""
+
+        with self._lock:
+            repository_root = self._process_lease_repository_root
+            lock_path = self._process_lease_lock_path
+            if repository_root is None or lock_path is None:
+                raise BleRuntimeProcessLeaseNotConfigured()
+            if lease.path != lock_path or lease.release_state != "owned":
+                raise BleRuntimeProcessLeaseBorrowInvalid()
+            return BleRuntimeProcessLeaseBorrowToken(
+                repository_root=repository_root,
+                lock_path=lock_path,
+                configuration_epoch=self._process_lease_configuration_epoch,
+                _lease=lease,
+            )
+
+    def deactivate_process_lease_borrow(
+        self,
+        borrow: BleRuntimeProcessLeaseBorrowToken,
+    ) -> None:
+        with self._lock:
+            borrow._active = False
+
+    def bind_owner_loop(self, loop: asyncio.AbstractEventLoop) -> int:
+        if loop.is_closed():
+            raise BleRuntimeOwnerLoopConflict()
+        with self._lock:
+            current = self._owner_loop
+            if current is loop:
+                # Facade dispatch binds every action so state/STOP/close remain
+                # available after an unproved BLE teardown. Only the next BLE
+                # lease acquisition is rejected while this loop stays owner.
+                return self._owner_epoch
+            if self._poisoned:
+                raise BleRuntimePoisoned()
+            if current is not None and not current.is_closed():
+                raise BleRuntimeOwnerLoopConflict()
+            if self._active is not None:
+                # Its done callback cannot be trusted after the owning loop has
+                # closed. Never admit a replacement native CoreBluetooth task.
+                self._poisoned = True
+                raise BleRuntimePoisoned()
+            self._owner_epoch += 1
+            self._owner_loop = loop
+            return self._owner_epoch
+
+    def owner_epoch_for_loop(self, loop: asyncio.AbstractEventLoop) -> int | None:
+        with self._lock:
+            if self._poisoned or self._owner_loop is not loop or loop.is_closed():
+                return None
+            return self._owner_epoch
+
+    def invalidate_owner_loop(
+        self,
+        expected_loop: asyncio.AbstractEventLoop | None,
+    ) -> int:
+        with self._lock:
+            if expected_loop is not None and self._owner_loop is not expected_loop:
+                return self._owner_epoch
+            self._owner_epoch += 1
+            self._owner_loop = None
+            if self._active is not None:
+                self._poisoned = True
+            return self._owner_epoch
+
+    def acquire(
+        self,
+        operation_kind: BleOperationKind,
+        *,
+        owner_epoch: int,
+        process_lease_borrow: BleRuntimeProcessLeaseBorrowToken | None,
+    ) -> int:
+        with self._lock:
+            if self._poisoned:
+                raise BleRuntimePoisoned()
+            if self._owner_epoch != owner_epoch:
+                raise BleRuntimeOwnerLoopConflict()
+            if self._active is not None:
+                raise BleRuntimeBusy(
+                    active_operation_kind=self._active.operation_kind,
+                    cleanup_pending=self._active.cleanup_pending,
+                )
+            repository_root = self._process_lease_repository_root
+            lock_path = self._process_lease_lock_path
+            if repository_root is None or lock_path is None:
+                raise BleRuntimeProcessLeaseNotConfigured()
+
+            owned_process_lease: ApplicationControlProcessLease | None = None
+            borrowed_process_lease: ApplicationControlProcessLease | None = None
+            if process_lease_borrow is None:
+                # Detect an environment-driven data-root change before opening
+                # or creating a lock at a different path than configuration.
+                if _process_lease_path(repository_root) != lock_path:
+                    raise BleRuntimeProcessLeaseConfigurationConflict()
+                owned_process_lease = ApplicationControlProcessLease.acquire(repository_root)
+                if owned_process_lease.path != lock_path:
+                    owned_process_lease.release()
+                    raise BleRuntimeProcessLeaseConfigurationConflict()
+            else:
+                if (
+                    not process_lease_borrow._active
+                    or process_lease_borrow.configuration_epoch
+                    != self._process_lease_configuration_epoch
+                    or process_lease_borrow.repository_root != repository_root
+                    or process_lease_borrow.lock_path != lock_path
+                    or process_lease_borrow._lease.path != lock_path
+                    or process_lease_borrow._lease.release_state != "owned"
+                ):
+                    raise BleRuntimeProcessLeaseBorrowInvalid()
+                borrowed_process_lease = process_lease_borrow._lease
+            self._next_token += 1
+            token = self._next_token
+            self._active = _ActiveLease(
+                token=token,
+                operation_kind=operation_kind,
+                owner_epoch=owner_epoch,
+                owned_process_lease=owned_process_lease,
+                borrowed_process_lease=borrowed_process_lease,
+            )
+            return token
+
+    def attach_task(self, token: int, task: asyncio.Task[Any]) -> None:
+        with self._lock:
+            if self._active is None or self._active.token != token:
+                raise RuntimeError("BLE runtime lease was invalidated before task attachment")
+            self._active.task = task
+
+    def mark_cleanup_pending(self, token: int) -> None:
+        with self._lock:
+            if self._active is not None and self._active.token == token:
+                self._active.cleanup_pending = True
+
+    def release(self, token: int) -> None:
+        """Finish one lease and drain callbacks before admitting another lease.
+
+        The completed task remains represented by ``_active`` while callbacks
+        run.  Consequently a concurrent local BLE acquisition still fails
+        closed until the external lifecycle barriers retained by those
+        callbacks have actually been released.  Callbacks must be bounded and
+        must not call back into this arbiter.
+        """
+
+        while True:
+            with self._lock:
+                active = self._active
+                if active is None or active.token != token or self._poisoned:
+                    return
+                callbacks = tuple(self._idle_callbacks)
+                self._idle_callbacks.clear()
+                if not callbacks:
+                    process_lease = active.owned_process_lease
+                    if process_lease is not None:
+                        try:
+                            # This is a non-blocking unlock/close. Keep the
+                            # arbiter lock held so neither a local acquisition
+                            # nor a newly registered external release can race
+                            # the final OS ownership transition.
+                            process_lease.release()
+                        except Exception:
+                            active.cleanup_pending = True
+                            self._poisoned = True
+                            return
+                    self._active = None
+                    return
+
+            for index, callback in enumerate(callbacks):
+                try:
+                    callback()
+                except Exception:
+                    # An external release that cannot be proven complete is a
+                    # process-wide safety failure.  Keep this lease and every
+                    # unexecuted callback retained until process restart; never
+                    # retry a callback whose partial effects are unknowable.
+                    with self._lock:
+                        current = self._active
+                        if current is not None and current.token == token:
+                            current.cleanup_pending = True
+                            self._poisoned = True
+                            self._idle_callbacks[0:0] = callbacks[index:]
+                    return
+
+    def poison_cleanup_failure(self, token: int) -> None:
+        """Retain ownership when native session teardown is not proven clean."""
+
+        with self._lock:
+            if self._active is not None and self._active.token == token:
+                self._active.cleanup_pending = True
+                self._poisoned = True
+
+    def defer_until_idle(
+        self,
+        callback: Callable[[], None],
+    ) -> BleRuntimeIdleCallbackDisposition:
+        """Run a bounded callback now or at the next proven-idle transition.
+
+        Registration and the idle decision are serialized with acquisition and
+        release.  When a BLE operation is active, its lease stays active while
+        the callback runs, so another local operation cannot enter the native
+        runtime between teardown and release of an external process lease.
+        Poisoned runtimes retain callbacks without ever invoking them: only a
+        process restart may safely release the corresponding OS ownership.
+
+        The callback must be quick, non-blocking, and must not call this arbiter.
+        """
+
+        with self._lock:
+            if self._poisoned:
+                self._idle_callbacks.append(callback)
+                return "poisoned"
+            if self._active is not None:
+                self._idle_callbacks.append(callback)
+                return "deferred"
+            try:
+                callback()
+            except Exception:
+                # There is no caller-independent way to know whether an
+                # external release partially succeeded.  Preserve the callback
+                # and prohibit a later BLE admission until process restart.
+                self._poisoned = True
+                self._idle_callbacks.append(callback)
+                raise
+            return "released"
+
+    def snapshot(self) -> BleRuntimeSnapshot:
+        with self._lock:
+            active = self._active
+            return {
+                "owner_epoch": self._owner_epoch,
+                "owner_loop_bound": (
+                    self._owner_loop is not None and not self._owner_loop.is_closed()
+                ),
+                "active_operation_kind": (active.operation_kind if active is not None else None),
+                "cleanup_pending": bool(active is not None and active.cleanup_pending),
+                "poisoned": self._poisoned,
+            }
+
+    def reset_for_tests(self) -> None:
+        """Forget singleton state without letting an old callback release a new lease."""
+
+        with self._lock:
+            owned_process_lease = (
+                self._active.owned_process_lease if self._active is not None else None
+            )
+            self._owner_epoch += 1
+            self._next_token += 1
+            self._owner_loop = None
+            self._active = None
+            self._poisoned = False
+            self._idle_callbacks.clear()
+            self._process_lease_configuration_epoch += 1
+            self._process_lease_repository_root = None
+            self._process_lease_lock_path = None
+        if owned_process_lease is not None:
+            # Test reset deliberately models process exit for synthetic poison.
+            owned_process_lease.release()
+
+
+_BLE_PROCESS_ARBITER = _ProcessBleRuntimeArbiter()
+_BLE_PROCESS_LEASE_BORROW_CONTEXT: ContextVar[
+    BleRuntimeProcessLeaseBorrowToken | None
+] = ContextVar("ble_runtime_process_lease_borrow", default=None)
+
+
+def _process_lease_path(repository_root: Path) -> Path:
+    return (
+        resolve_missioncore_data_dir(repository_root)
+        / "xgrids-k1"
+        / APPLICATION_CONTROL_LOCK_FILENAME
+    )
+
+
+def _annotate_cancellation(
+    exc: asyncio.CancelledError,
+    progress: BleOperationProgress,
+) -> None:
+    """Preserve side-effect facts when caller cancellation crosses BLE I/O."""
+
+    exc.operation_stage = progress.operation_stage  # type: ignore[attr-defined]
+    exc.device_write_attempted = progress.device_write_attempted  # type: ignore[attr-defined]
+    exc.device_write_confirmed = progress.device_write_confirmed  # type: ignore[attr-defined]
+
+
+def bind_ble_runtime_owner_loop(
+    loop: asyncio.AbstractEventLoop | None = None,
+) -> int:
+    """Bind process CoreBluetooth ownership to the current persistent loop."""
+
+    return _BLE_PROCESS_ARBITER.bind_owner_loop(loop or asyncio.get_running_loop())
+
+
+def ble_runtime_owner_epoch_for_current_loop() -> int | None:
+    """Return the owner epoch only when called on the bound live loop."""
+
+    try:
+        loop = asyncio.get_running_loop()
+    except RuntimeError:
+        return None
+    return _BLE_PROCESS_ARBITER.owner_epoch_for_loop(loop)
+
+
+def invalidate_ble_runtime_owner_loop(
+    loop: asyncio.AbstractEventLoop | None = None,
+) -> int:
+    """Invalidate process loop ownership without releasing an active BLE lease."""
+
+    return _BLE_PROCESS_ARBITER.invalidate_owner_loop(loop)
+
+
+def ble_runtime_snapshot() -> BleRuntimeSnapshot:
+    """Expose a non-secret process snapshot for diagnostics and focused tests."""
+
+    return _BLE_PROCESS_ARBITER.snapshot()
+
+
+def reset_ble_runtime_arbiter_for_tests() -> None:
+    """Reset process-global state between synthetic tests only."""
+
+    _BLE_PROCESS_ARBITER.reset_for_tests()
+    _BLE_PROCESS_LEASE_BORROW_CONTEXT.set(None)
+
+
+def configure_ble_runtime_process_lease(repository_root: Path) -> int:
+    """Configure the canonical OS lifecycle lock used by all BLE entrypoints."""
+
+    return _BLE_PROCESS_ARBITER.configure_process_lease(repository_root)
+
+
+@contextmanager
+def borrow_ble_runtime_process_lease(
+    lease: ApplicationControlProcessLease,
+) -> Iterator[BleRuntimeProcessLeaseBorrowToken]:
+    """Borrow a higher-level lifecycle lease without a second flock attempt.
+
+    The caller must keep ``lease`` owned until every BLE operation admitted in
+    this context has reached proven idle.  A copied context cannot admit a new
+    operation after this manager exits because the opaque token is invalidated.
+    """
+
+    borrow = _BLE_PROCESS_ARBITER.create_process_lease_borrow(lease)
+    context_token = _BLE_PROCESS_LEASE_BORROW_CONTEXT.set(borrow)
+    try:
+        yield borrow
+    finally:
+        _BLE_PROCESS_ARBITER.deactivate_process_lease_borrow(borrow)
+        _BLE_PROCESS_LEASE_BORROW_CONTEXT.reset(context_token)
+
+
+def defer_until_ble_runtime_idle(
+    callback: Callable[[], None],
+) -> BleRuntimeIdleCallbackDisposition:
+    """Release an external barrier only after native BLE ownership is idle.
+
+    ``"released"`` means the callback ran synchronously because no BLE lease
+    existed.  ``"deferred"`` means it is owned by the active lease and will run
+    exactly once after native task completion.  ``"poisoned"`` means it is
+    intentionally retained without execution until process restart.
+    """
+
+    return _BLE_PROCESS_ARBITER.defer_until_idle(callback)
+
+
+async def wait_for_ble_runtime_idle(timeout_seconds: float = 1.0) -> bool:
+    """Wait without blocking the owner loop until the active lease is released."""
+
+    if timeout_seconds < 0:
+        raise ValueError("timeout_seconds must be non-negative")
+    loop = asyncio.get_running_loop()
+    deadline = loop.time() + timeout_seconds
+    while ble_runtime_snapshot()["active_operation_kind"] is not None:
+        if loop.time() >= deadline:
+            return False
+        await asyncio.sleep(min(0.01, max(0.0, deadline - loop.time())))
+    return True
+
+
+async def run_ble_operation[T](
+    operation_kind: BleOperationKind,
+    *,
+    hard_timeout_seconds: float,
+    operation: Callable[[BleOperationProgress], Coroutine[Any, Any, T]],
+    progress: BleOperationProgress | None = None,
+) -> T:
+    """Run one BLE task behind a process lease and a non-blocking hard deadline."""
+
+    if hard_timeout_seconds <= 0:
+        raise ValueError("hard_timeout_seconds must be positive")
+    loop = asyncio.get_running_loop()
+    owner_epoch = bind_ble_runtime_owner_loop(loop)
+    operation_progress = progress or BleOperationProgress()
+    operation_progress.owner_epoch = owner_epoch
+    token = _BLE_PROCESS_ARBITER.acquire(
+        operation_kind,
+        owner_epoch=owner_epoch,
+        process_lease_borrow=_BLE_PROCESS_LEASE_BORROW_CONTEXT.get(),
+    )
+    try:
+        task: asyncio.Task[T] = loop.create_task(operation(operation_progress))
+    except BaseException:
+        _BLE_PROCESS_ARBITER.release(token)
+        raise
+    _BLE_PROCESS_ARBITER.attach_task(token, task)
+    cleanup_requested = False
+
+    def release_after_completion(completed: asyncio.Future[T]) -> None:
+        cleanup_failed = False
+        try:
+            completed_exception = None if completed.cancelled() else completed.exception()
+            cleanup_failed = cleanup_requested and completed_exception is not None
+        except BaseException:
+            # The caller observes the original result/exception. This callback
+            # only consumes detached cleanup outcomes and releases ownership.
+            cleanup_failed = cleanup_requested
+        finally:
+            if cleanup_failed:
+                _BLE_PROCESS_ARBITER.poison_cleanup_failure(token)
+            else:
+                _BLE_PROCESS_ARBITER.release(token)
+
+    task.add_done_callback(release_after_completion)
+    try:
+        completed, _ = await asyncio.wait({task}, timeout=hard_timeout_seconds)
+    except asyncio.CancelledError as exc:
+        _annotate_cancellation(exc, operation_progress)
+        if not task.done():
+            _BLE_PROCESS_ARBITER.mark_cleanup_pending(token)
+            cleanup_requested = True
+            task.cancel()
+        raise
+    if completed:
+        return task.result()
+
+    _BLE_PROCESS_ARBITER.mark_cleanup_pending(token)
+    cleanup_requested = True
+    task.cancel()
+    raise BleOperationHardTimeout(operation_kind, operation_progress)
+
+
+@asynccontextmanager
+async def run_ble_operation_session[T](
+    operation_kind: BleOperationKind,
+    *,
+    hard_setup_timeout_seconds: float,
+    hard_cleanup_timeout_seconds: float = 5.0,
+    operation: Callable[
+        [BleOperationProgress],
+        AbstractAsyncContextManager[T],
+    ],
+    progress: BleOperationProgress | None = None,
+) -> AsyncIterator[T]:
+    """Hold one process BLE lease across setup, caller work, and cleanup.
+
+    Only setup is subject to the hard deadline. Once the inner session yields,
+    its native BLE client remains alive until the caller leaves this context.
+    Cancellation never frees the process lease early: a detached native cleanup
+    continues to own the lease until its task has actually completed.
+    """
+
+    if hard_setup_timeout_seconds <= 0:
+        raise ValueError("hard_setup_timeout_seconds must be positive")
+    if hard_cleanup_timeout_seconds <= 0:
+        raise ValueError("hard_cleanup_timeout_seconds must be positive")
+    loop = asyncio.get_running_loop()
+    owner_epoch = bind_ble_runtime_owner_loop(loop)
+    operation_progress = progress or BleOperationProgress()
+    operation_progress.owner_epoch = owner_epoch
+    token = _BLE_PROCESS_ARBITER.acquire(
+        operation_kind,
+        owner_epoch=owner_epoch,
+        process_lease_borrow=_BLE_PROCESS_LEASE_BORROW_CONTEXT.get(),
+    )
+    ready: asyncio.Future[T] = loop.create_future()
+    release_requested = asyncio.Event()
+    setup_cleanup_requested = False
+
+    async def session_task() -> None:
+        async with operation(operation_progress) as value:
+            if not ready.done():
+                ready.set_result(value)
+            await release_requested.wait()
+
+    try:
+        task = loop.create_task(session_task())
+    except BaseException:
+        _BLE_PROCESS_ARBITER.release(token)
+        raise
+    _BLE_PROCESS_ARBITER.attach_task(token, task)
+
+    def release_after_completion(completed: asyncio.Future[None]) -> None:
+        cleanup_failed = False
+        try:
+            completed_exception = None if completed.cancelled() else completed.exception()
+            cleanup_failed = bool(
+                (
+                    ready.done()
+                    and not ready.cancelled()
+                    and release_requested.is_set()
+                    and (completed.cancelled() or completed_exception is not None)
+                )
+                or (setup_cleanup_requested and completed_exception is not None)
+            )
+        except BaseException:
+            # The active caller observes setup/cleanup failures directly. This
+            # callback only consumes detached outcomes before releasing ownership.
+            cleanup_failed = setup_cleanup_requested or bool(
+                ready.done() and not ready.cancelled() and release_requested.is_set()
+            )
+        finally:
+            if cleanup_failed:
+                _BLE_PROCESS_ARBITER.poison_cleanup_failure(token)
+            else:
+                _BLE_PROCESS_ARBITER.release(token)
+
+    task.add_done_callback(release_after_completion)
+    setup_waiters: set[asyncio.Future[Any]] = {ready, task}
+    try:
+        completed, _ = await asyncio.wait(
+            setup_waiters,
+            timeout=hard_setup_timeout_seconds,
+            return_when=asyncio.FIRST_COMPLETED,
+        )
+    except asyncio.CancelledError as exc:
+        _annotate_cancellation(exc, operation_progress)
+        if not task.done():
+            _BLE_PROCESS_ARBITER.mark_cleanup_pending(token)
+            setup_cleanup_requested = True
+            task.cancel()
+        raise
+
+    if ready not in completed:
+        if task in completed:
+            # Setup failed before the inner session became available.
+            task.result()
+            raise RuntimeError("BLE session ended before setup completed")
+        _BLE_PROCESS_ARBITER.mark_cleanup_pending(token)
+        setup_cleanup_requested = True
+        task.cancel()
+        raise BleOperationHardTimeout(operation_kind, operation_progress)
+
+    try:
+        try:
+            yield ready.result()
+        except asyncio.CancelledError as exc:
+            _annotate_cancellation(exc, operation_progress)
+            raise
+    finally:
+        # From this point the setup deadline no longer applies. The caller may
+        # perform a bounded host-side handoff while the same BLE client remains
+        # connected. On exit, ownership is retained until __aexit__ really ends.
+        # A wedged native disconnect must not hold the HTTP caller forever:
+        # detach it after the cleanup deadline while the task and process lease
+        # remain quarantined until CoreBluetooth actually acknowledges cleanup.
+        _BLE_PROCESS_ARBITER.mark_cleanup_pending(token)
+        release_requested.set()
+        completed, _ = await asyncio.wait(
+            {task},
+            timeout=hard_cleanup_timeout_seconds,
+        )
+        if completed:
+            task.result()
+        # Do not cancel a disconnect already in progress. A second cancellation
+        # can make an otherwise responsive ``__aexit__`` finish as cancelled
+        # without proving that CoreBluetooth acknowledged the native teardown.
+        # The detached task therefore keeps the process lease until its natural
+        # completion callback releases it.

src/k1link/device_plugins/xgrids_k1/ble/wifi_provisioning.py

@@ -2,6 +2,8 @@ from __future__ import annotations
 
 import asyncio
 import ipaddress
+import math
+from collections.abc import Callable
 from importlib.metadata import version
 from time import monotonic
 from typing import Literal, TypedDict
@@ -10,8 +12,20 @@ from bleak import BleakClient, BleakScanner
 from bleak.exc import BleakDeviceNotFoundError, BleakError, BleakGATTProtocolError
 
 from k1link.artifacts import utc_now_iso
+from k1link.device_plugins.xgrids_k1.ble.runtime_arbiter import (
+    BleOperationProgress,
+    run_ble_operation,
+)
 from k1link.device_plugins.xgrids_k1.ble.scanner import (
+    CapturedDiscoveredDevice,
+    captured_device_handle,
+    connected_device_capture,
+    demote_connected_device_handle_after_gatt_failure,
+    discover_known_device_capture_for_status_read,
     discovered_device_selection,
+    mark_captured_device_gatt_validated,
+    retrieve_connected_device_capture,
+    retrieve_known_device_capture_for_status_read,
 )
 
 PROFILE_ID = "xgrids-k1-fw3-wifi-v1"
@@ -22,6 +36,8 @@ FRAME_LENGTH = 99
 SSID_SLOT_LENGTH = 32
 PASSWORD_SLOT_LENGTH = 64
 AP_FALLBACK_IPV4 = "192.168.56.1"
+BLE_STATUS_HARD_TIMEOUT_GRACE_SECONDS = 5.0
+BLE_PROVISION_HARD_TIMEOUT_GRACE_SECONDS = 25.0
 ProvisioningOutcome = Literal[
     "lan_address_observed",
     "status_changed",
@@ -38,11 +54,28 @@ BleOperationStage = Literal[
     "gatt-write",
     "status-poll",
 ]
+StatusReadOperationStage = Literal[
+    "resolution",
+    "exact-uuid-scan",
+    "connect",
+    "gatt-contract",
+    "status-read",
+]
+_STATUS_READ_OPERATION_STAGES: frozenset[str] = frozenset(
+    {
+        "resolution",
+        "exact-uuid-scan",
+        "connect",
+        "gatt-contract",
+        "status-read",
+    }
+)
 
 
 class WifiStatus(TypedDict):
     value_length: int
     mode: str | None
+    network_name: str | None
     ipv4: str | None
     status_code: int
     reserved: int | None
@@ -70,7 +103,7 @@ class WifiProvisioningResult(TypedDict):
     operation: str
     write_mode: ResolvedWriteMode
     write_without_response_advertised: bool
-    max_write_without_response_size: int
+    max_write_without_response_size: int | None
     frame_length: int
     baseline_status: WifiStatus
     observations: list[StatusObservation]
@@ -86,6 +119,10 @@ class WifiStatusReadResult(TypedDict):
     device_macos_uuid: str
     device_name: str
     service_uuid: str
+    write_characteristic_uuid: str
+    write_characteristic_properties: list[str]
+    max_write_without_response_size: int | None
+    mtu_size: int | None
     status_characteristic_uuid: str
     operation: Literal["single_reviewed_wifi_status_read"]
     write_performed: Literal[False]
@@ -98,17 +135,61 @@ def _annotate_ble_operation_error(
     operation_stage: BleOperationStage,
     device_write_attempted: bool,
     device_write_confirmed: bool,
+    resolved_write_mode: ResolvedWriteMode | None,
+    write_characteristic_properties: tuple[str, ...] | None,
+    max_write_without_response_size: int | None,
+    frame_length: int,
 ) -> None:
     """Attach non-secret transport facts while preserving the exception type."""
 
     exc.operation_stage = operation_stage  # type: ignore[attr-defined]
     exc.device_write_attempted = device_write_attempted  # type: ignore[attr-defined]
     exc.device_write_confirmed = device_write_confirmed  # type: ignore[attr-defined]
+    exc.resolved_write_mode = resolved_write_mode  # type: ignore[attr-defined]
+    exc.write_characteristic_properties = (  # type: ignore[attr-defined]
+        write_characteristic_properties
+    )
+    exc.max_write_without_response_size = (  # type: ignore[attr-defined]
+        max_write_without_response_size
+    )
+    exc.frame_length = frame_length  # type: ignore[attr-defined]
     if isinstance(exc, BleakGATTProtocolError):
         exc.att_error_code = int(exc.code)  # type: ignore[attr-defined]
         exc.att_error_name = exc.code.name  # type: ignore[attr-defined]
 
 
+def _annotate_status_read_error(exc: Exception, operation_stage: str) -> None:
+    """Attach only one sanitized read-only stage to a transport failure."""
+
+    try:
+        existing_stage = getattr(exc, "operation_stage", None)
+    except Exception:
+        existing_stage = None
+    if existing_stage in _STATUS_READ_OPERATION_STAGES:
+        return
+    sanitized_stage = (
+        operation_stage
+        if operation_stage in _STATUS_READ_OPERATION_STAGES
+        else "resolution"
+    )
+    try:
+        exc.operation_stage = sanitized_stage  # type: ignore[attr-defined]
+    except Exception:
+        # A third-party exception may forbid dynamic attributes. Preserve its
+        # original type and traceback rather than replacing the BLE failure.
+        return
+
+
+def _optional_int_attribute(source: object, name: str) -> int | None:
+    """Read optional backend metadata without making diagnostics operationally required."""
+
+    try:
+        value = getattr(source, name, None)
+    except Exception:
+        return None
+    return value if isinstance(value, int) and not isinstance(value, bool) else None
+
+
 def build_wifi_provisioning_frame(ssid: str, password: str) -> bytearray:
     """Build the deterministic 99-byte frame used by LixelGO for K1 Wi-Fi setup."""
     ssid_bytes = ssid.encode("utf-8")
@@ -133,17 +214,40 @@ def build_wifi_provisioning_frame(ssid: str, password: str) -> bytearray:
 
 
 def parse_wifi_status(value: bytes) -> WifiStatus:
-    """Parse the non-secret status frame returned by the K1 read characteristic."""
+    """Parse the status frame returned by the K1 read characteristic.
+
+    The first 32-byte text slot is not a mode enum in station mode.  Physical
+    K1 FW 3.0.2 evidence shows that it contains the joined Wi-Fi network name
+    (for example a lab SSID), while AP mode uses the control literal
+    ``WIFI_AP``.  Keep ``mode`` as the normalized semantic family so callers do
+    not have to mistake an operator network name for a protocol enum, and
+    expose ``network_name`` only for exact, in-process target comparison.
+    """
     if len(value) < 51:
         raise ValueError("K1 Wi-Fi status must contain at least 51 bytes")
 
-    mode_length = value[0]
-    if mode_length > SSID_SLOT_LENGTH:
-        raise ValueError("K1 Wi-Fi status mode length is invalid")
+    text_length = value[0]
+    if text_length > SSID_SLOT_LENGTH:
+        raise ValueError("K1 Wi-Fi status text length is invalid")
     try:
-        mode = value[1 : 1 + mode_length].decode("utf-8") if mode_length else None
+        status_text = (
+            value[1 : 1 + text_length].decode("utf-8") if text_length else None
+        )
     except UnicodeDecodeError as exc:
-        raise ValueError("K1 Wi-Fi status mode is not valid UTF-8") from exc
+        raise ValueError("K1 Wi-Fi status text is not valid UTF-8") from exc
+
+    if status_text == "WIFI_AP":
+        mode = "WIFI_AP"
+        network_name = None
+    elif status_text:
+        # Older synthetic fixtures and possible legacy firmware may still
+        # report the literal WIFI_CLIENT.  It identifies the station family
+        # but supplies no exact network discriminator.
+        mode = "WIFI_CLIENT"
+        network_name = None if status_text == "WIFI_CLIENT" else status_text
+    else:
+        mode = None
+        network_name = None
 
     address_length = value[33]
     address_start = 34
@@ -163,6 +267,7 @@ def parse_wifi_status(value: bytes) -> WifiStatus:
     return {
         "value_length": len(value),
         "mode": mode,
+        "network_name": network_name,
         "ipv4": ipv4,
         "status_code": value[50],
         "reserved": value[51] if len(value) > 51 else None,
@@ -186,25 +291,95 @@ def _outcome(
     return "no_status_change_before_timeout"
 
 
-async def read_wifi_status_once(
+async def _read_wifi_status_impl(
     device_macos_uuid: str,
     *,
-    timeout_seconds: float = 20.0,
-    rediscover: bool = False,
+    timeout_seconds: float,
+    exact_scan_timeout_seconds: float,
+    rediscover: bool,
+    captured_device: CapturedDiscoveredDevice | None,
+    recovery_device_session_id: str | None,
+    allow_known_device_retrieval: bool,
+    on_gatt_validated: Callable[[CapturedDiscoveredDevice], None] | None,
+    progress: BleOperationProgress,
 ) -> WifiStatusReadResult:
-    """Read the K1's current DHCP status over BLE without writing a characteristic."""
-
-    if timeout_seconds <= 0:
-        raise ValueError("timeout_seconds must be positive")
-    async with asyncio.timeout(timeout_seconds + 5.0):
-        # A still-live explicit scan lease is authoritative even for a caller
-        # requesting recovery. Physical acceptance proved that immediately
-        # looking the same CoreBluetooth UUID up again can lose a present K1.
-        # ``rediscover`` therefore permits fallback only after that short lease
-        # has expired; it never discards a fresh retained BLEDevice.
-        selection = discovered_device_selection(device_macos_uuid)
-        device = selection.device
-        if device is None and not selection.from_fresh_scan:
+    active_captured_device: CapturedDiscoveredDevice | None = None
+    gatt_baseline_validated = False
+    try:
+        progress.operation_stage = "resolution"
+        if captured_device is not None:
+            # Explicit recovery is fail-closed: only the exact retrieved
+            # CoreBluetooth object may be used. Never replace it with a scan,
+            # UUID lookup, or automatic retry behind the operator's back.
+            device = (
+                captured_device_handle(captured_device)
+                if captured_device.macos_uuid.casefold()
+                == device_macos_uuid.casefold()
+                else None
+            )
+            if device is not None:
+                active_captured_device = captured_device
+            selection_from_fresh_scan = False
+        elif recovery_device_session_id is not None:
+            # Same-process recovery is resolved inside the status-read BLE
+            # lease.  Retrieval, connect, reviewed GATT contract and 7f02 are
+            # therefore one serialized operation; no scan row or UUID lookup
+            # can race between those steps.
+            active_captured_device = connected_device_capture(
+                device_macos_uuid,
+                device_session_id=recovery_device_session_id,
+            )
+            if active_captured_device is None:
+                active_captured_device = await retrieve_connected_device_capture(
+                    device_macos_uuid,
+                    device_session_id=recovery_device_session_id,
+                )
+            device = (
+                captured_device_handle(active_captured_device)
+                if active_captured_device is not None
+                else None
+            )
+            selection_from_fresh_scan = False
+        elif allow_known_device_retrieval:
+            # Durable physical recovery deliberately bypasses public scan
+            # generations. ``rediscover`` means one primary, unfiltered
+            # advertisement wait for this exact CoreBluetooth UUID; it never
+            # attempts the potentially stale cached peripheral first and never
+            # falls back to it after a timeout or failed connect.
+            if rediscover:
+                progress.operation_stage = "exact-uuid-scan"
+                active_captured_device = (
+                    await discover_known_device_capture_for_status_read(
+                        device_macos_uuid,
+                        timeout_seconds=exact_scan_timeout_seconds,
+                    )
+                )
+            else:
+                active_captured_device = (
+                    await retrieve_known_device_capture_for_status_read(
+                        device_macos_uuid,
+                    )
+                )
+            device = (
+                captured_device_handle(active_captured_device)
+                if active_captured_device is not None
+                else None
+            )
+            selection_from_fresh_scan = False
+        else:
+            # A still-live explicit scan lease is authoritative. Physical
+            # acceptance proved that immediately looking the same UUID up
+            # again can lose a present K1.
+            selection = discovered_device_selection(device_macos_uuid)
+            device = selection.device
+            selection_from_fresh_scan = selection.from_fresh_scan
+        if (
+            device is None
+            and captured_device is None
+            and recovery_device_session_id is None
+            and not selection_from_fresh_scan
+            and not allow_known_device_retrieval
+        ):
             device = await BleakScanner.find_device_by_address(
                 device_macos_uuid,
                 timeout=timeout_seconds,
@@ -212,23 +387,61 @@ async def read_wifi_status_once(
         if device is None:
             raise BleakDeviceNotFoundError(
                 device_macos_uuid,
-                "Device was not rediscovered; keep the K1 powered and nearby.",
+                "Exact BLE device is unavailable; run an explicit recovery or scan.",
             )
 
+        progress.operation_stage = "connect"
         async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
+            progress.operation_stage = "gatt-contract"
             service = client.services.get_service(SERVICE_UUID)
-            status_characteristic = client.services.get_characteristic(STATUS_CHARACTERISTIC_UUID)
+            write_characteristic = client.services.get_characteristic(
+                WRITE_CHARACTERISTIC_UUID
+            )
+            status_characteristic = client.services.get_characteristic(
+                STATUS_CHARACTERISTIC_UUID
+            )
             if service is None:
                 raise ValueError(f"Reviewed K1 service not found: {SERVICE_UUID}")
+            if write_characteristic is None:
+                raise ValueError(
+                    "Reviewed K1 write characteristic not found: "
+                    f"{WRITE_CHARACTERISTIC_UUID}"
+                )
             if status_characteristic is None:
                 raise ValueError(
-                    f"Reviewed K1 status characteristic not found: {STATUS_CHARACTERISTIC_UUID}"
+                    "Reviewed K1 status characteristic not found: "
+                    f"{STATUS_CHARACTERISTIC_UUID}"
+                )
+            if write_characteristic.service_uuid != service.uuid:
+                raise ValueError(
+                    "K1 write characteristic is attached to an unexpected service"
                 )
             if status_characteristic.service_uuid != service.uuid:
-                raise ValueError("K1 status characteristic is attached to an unexpected service")
+                raise ValueError(
+                    "K1 status characteristic is attached to an unexpected service"
+                )
             if "read" not in status_characteristic.properties:
                 raise ValueError("Reviewed K1 status characteristic is not readable")
+            write_properties = sorted(
+                {str(item) for item in write_characteristic.properties}
+            )
+            max_without_response = _optional_int_attribute(
+                write_characteristic,
+                "max_write_without_response_size",
+            )
+            mtu_size = _optional_int_attribute(client, "mtu_size")
+            progress.operation_stage = "status-read"
             value = bytes(await client.read_gatt_char(status_characteristic))
+            status = parse_wifi_status(value)
+            if active_captured_device is not None and not (
+                mark_captured_device_gatt_validated(active_captured_device)
+            ):
+                raise RuntimeError(
+                    "Exact BLE recovery handle changed before status validation"
+                )
+            gatt_baseline_validated = True
+            if active_captured_device is not None and on_gatt_validated is not None:
+                on_gatt_validated(active_captured_device)
             return {
                 "schema_version": 1,
                 "profile_id": PROFILE_ID,
@@ -238,11 +451,325 @@ async def read_wifi_status_once(
                 "device_macos_uuid": device_macos_uuid,
                 "device_name": client.name or "",
                 "service_uuid": service.uuid,
+                "write_characteristic_uuid": write_characteristic.uuid,
+                "write_characteristic_properties": write_properties,
+                "max_write_without_response_size": max_without_response,
+                "mtu_size": mtu_size,
                 "status_characteristic_uuid": status_characteristic.uuid,
                 "operation": "single_reviewed_wifi_status_read",
                 "write_performed": False,
-                "status": parse_wifi_status(value),
+                "status": status,
             }
+    except Exception as exc:
+        _annotate_status_read_error(exc, progress.operation_stage)
+        raise
+    finally:
+        # Connection/contract/read failure revokes only this transport lease;
+        # the process-scoped UUID/session token remains available for another
+        # explicit CoreBluetooth retrieval attempt.
+        if active_captured_device is not None and not gatt_baseline_validated:
+            demote_connected_device_handle_after_gatt_failure(active_captured_device)
+
+
+async def read_wifi_status_once(
+    device_macos_uuid: str,
+    *,
+    timeout_seconds: float = 20.0,
+    exact_scan_timeout_seconds: float = 30.0,
+    rediscover: bool = False,
+    captured_device: CapturedDiscoveredDevice | None = None,
+    recovery_device_session_id: str | None = None,
+    allow_known_device_retrieval: bool = False,
+    on_gatt_validated: Callable[[CapturedDiscoveredDevice], None] | None = None,
+) -> WifiStatusReadResult:
+    """Read the K1's current DHCP status over BLE without writing a characteristic."""
+
+    if timeout_seconds <= 0:
+        raise ValueError("timeout_seconds must be positive")
+    if captured_device is not None and recovery_device_session_id is not None:
+        raise ValueError(
+            "captured_device and recovery_device_session_id are mutually exclusive"
+        )
+    if recovery_device_session_id == "":
+        raise ValueError("recovery_device_session_id must not be empty")
+    if recovery_device_session_id is not None and allow_known_device_retrieval:
+        raise ValueError(
+            "same-process recovery and durable known-device retrieval are mutually exclusive"
+        )
+    use_exact_uuid_scan = bool(
+        rediscover
+        and allow_known_device_retrieval
+        and captured_device is None
+        and recovery_device_session_id is None
+    )
+    if use_exact_uuid_scan and (
+        not math.isfinite(exact_scan_timeout_seconds)
+        or exact_scan_timeout_seconds <= 0
+    ):
+        raise ValueError("exact_scan_timeout_seconds must be positive and finite")
+    exact_scan_budget = exact_scan_timeout_seconds if use_exact_uuid_scan else 0.0
+    return await run_ble_operation(
+        "status-read",
+        hard_timeout_seconds=(
+            exact_scan_budget
+            + timeout_seconds
+            + BLE_STATUS_HARD_TIMEOUT_GRACE_SECONDS
+        ),
+        operation=lambda progress: _read_wifi_status_impl(
+            device_macos_uuid,
+            timeout_seconds=timeout_seconds,
+            exact_scan_timeout_seconds=exact_scan_timeout_seconds,
+            rediscover=rediscover,
+            captured_device=captured_device,
+            recovery_device_session_id=recovery_device_session_id,
+            allow_known_device_retrieval=allow_known_device_retrieval,
+            on_gatt_validated=on_gatt_validated,
+            progress=progress,
+        ),
+        progress=BleOperationProgress(operation_stage="resolution"),
+    )
+
+
+async def _provision_wifi_impl(
+    device_macos_uuid: str,
+    ssid: str,
+    password: str,
+    *,
+    timeout_seconds: float,
+    poll_interval_seconds: float,
+    write_mode: WriteMode,
+    captured_device: CapturedDiscoveredDevice | None,
+    recovery_device_session_id: str | None,
+    on_write_dispatch: Callable[[WifiStatus, ResolvedWriteMode], None] | None,
+    progress: BleOperationProgress,
+) -> WifiProvisioningResult:
+    frame = build_wifi_provisioning_frame(ssid, password)
+    started_at = utc_now_iso()
+    observations: list[StatusObservation] = []
+    disconnected = False
+    operation_stage: BleOperationStage = "resolution"
+    device_write_attempted = False
+    device_write_confirmed = False
+    resolved_write_mode_for_error: ResolvedWriteMode | None = None
+    write_characteristic_properties: tuple[str, ...] | None = None
+    max_without_response: int | None = None
+    active_captured_device: CapturedDiscoveredDevice | None = None
+    gatt_baseline_validated = False
+
+    try:
+        progress.operation_stage = operation_stage
+        # The explicit UI scan and its selected network action are one
+        # CoreBluetooth lifecycle. A supplied capture is fail-closed: never
+        # replace an expired/mismatched object with a scan or UUID lookup.
+        if captured_device is not None:
+            device = (
+                captured_device_handle(captured_device)
+                if captured_device.macos_uuid.casefold()
+                == device_macos_uuid.casefold()
+                else None
+            )
+            if device is not None:
+                active_captured_device = captured_device
+            selection_from_fresh_scan = False
+        elif recovery_device_session_id is not None:
+            # Resolve an exact retained UUID/session token only after the
+            # wifi-provision arbiter lease has been admitted. This prevents a
+            # concurrent scan/status operation and removes the preflight-to-
+            # write lease gap.
+            active_captured_device = connected_device_capture(
+                device_macos_uuid,
+                device_session_id=recovery_device_session_id,
+            )
+            if active_captured_device is None:
+                active_captured_device = await retrieve_connected_device_capture(
+                    device_macos_uuid,
+                    device_session_id=recovery_device_session_id,
+                )
+            device = (
+                captured_device_handle(active_captured_device)
+                if active_captured_device is not None
+                else None
+            )
+            selection_from_fresh_scan = False
+        else:
+            selection = discovered_device_selection(device_macos_uuid)
+            device = selection.device
+            selection_from_fresh_scan = selection.from_fresh_scan
+        if (
+            device is None
+            and captured_device is None
+            and recovery_device_session_id is None
+            and not selection_from_fresh_scan
+        ):
+            # Non-UI callers without a current explicit scan retain the
+            # bounded lookup fallback. A fresh scan missing this device is
+            # authoritative and must not be silently replaced here.
+            device = await BleakScanner.find_device_by_address(
+                device_macos_uuid,
+                timeout=min(20.0, timeout_seconds),
+            )
+        if device is None:
+            raise BleakDeviceNotFoundError(
+                device_macos_uuid,
+                "Device was not rediscovered; keep the K1 powered and nearby.",
+            )
+
+        operation_stage = "connect"
+        progress.operation_stage = operation_stage
+        async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
+            device_name = client.name
+            operation_stage = "gatt-contract"
+            progress.operation_stage = operation_stage
+            service = client.services.get_service(SERVICE_UUID)
+            write_characteristic = client.services.get_characteristic(WRITE_CHARACTERISTIC_UUID)
+            status_characteristic = client.services.get_characteristic(STATUS_CHARACTERISTIC_UUID)
+            if service is None:
+                raise ValueError(f"Reviewed K1 service not found: {SERVICE_UUID}")
+            if write_characteristic is None:
+                raise ValueError(
+                    f"Reviewed K1 write characteristic not found: {WRITE_CHARACTERISTIC_UUID}"
+                )
+            if status_characteristic is None:
+                raise ValueError(
+                    f"Reviewed K1 status characteristic not found: {STATUS_CHARACTERISTIC_UUID}"
+                )
+            if write_characteristic.service_uuid != service.uuid:
+                raise ValueError("K1 write characteristic is attached to an unexpected service")
+            if status_characteristic.service_uuid != service.uuid:
+                raise ValueError("K1 status characteristic is attached to an unexpected service")
+            if "read" not in status_characteristic.properties:
+                raise ValueError("Reviewed K1 status characteristic is not readable")
+
+            properties = set(write_characteristic.properties)
+            write_characteristic_properties = tuple(sorted(properties))
+            max_without_response = _optional_int_attribute(
+                write_characteristic,
+                "max_write_without_response_size",
+            )
+            resolved_write_mode: ResolvedWriteMode
+            if write_mode == "auto":
+                if "write-without-response" in properties:
+                    resolved_write_mode = "without_response"
+                elif "write" in properties:
+                    resolved_write_mode = "with_response"
+                else:
+                    raise ValueError("Reviewed K1 characteristic is not writable")
+            elif write_mode == "with_response":
+                if "write" not in properties:
+                    raise ValueError(
+                        "Reviewed K1 characteristic does not advertise writes with response"
+                    )
+                resolved_write_mode = "with_response"
+            else:
+                resolved_write_mode = "without_response"
+
+            resolved_write_mode_for_error = resolved_write_mode
+            if resolved_write_mode == "without_response":
+                if max_without_response is None:
+                    raise ValueError("Negotiated write-without-response size is unavailable")
+                if len(frame) > max_without_response:
+                    raise ValueError(
+                        "Provisioning frame exceeds the negotiated write-without-response size"
+                    )
+
+            operation_stage = "baseline-read"
+            progress.operation_stage = operation_stage
+            baseline_value = bytes(await client.read_gatt_char(status_characteristic))
+            baseline = parse_wifi_status(baseline_value)
+            if active_captured_device is not None and not (
+                mark_captured_device_gatt_validated(active_captured_device)
+            ):
+                raise RuntimeError(
+                    "Exact BLE recovery handle changed before provisioning validation"
+                )
+            gatt_baseline_validated = True
+
+            operation_stage = "gatt-write"
+            # This callback is the durable side-effect boundary. It must finish
+            # before CoreBluetooth receives the frame, so a process crash can
+            # only create a conservative false-positive fence, never an unsafe
+            # forgotten write. It receives no SSID, password, or frame bytes.
+            if on_write_dispatch is not None:
+                on_write_dispatch(baseline, resolved_write_mode)
+            device_write_attempted = True
+            progress.operation_stage = operation_stage
+            progress.device_write_attempted = True
+            await client.write_gatt_char(
+                write_characteristic,
+                frame,
+                response=resolved_write_mode == "with_response",
+            )
+            device_write_confirmed = resolved_write_mode == "with_response"
+            progress.device_write_confirmed = device_write_confirmed
+            write_completed = monotonic()
+            deadline = write_completed + timeout_seconds
+
+            operation_stage = "status-poll"
+            progress.operation_stage = operation_stage
+            while monotonic() < deadline:
+                try:
+                    value = bytes(await client.read_gatt_char(status_characteristic))
+                except BleakError:
+                    if not client.is_connected:
+                        disconnected = True
+                        break
+                    raise
+                status = parse_wifi_status(value)
+                observation: StatusObservation = {
+                    "observed_at_utc": utc_now_iso(),
+                    "seconds_after_write": round(monotonic() - write_completed, 3),
+                    "status": status,
+                }
+                if not observations or status != observations[-1]["status"]:
+                    observations.append(observation)
+                if status["ipv4"] not in (None, AP_FALLBACK_IPV4):
+                    break
+                await asyncio.sleep(poll_interval_seconds)
+
+            return {
+                "schema_version": 1,
+                "profile_id": PROFILE_ID,
+                "started_at_utc": started_at,
+                "completed_at_utc": utc_now_iso(),
+                "adapter": "CoreBluetooth",
+                "bleak_version": version("bleak"),
+                "device_macos_uuid": device_macos_uuid,
+                "device_name": device_name,
+                "service_uuid": service.uuid,
+                "write_characteristic_uuid": write_characteristic.uuid,
+                "status_characteristic_uuid": status_characteristic.uuid,
+                "operation": "single_reviewed_wifi_provisioning_write",
+                "write_mode": resolved_write_mode,
+                "write_without_response_advertised": ("write-without-response" in properties),
+                "max_write_without_response_size": max_without_response,
+                "frame_length": len(frame),
+                "baseline_status": baseline,
+                "observations": observations,
+                "outcome": _outcome(baseline, observations, disconnected),
+            }
+    except Exception as exc:
+        if active_captured_device is not None and not gatt_baseline_validated:
+            demote_connected_device_handle_after_gatt_failure(active_captured_device)
+        _annotate_ble_operation_error(
+            exc,
+            operation_stage=operation_stage,
+            device_write_attempted=device_write_attempted,
+            device_write_confirmed=device_write_confirmed,
+            resolved_write_mode=resolved_write_mode_for_error,
+            write_characteristic_properties=write_characteristic_properties,
+            max_write_without_response_size=max_without_response,
+            frame_length=len(frame),
+        )
+        raise
+    finally:
+        # Hard-timeout cancellation may bypass ``except Exception``.  The
+        # failed live transport lease must still be demoted when the exact
+        # captured object never completed the reviewed baseline read; the
+        # UUID/session recovery token itself remains available for a later
+        # explicit attempt.
+        if active_captured_device is not None and not gatt_baseline_validated:
+            demote_connected_device_handle_after_gatt_failure(active_captured_device)
+        frame[:] = b"\x00" * len(frame)
 
 
 async def provision_wifi_once(
@@ -252,160 +779,40 @@ async def provision_wifi_once(
     timeout_seconds: float = 45.0,
     poll_interval_seconds: float = 1.0,
     write_mode: WriteMode = "auto",
+    *,
+    captured_device: CapturedDiscoveredDevice | None = None,
+    recovery_device_session_id: str | None = None,
+    on_write_dispatch: Callable[[WifiStatus, ResolvedWriteMode], None] | None = None,
 ) -> WifiProvisioningResult:
     """Perform one reviewed provisioning write and poll the K1 status characteristic."""
+
     if timeout_seconds <= 0:
         raise ValueError("timeout_seconds must be positive")
     if poll_interval_seconds <= 0:
         raise ValueError("poll_interval_seconds must be positive")
     if write_mode not in ("auto", "with_response", "without_response"):
         raise ValueError(f"Unsupported write mode: {write_mode}")
-
-    frame = build_wifi_provisioning_frame(ssid, password)
-    started_at = utc_now_iso()
-    observations: list[StatusObservation] = []
-    disconnected = False
-    operation_stage: BleOperationStage = "resolution"
-    device_write_attempted = False
-    device_write_confirmed = False
-
-    try:
-        async with asyncio.timeout(timeout_seconds + 25.0):
-            # The explicit UI scan and its selected network action are one
-            # CoreBluetooth lifecycle. Physical acceptance proved that a
-            # second UUID lookup can fail moments after a successful scan, so
-            # use the exact retained handle while its short lease is fresh.
-            selection = discovered_device_selection(device_macos_uuid)
-            device = selection.device
-            if device is None and not selection.from_fresh_scan:
-                # Non-UI callers without a current explicit scan retain the
-                # bounded lookup fallback. A fresh scan missing this device is
-                # authoritative and must not be silently replaced here.
-                device = await BleakScanner.find_device_by_address(
-                    device_macos_uuid,
-                    timeout=min(20.0, timeout_seconds),
-                )
-            if device is None:
-                raise BleakDeviceNotFoundError(
-                    device_macos_uuid,
-                    "Device was not rediscovered; keep the K1 powered and nearby.",
-                )
-
-            operation_stage = "connect"
-            async with BleakClient(device, timeout=timeout_seconds, pair=False) as client:
-                device_name = client.name
-                operation_stage = "gatt-contract"
-                service = client.services.get_service(SERVICE_UUID)
-                write_characteristic = client.services.get_characteristic(WRITE_CHARACTERISTIC_UUID)
-                status_characteristic = client.services.get_characteristic(
-                    STATUS_CHARACTERISTIC_UUID
-                )
-                if service is None:
-                    raise ValueError(f"Reviewed K1 service not found: {SERVICE_UUID}")
-                if write_characteristic is None:
-                    raise ValueError(
-                        f"Reviewed K1 write characteristic not found: {WRITE_CHARACTERISTIC_UUID}"
-                    )
-                if status_characteristic is None:
-                    raise ValueError(
-                        f"Reviewed K1 status characteristic not found: {STATUS_CHARACTERISTIC_UUID}"
-                    )
-                if write_characteristic.service_uuid != service.uuid:
-                    raise ValueError("K1 write characteristic is attached to an unexpected service")
-                if status_characteristic.service_uuid != service.uuid:
-                    raise ValueError(
-                        "K1 status characteristic is attached to an unexpected service"
-                    )
-                if "read" not in status_characteristic.properties:
-                    raise ValueError("Reviewed K1 status characteristic is not readable")
-
-                properties = set(write_characteristic.properties)
-                max_without_response = write_characteristic.max_write_without_response_size
-                resolved_write_mode: ResolvedWriteMode
-                if write_mode == "auto":
-                    if "write-without-response" in properties:
-                        resolved_write_mode = "without_response"
-                    elif "write" in properties:
-                        resolved_write_mode = "with_response"
-                    else:
-                        raise ValueError("Reviewed K1 characteristic is not writable")
-                elif write_mode == "with_response":
-                    if "write" not in properties:
-                        raise ValueError(
-                            "Reviewed K1 characteristic does not advertise writes with response"
-                        )
-                    resolved_write_mode = "with_response"
-                else:
-                    if len(frame) > max_without_response:
-                        raise ValueError(
-                            "Provisioning frame exceeds the negotiated write-without-response size"
-                        )
-                    resolved_write_mode = "without_response"
-
-                operation_stage = "baseline-read"
-                baseline_value = bytes(await client.read_gatt_char(status_characteristic))
-                baseline = parse_wifi_status(baseline_value)
-
-                operation_stage = "gatt-write"
-                device_write_attempted = True
-                await client.write_gatt_char(
-                    write_characteristic,
-                    frame,
-                    response=resolved_write_mode == "with_response",
-                )
-                device_write_confirmed = resolved_write_mode == "with_response"
-                write_completed = monotonic()
-                deadline = write_completed + timeout_seconds
-
-                operation_stage = "status-poll"
-                while monotonic() < deadline:
-                    try:
-                        value = bytes(await client.read_gatt_char(status_characteristic))
-                    except BleakError:
-                        if not client.is_connected:
-                            disconnected = True
-                            break
-                        raise
-                    status = parse_wifi_status(value)
-                    observation: StatusObservation = {
-                        "observed_at_utc": utc_now_iso(),
-                        "seconds_after_write": round(monotonic() - write_completed, 3),
-                        "status": status,
-                    }
-                    if not observations or status != observations[-1]["status"]:
-                        observations.append(observation)
-                    if status["ipv4"] not in (None, AP_FALLBACK_IPV4):
-                        break
-                    await asyncio.sleep(poll_interval_seconds)
-
-                return {
-                    "schema_version": 1,
-                    "profile_id": PROFILE_ID,
-                    "started_at_utc": started_at,
-                    "completed_at_utc": utc_now_iso(),
-                    "adapter": "CoreBluetooth",
-                    "bleak_version": version("bleak"),
-                    "device_macos_uuid": device_macos_uuid,
-                    "device_name": device_name,
-                    "service_uuid": service.uuid,
-                    "write_characteristic_uuid": write_characteristic.uuid,
-                    "status_characteristic_uuid": status_characteristic.uuid,
-                    "operation": "single_reviewed_wifi_provisioning_write",
-                    "write_mode": resolved_write_mode,
-                    "write_without_response_advertised": ("write-without-response" in properties),
-                    "max_write_without_response_size": max_without_response,
-                    "frame_length": len(frame),
-                    "baseline_status": baseline,
-                    "observations": observations,
-                    "outcome": _outcome(baseline, observations, disconnected),
-                }
-    except Exception as exc:
-        _annotate_ble_operation_error(
-            exc,
-            operation_stage=operation_stage,
-            device_write_attempted=device_write_attempted,
-            device_write_confirmed=device_write_confirmed,
+    if captured_device is not None and recovery_device_session_id is not None:
+        raise ValueError(
+            "captured_device and recovery_device_session_id are mutually exclusive"
         )
-        raise
-    finally:
-        frame[:] = b"\x00" * len(frame)
+    if recovery_device_session_id == "":
+        raise ValueError("recovery_device_session_id must not be empty")
+    progress = BleOperationProgress(operation_stage="resolution")
+    return await run_ble_operation(
+        "wifi-provision",
+        hard_timeout_seconds=(timeout_seconds + BLE_PROVISION_HARD_TIMEOUT_GRACE_SECONDS),
+        operation=lambda operation_progress: _provision_wifi_impl(
+            device_macos_uuid,
+            ssid,
+            password,
+            timeout_seconds=timeout_seconds,
+            poll_interval_seconds=poll_interval_seconds,
+            write_mode=write_mode,
+            captured_device=captured_device,
+            recovery_device_session_id=recovery_device_session_id,
+            on_write_dispatch=on_write_dispatch,
+            progress=operation_progress,
+        ),
+        progress=progress,
+    )

src/k1link/device_plugins/xgrids_k1/cli.py

@@ -1,13 +1,21 @@
 from __future__ import annotations
 
 import asyncio
+import fcntl
+import http.client
 import json
+import os
 import platform
 import shutil
+import stat
 import subprocess
 import sys
+from collections.abc import Iterator
+from contextlib import contextmanager
+from dataclasses import dataclass
 from pathlib import Path
-from typing import Annotated, TypedDict
+from typing import Annotated, Literal, TypedDict
+from uuid import UUID
 
 import typer
 import uvicorn
@@ -35,8 +43,17 @@ from k1link.device_plugins.xgrids_k1.analyze import (
     run_calibrated_overlay_experiment,
     summarize_mqtt_streams,
 )
+from k1link.device_plugins.xgrids_k1.application_control_process_lease import (
+    ApplicationControlProcessLease,
+    ApplicationControlProcessLeaseError,
+)
 from k1link.device_plugins.xgrids_k1.archive import discover_legacy_viewer_sessions
 from k1link.device_plugins.xgrids_k1.ble.gatt import dump_metadata
+from k1link.device_plugins.xgrids_k1.ble.runtime_arbiter import (
+    borrow_ble_runtime_process_lease,
+    configure_ble_runtime_process_lease,
+    defer_until_ble_runtime_idle,
+)
 from k1link.device_plugins.xgrids_k1.ble.scanner import scan
 from k1link.device_plugins.xgrids_k1.ble.wifi_provisioning import (
     PROFILE_ID,
@@ -54,6 +71,10 @@ from k1link.device_plugins.xgrids_k1.mqtt import (
     capture_mqtt,
 )
 from k1link.device_plugins.xgrids_k1.net.snapshot import snapshot
+from k1link.device_plugins.xgrids_k1.physical_command_ledger import (
+    PhysicalCommandLedger,
+    active_operator_retirements,
+)
 from k1link.device_plugins.xgrids_k1.protocol.application_authority import (
     ApplicationAuthorityLoadError,
     MacOSKeychainApplicationAuthorityProvisioner,
@@ -93,6 +114,243 @@ app.add_typer(compute_app, name="compute")
 app.add_typer(lab_app, name="lab")
 app.add_typer(artifact_app, name="artifact")
 
+_CANONICAL_MISSION_CORE_PORT = 8000
+_MISSION_CORE_SERVE_LOCK_FILENAME = ".serve.lock"
+
+
+class _MissionCoreServeLeaseError(RuntimeError):
+    """The canonical backend singleton lock cannot be trusted."""
+
+
+class _MissionCoreServeLeaseUnavailable(_MissionCoreServeLeaseError):
+    """Another live process owns canonical backend startup or runtime."""
+
+
+@dataclass(slots=True)
+class _MissionCoreServeLease:
+    """Stable OS-owned lease for the complete canonical Uvicorn lifetime."""
+
+    path: Path
+    _descriptor: int
+    _identity: tuple[int, int]
+    _released: bool = False
+
+    @classmethod
+    def acquire(cls, repository_root: Path) -> _MissionCoreServeLease:
+        runtime_dir = repository_root.expanduser().resolve() / ".runtime"
+        _ensure_serve_runtime_directory(runtime_dir)
+        lock_dir = runtime_dir / "mission-core"
+        _ensure_private_serve_lock_directory(lock_dir)
+        path = lock_dir / _MISSION_CORE_SERVE_LOCK_FILENAME
+        common_flags = os.O_RDWR | getattr(os, "O_CLOEXEC", 0)
+        common_flags |= getattr(os, "O_NOFOLLOW", 0)
+        try:
+            descriptor = os.open(path, common_flags | os.O_CREAT | os.O_EXCL, 0o600)
+        except FileExistsError:
+            try:
+                descriptor = os.open(path, common_flags)
+            except OSError as exc:
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock cannot be opened safely"
+                ) from exc
+        except OSError as exc:
+            raise _MissionCoreServeLeaseError(
+                "Mission Core serve lock cannot be created safely"
+            ) from exc
+
+        locked = False
+        try:
+            opened = os.fstat(descriptor)
+            _validate_private_serve_lock_file(opened)
+            try:
+                current = path.lstat()
+            except OSError as exc:
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock identity is unavailable"
+                ) from exc
+            _validate_private_serve_lock_file(current)
+            identity = (opened.st_dev, opened.st_ino)
+            if identity != (current.st_dev, current.st_ino):
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock changed while opening"
+                )
+            try:
+                fcntl.flock(descriptor, fcntl.LOCK_EX | fcntl.LOCK_NB)
+                locked = True
+            except BlockingIOError as exc:
+                raise _MissionCoreServeLeaseUnavailable(
+                    "another process owns Mission Core startup or runtime"
+                ) from exc
+            except OSError as exc:
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock cannot be acquired safely"
+                ) from exc
+            try:
+                locked_path = path.lstat()
+            except OSError as exc:
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock disappeared after acquisition"
+                ) from exc
+            _validate_private_serve_lock_file(locked_path)
+            if identity != (locked_path.st_dev, locked_path.st_ino):
+                raise _MissionCoreServeLeaseError(
+                    "Mission Core serve lock changed during acquisition"
+                )
+            return cls(path=path, _descriptor=descriptor, _identity=identity)
+        except BaseException:
+            if locked:
+                _unlock_serve_descriptor(descriptor)
+            os.close(descriptor)
+            raise
+
+    def release(self) -> None:
+        if self._released:
+            return
+        try:
+            _unlock_serve_descriptor(self._descriptor)
+        finally:
+            os.close(self._descriptor)
+            self._released = True
+
+    def __enter__(self) -> _MissionCoreServeLease:
+        return self
+
+    def __exit__(self, *_: object) -> None:
+        self.release()
+
+
+def _ensure_serve_runtime_directory(path: Path) -> None:
+    try:
+        path.mkdir(mode=0o700, parents=False, exist_ok=False)
+    except FileExistsError:
+        pass
+    except OSError as exc:
+        raise _MissionCoreServeLeaseError(
+            "Mission Core runtime directory is unavailable"
+        ) from exc
+    try:
+        metadata = path.lstat()
+    except OSError as exc:
+        raise _MissionCoreServeLeaseError(
+            "Mission Core runtime directory identity is unavailable"
+        ) from exc
+    if not stat.S_ISDIR(metadata.st_mode):
+        raise _MissionCoreServeLeaseError(
+            "Mission Core runtime directory is not a regular directory"
+        )
+
+
+def _ensure_private_serve_lock_directory(path: Path) -> None:
+    try:
+        path.mkdir(mode=0o700, parents=False, exist_ok=False)
+    except FileExistsError:
+        pass
+    except OSError as exc:
+        raise _MissionCoreServeLeaseError(
+            "Mission Core serve lock directory is unavailable"
+        ) from exc
+    try:
+        metadata = path.lstat()
+    except OSError as exc:
+        raise _MissionCoreServeLeaseError(
+            "Mission Core serve lock directory identity is unavailable"
+        ) from exc
+    if not stat.S_ISDIR(metadata.st_mode) or stat.S_IMODE(metadata.st_mode) != 0o700:
+        raise _MissionCoreServeLeaseError(
+            "Mission Core serve lock directory is not private"
+        )
+
+
+def _validate_private_serve_lock_file(metadata: os.stat_result) -> None:
+    if (
+        not stat.S_ISREG(metadata.st_mode)
+        or stat.S_IMODE(metadata.st_mode) != 0o600
+        or metadata.st_nlink != 1
+    ):
+        raise _MissionCoreServeLeaseError(
+            "Mission Core serve lock is not a private regular file"
+        )
+
+
+def _unlock_serve_descriptor(descriptor: int) -> None:
+    try:
+        fcntl.flock(descriptor, fcntl.LOCK_UN)
+    except OSError:
+        return
+
+
+def _acquire_mission_core_serve_lease(repository_root: Path) -> _MissionCoreServeLease:
+    return _MissionCoreServeLease.acquire(repository_root)
+
+
+def _configure_ble_process_lease() -> Path:
+    """Pin every standalone CLI BLE call to Mission Core's global K1 lock."""
+
+    repository_root = Path(__file__).resolve().parents[4]
+    configure_ble_runtime_process_lease(repository_root)
+    return repository_root
+
+
+class _CliWifiProvisioningBlocked(RuntimeError):
+    """The standalone mutating BLE command lacks safe physical authority."""
+
+
+def _canonical_corebluetooth_uuid(value: str) -> str:
+    try:
+        return str(UUID(str(value).strip())).upper()
+    except (AttributeError, ValueError) as exc:
+        raise _CliWifiProvisioningBlocked(
+            "BLE device target is not a canonical CoreBluetooth UUID"
+        ) from exc
+
+
+def _require_cli_wifi_target_not_retired(
+    repository_root: Path,
+    device: str,
+) -> str:
+    canonical_device = _canonical_corebluetooth_uuid(device)
+    snapshot = PhysicalCommandLedger(repository_root).snapshot()
+    if snapshot.status == "corrupt":
+        raise _CliWifiProvisioningBlocked(
+            "physical command audit is unavailable; Wi-Fi write is blocked"
+        )
+    record = snapshot.record
+    if record is not None:
+        for retirement in active_operator_retirements(record):
+            retired_device = _canonical_corebluetooth_uuid(
+                retirement.retired_transport_ref
+            )
+            if retired_device == canonical_device:
+                raise _CliWifiProvisioningBlocked(
+                    "the selected BLE UUID belongs to an operator-retired K1"
+                )
+    return canonical_device
+
+
+@contextmanager
+def _cli_wifi_mutation_lease(device: str) -> Iterator[str]:
+    """Fence one explicit CLI Wi-Fi write against retirement and Mission Core."""
+
+    repository_root = _configure_ble_process_lease()
+    lease = ApplicationControlProcessLease.acquire(repository_root)
+    try:
+        # This durable check is intentionally after acquiring the same global
+        # flock as the backend retirement action.  Retirement either commits
+        # first and this write is denied, or this explicit write owns the fence
+        # through credential entry and the complete native BLE lifecycle.
+        canonical_device = _require_cli_wifi_target_not_retired(
+            repository_root,
+            device,
+        )
+        with borrow_ble_runtime_process_lease(lease):
+            yield canonical_device
+    finally:
+        # A hard CoreBluetooth timeout may return before its native cleanup
+        # task is terminal.  Never expose the retirement boundary until the
+        # shared BLE arbiter proves idle; poison intentionally keeps it held
+        # until process restart.
+        defer_until_ble_runtime_idle(lease.release)
+
 
 class ToolStatus(TypedDict):
     name: str
@@ -837,25 +1095,133 @@ def publish_e26_lab(
 def serve_console(
     port: Annotated[
         int,
-        typer.Option(min=1024, max=65535, help="Loopback HTTP port for the local console."),
+        typer.Option(
+            min=1024,
+            max=65535,
+            help="Canonical loopback HTTP port; only 8000 is accepted.",
+        ),
     ] = 8000,
 ) -> None:
     """Serve the built Mission Core Control Station and loopback control API."""
-    frontend = Path(__file__).resolve().parents[4] / "apps" / "control-station" / "dist"
-    if not frontend.is_dir():
+    if port != _CANONICAL_MISSION_CORE_PORT:
         console.print(
-            "[red]Frontend build is missing.[/red] Run npm install && npm run build "
-            "inside apps/control-station."
+            f"[red]Mission Core запускается только на каноническом порту "
+            f"{_CANONICAL_MISSION_CORE_PORT}.[/red]"
         )
+        console.print("Другой локальный backend не создан.")
         raise typer.Exit(code=2)
-    console.print(f"NODEDC MISSION CORE: http://127.0.0.1:{port}")
-    console.print("The credential endpoint is bound to this Mac only.")
-    uvicorn.run(
-        "k1link.web.app:app",
-        host="127.0.0.1",
-        port=port,
-        log_level="info",
-        access_log=True,
+
+    repository_root = Path(__file__).resolve().parents[4]
+    try:
+        lease = _acquire_mission_core_serve_lease(repository_root)
+    except _MissionCoreServeLeaseUnavailable:
+        local_server = _local_server_status(_CANONICAL_MISSION_CORE_PORT)
+        if local_server == "mission-core":
+            _print_existing_mission_core(_CANONICAL_MISSION_CORE_PORT)
+            return
+        if local_server == "free":
+            console.print(
+                "[yellow]Mission Core уже запускается или завершает работу.[/yellow]"
+            )
+        else:
+            console.print(
+                "[red]Запуск Mission Core уже выполняется, но канонический health endpoint "
+                "пока не подтверждён.[/red]"
+            )
+        console.print(
+            "Второй backend не создан. Повторите запуск после завершения текущего перехода."
+        )
+        raise typer.Exit(code=2) from None
+    except _MissionCoreServeLeaseError as exc:
+        console.print(
+            "[red]Не удалось безопасно получить блокировку канонического Mission Core.[/red]"
+        )
+        console.print(f"Второй backend не создан: {exc}")
+        raise typer.Exit(code=2) from exc
+
+    with lease:
+        local_server = _local_server_status(_CANONICAL_MISSION_CORE_PORT)
+        if local_server == "mission-core":
+            _print_existing_mission_core(_CANONICAL_MISSION_CORE_PORT)
+            return
+        if local_server == "occupied":
+            console.print(
+                f"[red]Порт 127.0.0.1:{_CANONICAL_MISSION_CORE_PORT} уже занят другим "
+                "или неготовым процессом.[/red]"
+            )
+            console.print(
+                "Mission Core не стал создавать второй backend. Остановите точный "
+                "процесс-владелец порта и повторите запуск."
+            )
+            raise typer.Exit(code=2)
+
+        frontend = repository_root / "apps" / "control-station" / "dist"
+        if not frontend.is_dir():
+            console.print(
+                "[red]Frontend build is missing.[/red] Run npm install && npm run build "
+                "inside apps/control-station."
+            )
+            raise typer.Exit(code=2)
+        console.print(
+            f"NODEDC MISSION CORE: http://127.0.0.1:{_CANONICAL_MISSION_CORE_PORT}"
+        )
+        console.print("The credential endpoint is bound to this Mac only.")
+        uvicorn.run(
+            "k1link.web.app:app",
+            host="127.0.0.1",
+            port=_CANONICAL_MISSION_CORE_PORT,
+            log_level="info",
+            access_log=True,
+        )
+
+
+def _print_existing_mission_core(port: int) -> None:
+    console.print(
+        f"[green]NODEDC MISSION CORE уже запущен:[/green] http://127.0.0.1:{port}"
+    )
+    console.print("Используется единственный локальный backend; второй процесс не создан.")
+
+
+def _local_server_status(port: int) -> Literal["free", "mission-core", "occupied"]:
+    """Classify the loopback listener before starting the singleton backend.
+
+    A healthy Mission Core listener makes ``k1link serve`` idempotent.  Any
+    other listener fails with a short operator-facing error instead of letting
+    Uvicorn start application resources and then emit an opaque ``Errno 48``.
+    The check is observational only and never kills or replaces a process.
+    """
+
+    connection = http.client.HTTPConnection("127.0.0.1", port, timeout=0.75)
+    try:
+        connection.request(
+            "GET",
+            "/api/health",
+            headers={"Accept": "application/json", "Connection": "close"},
+        )
+        response = connection.getresponse()
+        if response.status != 200:
+            return "occupied"
+        body = response.read(16_385)
+        if len(body) > 16_384:
+            return "occupied"
+        payload = json.loads(body.decode("utf-8"))
+    except ConnectionRefusedError:
+        return "free"
+    except (
+        OSError,
+        TimeoutError,
+        UnicodeDecodeError,
+        json.JSONDecodeError,
+        http.client.HTTPException,
+    ):
+        return "occupied"
+    finally:
+        connection.close()
+    return (
+        "mission-core"
+        if isinstance(payload, dict)
+        and payload.get("service") == "mission-core-control-plane"
+        else "occupied"
     )
 
 
@@ -871,6 +1237,7 @@ def ble_scan(
     ] = 30.0,
 ) -> None:
     """Discover BLE advertisements without connecting or changing device configuration."""
+    _configure_ble_process_lease()
     try:
         result = asyncio.run(scan(duration))
     except (BleakError, OSError, ValueError) as exc:
@@ -908,6 +1275,7 @@ def ble_gatt_dump(
     ] = 45.0,
 ) -> None:
     """Enumerate GATT metadata only: no characteristic reads, subscriptions or writes."""
+    _configure_ble_process_lease()
     console.print(
         "Connecting for service discovery only; no characteristic values will be read or written."
     )
@@ -963,28 +1331,40 @@ def ble_wifi_configure(
         )
         raise typer.Exit(code=2)
 
-    console.print(
-        "Two local macOS dialogs will request the Wi-Fi name and hidden password. "
-        "The password is never printed, logged, or written to the result file; "
-        "the K1 may echo the SSID in the ignored sensitive status result."
-    )
+    ssid = ""
+    password = ""
     try:
-        ssid, password = prompt_wifi_credentials()
+        with _cli_wifi_mutation_lease(device) as canonical_device:
+            console.print(
+                "Two local macOS dialogs will request the Wi-Fi name and hidden password. "
+                "The password is never printed, logged, or written to the result file; "
+                "the K1 may echo the SSID in the ignored sensitive status result."
+            )
+            ssid, password = prompt_wifi_credentials()
+            console.print(
+                "Credentials accepted locally. Starting the single reviewed BLE write."
+            )
+            result = asyncio.run(
+                provision_wifi_once(
+                    canonical_device,
+                    ssid,
+                    password,
+                    timeout_seconds=timeout,
+                    write_mode=write_mode,
+                )
+            )
     except CredentialDialogError as exc:
         console.print(f"[red]Credential entry failed:[/red] {exc}")
         raise typer.Exit(code=2) from exc
-
-    console.print("Credentials accepted locally. Starting the single reviewed BLE write.")
-    try:
-        result = asyncio.run(
-            provision_wifi_once(
-                device,
-                ssid,
-                password,
-                timeout_seconds=timeout,
-                write_mode=write_mode,
-            )
+    except (
+        _CliWifiProvisioningBlocked,
+        ApplicationControlProcessLeaseError,
+    ) as exc:
+        console.print(
+            "[red]Wi-Fi provisioning blocked before credential or device access:[/red] "
+            f"{exc}"
         )
+        raise typer.Exit(code=2) from exc
     except (BleakError, OSError, TimeoutError, ValueError) as exc:
         console.print(f"[red]Wi-Fi provisioning failed:[/red] {type(exc).__name__}: {exc}")
         console.print("No automatic retry was attempted.")

src/k1link/device_plugins/xgrids_k1/device_identity_pin_store.py

@@ -0,0 +1,641 @@
+from __future__ import annotations
+
+import fcntl
+import hmac
+import json
+import os
+import re
+import stat
+import tempfile
+import threading
+from collections.abc import Iterator, Mapping
+from contextlib import contextmanager
+from dataclasses import dataclass
+from pathlib import Path
+from typing import IO, Literal, cast
+
+from k1link.sessions.store import resolve_missioncore_data_dir
+
+DEVICE_IDENTITY_PIN_SCHEMA = "missioncore.xgrids-k1-device-identity-pins/v1"
+DEVICE_IDENTITY_PIN_FILENAME = "device-identity-pins.json"
+DEVICE_IDENTITY_PIN_LOCK_FILENAME = ".device-identity-pins.lock"
+DEVICE_IDENTITY_PIN_MAX_BYTES = 64 * 1024
+DEVICE_IDENTITY_PIN_MAX_COUNT = 256
+DEVICE_IDENTITY_PIN_MAX_REVISION = (1 << 63) - 1
+
+DeviceIdentityPinStoreStatus = Literal["empty", "available", "corrupt"]
+
+_SAFE_IDENTIFIER = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:+-]{0,159}$")
+_MAX_VENDOR_DEVICE_ID_BYTES = 4 * 1024
+
+
+class DeviceIdentityPinStoreError(RuntimeError):
+    """Base error for the durable BLE-to-K1 identity binding."""
+
+    reason_code = "device-identity-pin-store-error"
+
+
+class DeviceIdentityPinStoreCorrupt(DeviceIdentityPinStoreError):
+    """The private identity store cannot be trusted and fails closed."""
+
+    reason_code = "device-identity-pin-store-corrupt"
+
+
+class DeviceIdentityPinMismatch(DeviceIdentityPinStoreError):
+    """Live DeviceInfo identity does not match the first-contact pin."""
+
+    reason_code = "device-identity-pin-mismatch"
+
+    def __init__(
+        self,
+        *,
+        transport_ref: str,
+        expected_vendor_device_id: str,
+        observed_vendor_device_id: str,
+        expected_compatibility_profile_id: str,
+        observed_compatibility_profile_id: str,
+    ) -> None:
+        super().__init__(
+            "live K1 identity/profile does not match the durable BLE transport pin"
+        )
+        self.transport_ref = transport_ref
+        self.expected_vendor_device_id = expected_vendor_device_id
+        self.observed_vendor_device_id = observed_vendor_device_id
+        self.expected_compatibility_profile_id = expected_compatibility_profile_id
+        self.observed_compatibility_profile_id = observed_compatibility_profile_id
+
+
+@dataclass(frozen=True, slots=True)
+class DeviceIdentityPin:
+    """Immutable first-contact binding for one CoreBluetooth transport."""
+
+    transport_ref: str
+    vendor_device_id: str
+    compatibility_profile_id: str
+
+    def as_dict(self) -> dict[str, str]:
+        return {
+            "transport_ref": self.transport_ref,
+            "vendor_device_id": self.vendor_device_id,
+            "compatibility_profile_id": self.compatibility_profile_id,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class DeviceIdentityPinDecision:
+    pin: DeviceIdentityPin
+    created: bool
+    revision: int
+
+
+@dataclass(frozen=True, slots=True)
+class DeviceIdentityPinStoreSnapshot:
+    status: DeviceIdentityPinStoreStatus
+    revision: int | None
+    pins: tuple[DeviceIdentityPin, ...]
+    reason_code: str | None
+
+    def for_transport(self, transport_ref: str) -> DeviceIdentityPin | None:
+        return next(
+            (pin for pin in self.pins if pin.transport_ref == transport_ref),
+            None,
+        )
+
+    def as_public_dict(self) -> dict[str, object]:
+        """Expose store health without publishing durable device identifiers."""
+
+        return {
+            "schema_version": DEVICE_IDENTITY_PIN_SCHEMA,
+            "status": self.status,
+            "revision": self.revision,
+            "pin_count": len(self.pins),
+            "reason_code": self.reason_code,
+        }
+
+
+class DeviceIdentityPinStore:
+    """Private, atomic first-contact K1 identity pins.
+
+    A pin binds one exact BLE ``transport_ref`` to the logical vendor
+    ``device_id`` proved by DeviceInfo and the reviewed compatibility profile.
+    An IP address is intentionally absent: TCP reachability can never substitute
+    for this identity. The stable flock serializes the complete
+    reload/check/publish transaction across Mission Core processes.
+    """
+
+    def __init__(self, repository_root: Path) -> None:
+        data_dir = resolve_missioncore_data_dir(repository_root)
+        self.path = data_dir / "xgrids-k1" / DEVICE_IDENTITY_PIN_FILENAME
+        self._lock_path = data_dir / "xgrids-k1" / DEVICE_IDENTITY_PIN_LOCK_FILENAME
+        self._data_dir = data_dir
+        self._thread_lock = threading.RLock()
+        self._revision = 0
+        self._pins: dict[str, DeviceIdentityPin] = {}
+        self._corrupt = False
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+
+    def snapshot(self) -> DeviceIdentityPinStoreSnapshot:
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt:
+                return DeviceIdentityPinStoreSnapshot(
+                    status="corrupt",
+                    revision=None,
+                    pins=(),
+                    reason_code=DeviceIdentityPinStoreCorrupt.reason_code,
+                )
+            pins = tuple(self._pins[key] for key in sorted(self._pins))
+            if not pins:
+                return DeviceIdentityPinStoreSnapshot(
+                    status="empty",
+                    revision=None,
+                    pins=(),
+                    reason_code=None,
+                )
+            return DeviceIdentityPinStoreSnapshot(
+                status="available",
+                revision=self._revision,
+                pins=pins,
+                reason_code=None,
+            )
+
+    def pin_or_match(
+        self,
+        *,
+        transport_ref: str,
+        vendor_device_id: str,
+        compatibility_profile_id: str,
+    ) -> DeviceIdentityPinDecision:
+        """Create the first pin or verify an exact existing pin.
+
+        Matching an existing pin is read-only and does not bump the revision or
+        rewrite the file. Any identity or profile mismatch is a typed,
+        fail-closed error and preserves the original bytes.
+        """
+
+        _validate_identifier(transport_ref, field_name="transport_ref")
+        _validate_vendor_device_id(vendor_device_id)
+        _validate_identifier(
+            compatibility_profile_id,
+            field_name="compatibility_profile_id",
+        )
+
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt:
+                raise DeviceIdentityPinStoreCorrupt(
+                    "device identity pin store is corrupt; live K1 identity was not adopted"
+                )
+
+            current = self._pins.get(transport_ref)
+            if current is not None:
+                vendor_matches = hmac.compare_digest(
+                    current.vendor_device_id,
+                    vendor_device_id,
+                )
+                profile_matches = hmac.compare_digest(
+                    current.compatibility_profile_id,
+                    compatibility_profile_id,
+                )
+                if not vendor_matches or not profile_matches:
+                    raise DeviceIdentityPinMismatch(
+                        transport_ref=transport_ref,
+                        expected_vendor_device_id=current.vendor_device_id,
+                        observed_vendor_device_id=vendor_device_id,
+                        expected_compatibility_profile_id=(
+                            current.compatibility_profile_id
+                        ),
+                        observed_compatibility_profile_id=compatibility_profile_id,
+                    )
+                return DeviceIdentityPinDecision(
+                    pin=current,
+                    created=False,
+                    revision=self._revision,
+                )
+
+            if len(self._pins) >= DEVICE_IDENTITY_PIN_MAX_COUNT:
+                raise DeviceIdentityPinStoreCorrupt(
+                    "device identity pin store reached its bounded pin count"
+                )
+            if self._revision >= DEVICE_IDENTITY_PIN_MAX_REVISION:
+                raise DeviceIdentityPinStoreCorrupt(
+                    "device identity pin store revision is exhausted"
+                )
+
+            pin = DeviceIdentityPin(
+                transport_ref=transport_ref,
+                vendor_device_id=vendor_device_id,
+                compatibility_profile_id=compatibility_profile_id,
+            )
+            next_pins = dict(self._pins)
+            next_pins[transport_ref] = pin
+            revision = self._revision + 1
+            self._persist_locked(revision=revision, pins=next_pins)
+            return DeviceIdentityPinDecision(
+                pin=pin,
+                created=True,
+                revision=revision,
+            )
+
+    @contextmanager
+    def _process_lock_locked(self) -> Iterator[None]:
+        data_dir_created = _ensure_private_directory(self._data_dir, parents=True)
+        if data_dir_created:
+            _fsync_directory(self._data_dir.parent)
+        store_dir_created = _ensure_private_directory(self.path.parent, parents=False)
+        if store_dir_created:
+            _fsync_directory(self._data_dir)
+
+        flags = os.O_RDWR | os.O_CREAT | getattr(os, "O_CLOEXEC", 0)
+        flags |= getattr(os, "O_NOFOLLOW", 0)
+        try:
+            descriptor = os.open(self._lock_path, flags, 0o600)
+        except OSError as exc:
+            raise DeviceIdentityPinStoreCorrupt(
+                "device identity pin lock cannot be opened safely"
+            ) from exc
+        stream: IO[bytes] | None = None
+        try:
+            try:
+                _validate_private_open_file(
+                    descriptor,
+                    self._lock_path,
+                    label="device identity pin lock",
+                    require_empty=True,
+                )
+            except ValueError as exc:
+                raise DeviceIdentityPinStoreCorrupt(
+                    "device identity pin lock is not a stable private file"
+                ) from exc
+            stream = os.fdopen(descriptor, "r+b", closefd=True)
+            descriptor = -1
+            fcntl.flock(stream.fileno(), fcntl.LOCK_EX)
+            try:
+                try:
+                    _validate_private_open_file(
+                        stream.fileno(),
+                        self._lock_path,
+                        label="device identity pin lock",
+                        require_empty=True,
+                    )
+                except ValueError as exc:
+                    raise DeviceIdentityPinStoreCorrupt(
+                        "device identity pin lock changed while being acquired"
+                    ) from exc
+                _fsync_directory(self.path.parent)
+                yield
+            finally:
+                fcntl.flock(stream.fileno(), fcntl.LOCK_UN)
+        finally:
+            if stream is not None:
+                stream.close()
+            elif descriptor >= 0:
+                os.close(descriptor)
+
+    def _persist_locked(
+        self,
+        *,
+        revision: int,
+        pins: Mapping[str, DeviceIdentityPin],
+    ) -> None:
+        ordered = tuple(pins[key] for key in sorted(pins))
+        payload: dict[str, object] = {
+            "schema_version": DEVICE_IDENTITY_PIN_SCHEMA,
+            "revision": revision,
+            "pins": [pin.as_dict() for pin in ordered],
+        }
+        _write_private_json_atomic(
+            self.path,
+            payload,
+            data_dir=self._data_dir,
+        )
+        self._revision = revision
+        self._pins = dict(pins)
+        self._corrupt = False
+
+    def _reload_locked(self) -> None:
+        try:
+            payload = _read_private_json(self.path)
+        except FileNotFoundError:
+            self._revision = 0
+            self._pins = {}
+            self._corrupt = False
+            return
+        except (OSError, UnicodeError, json.JSONDecodeError, TypeError, ValueError):
+            self._revision = 0
+            self._pins = {}
+            self._corrupt = True
+            return
+        try:
+            revision, pins = _document_from_mapping(payload)
+        except (TypeError, ValueError):
+            self._revision = 0
+            self._pins = {}
+            self._corrupt = True
+            return
+        self._revision = revision
+        self._pins = {pin.transport_ref: pin for pin in pins}
+        self._corrupt = False
+
+
+def _read_private_json(path: Path) -> object:
+    try:
+        initial = path.lstat()
+    except FileNotFoundError:
+        raise
+    except OSError as exc:
+        raise ValueError("device identity pin file cannot be inspected safely") from exc
+    _validate_private_metadata(
+        initial,
+        label="device identity pin file",
+        require_empty=False,
+    )
+
+    flags = os.O_RDONLY | getattr(os, "O_CLOEXEC", 0)
+    flags |= getattr(os, "O_NOFOLLOW", 0) | getattr(os, "O_NONBLOCK", 0)
+    try:
+        descriptor = os.open(path, flags)
+    except OSError as exc:
+        raise ValueError("device identity pin file cannot be opened safely") from exc
+    try:
+        metadata = _validate_private_open_file(
+            descriptor,
+            path,
+            label="device identity pin file",
+            require_empty=False,
+        )
+        if (initial.st_dev, initial.st_ino) != (metadata.st_dev, metadata.st_ino):
+            raise ValueError("device identity pin file changed while opening")
+        if metadata.st_size > DEVICE_IDENTITY_PIN_MAX_BYTES:
+            raise ValueError("device identity pin file exceeds the bounded size")
+        chunks: list[bytes] = []
+        remaining = DEVICE_IDENTITY_PIN_MAX_BYTES + 1
+        while remaining > 0:
+            chunk = os.read(descriptor, min(remaining, 64 * 1024))
+            if not chunk:
+                break
+            chunks.append(chunk)
+            remaining -= len(chunk)
+        raw = b"".join(chunks)
+        if len(raw) > DEVICE_IDENTITY_PIN_MAX_BYTES:
+            raise ValueError("device identity pin file exceeds the bounded size")
+    finally:
+        os.close(descriptor)
+    return json.loads(raw.decode("utf-8"), object_pairs_hook=_unique_json_object)
+
+
+def _write_private_json_atomic(
+    path: Path,
+    payload: Mapping[str, object],
+    *,
+    data_dir: Path,
+) -> None:
+    serialized = (
+        json.dumps(
+            payload,
+            ensure_ascii=False,
+            sort_keys=True,
+            separators=(",", ":"),
+            allow_nan=False,
+        )
+        + "\n"
+    ).encode("utf-8")
+    if len(serialized) > DEVICE_IDENTITY_PIN_MAX_BYTES:
+        raise ValueError("device identity pin payload exceeds the bounded size")
+
+    _ensure_private_directory(data_dir, parents=True)
+    _ensure_private_directory(path.parent, parents=False)
+    previous_identity = _existing_private_file_identity(path)
+
+    descriptor, temp_name = tempfile.mkstemp(
+        dir=path.parent,
+        prefix=f".{path.name}.",
+        suffix=".tmp",
+    )
+    temp_path = Path(temp_name)
+    try:
+        os.fchmod(descriptor, 0o600)
+        with os.fdopen(descriptor, "wb") as stream:
+            descriptor = -1
+            stream.write(serialized)
+            stream.flush()
+            os.fsync(stream.fileno())
+        _require_unchanged_existing_path(path, previous_identity)
+        os.replace(temp_path, path)
+        _fsync_directory(path.parent)
+    finally:
+        if descriptor >= 0:
+            os.close(descriptor)
+        temp_path.unlink(missing_ok=True)
+
+
+def _existing_private_file_identity(path: Path) -> tuple[int, int] | None:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        return None
+    _validate_private_metadata(
+        metadata,
+        label="device identity pin file",
+        require_empty=False,
+    )
+    return metadata.st_dev, metadata.st_ino
+
+
+def _require_unchanged_existing_path(
+    path: Path,
+    expected: tuple[int, int] | None,
+) -> None:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        if expected is None:
+            return
+        raise ValueError("device identity pin file disappeared during publication") from None
+    _validate_private_metadata(
+        metadata,
+        label="device identity pin file",
+        require_empty=False,
+    )
+    observed = metadata.st_dev, metadata.st_ino
+    if expected is None or observed != expected:
+        raise ValueError("device identity pin file changed during publication")
+
+
+def _validate_private_open_file(
+    descriptor: int,
+    path: Path,
+    *,
+    label: str,
+    require_empty: bool,
+) -> os.stat_result:
+    metadata = os.fstat(descriptor)
+    _validate_private_metadata(metadata, label=label, require_empty=require_empty)
+    try:
+        path_metadata = path.lstat()
+    except OSError as exc:
+        raise ValueError(f"{label} path cannot be verified") from exc
+    if (path_metadata.st_dev, path_metadata.st_ino) != (
+        metadata.st_dev,
+        metadata.st_ino,
+    ):
+        raise ValueError(f"{label} path does not reference the opened inode")
+    _validate_private_metadata(path_metadata, label=label, require_empty=False)
+    return metadata
+
+
+def _validate_private_metadata(
+    metadata: os.stat_result,
+    *,
+    label: str,
+    require_empty: bool,
+) -> None:
+    if not stat.S_ISREG(metadata.st_mode):
+        raise ValueError(f"{label} is not a regular file")
+    if stat.S_IMODE(metadata.st_mode) != 0o600:
+        raise ValueError(f"{label} is not private")
+    if metadata.st_nlink != 1:
+        raise ValueError(f"{label} has an unsafe hard link")
+    if require_empty and metadata.st_size != 0:
+        raise ValueError(f"{label} must remain empty")
+
+
+def _ensure_private_directory(path: Path, *, parents: bool) -> bool:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        try:
+            path.mkdir(mode=0o700, parents=parents, exist_ok=False)
+        except FileExistsError:
+            metadata = path.lstat()
+        else:
+            path.chmod(0o700)
+            return True
+    except OSError as exc:
+        raise DeviceIdentityPinStoreCorrupt(
+            "device identity pin directory is unavailable"
+        ) from exc
+    if not stat.S_ISDIR(metadata.st_mode):
+        raise DeviceIdentityPinStoreCorrupt(
+            "device identity pin directory is not a regular private directory"
+        )
+    if stat.S_IMODE(metadata.st_mode) != 0o700:
+        raise DeviceIdentityPinStoreCorrupt(
+            "device identity pin directory permissions are not private"
+        )
+    return False
+
+
+def _fsync_directory(path: Path) -> None:
+    flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0)
+    descriptor = os.open(path, flags)
+    try:
+        os.fsync(descriptor)
+    finally:
+        os.close(descriptor)
+
+
+def _unique_json_object(pairs: list[tuple[str, object]]) -> dict[str, object]:
+    document: dict[str, object] = {}
+    for key, value in pairs:
+        if key in document:
+            raise ValueError("device identity pin store contains duplicate fields")
+        document[key] = value
+    return document
+
+
+def _document_from_mapping(value: object) -> tuple[int, tuple[DeviceIdentityPin, ...]]:
+    document = _exact_mapping(
+        value,
+        {"schema_version", "revision", "pins"},
+        label="device identity pin document",
+    )
+    if document["schema_version"] != DEVICE_IDENTITY_PIN_SCHEMA:
+        raise ValueError("unsupported device identity pin schema")
+    revision = _positive_revision(document["revision"])
+    raw_pins = document["pins"]
+    if not isinstance(raw_pins, list) or not 1 <= len(raw_pins) <= DEVICE_IDENTITY_PIN_MAX_COUNT:
+        raise ValueError("device identity pin list has an invalid bounded size")
+
+    pins: list[DeviceIdentityPin] = []
+    seen_transport_refs: set[str] = set()
+    for raw_pin in raw_pins:
+        pin_document = _exact_mapping(
+            raw_pin,
+            {"transport_ref", "vendor_device_id", "compatibility_profile_id"},
+            label="device identity pin",
+        )
+        transport_ref = _required_string(
+            pin_document["transport_ref"],
+            field_name="transport_ref",
+        )
+        vendor_device_id = _required_string(
+            pin_document["vendor_device_id"],
+            field_name="vendor_device_id",
+        )
+        compatibility_profile_id = _required_string(
+            pin_document["compatibility_profile_id"],
+            field_name="compatibility_profile_id",
+        )
+        _validate_identifier(transport_ref, field_name="transport_ref")
+        _validate_vendor_device_id(vendor_device_id)
+        _validate_identifier(
+            compatibility_profile_id,
+            field_name="compatibility_profile_id",
+        )
+        if transport_ref in seen_transport_refs:
+            raise ValueError("device identity pin transport_ref is duplicated")
+        seen_transport_refs.add(transport_ref)
+        pins.append(
+            DeviceIdentityPin(
+                transport_ref=transport_ref,
+                vendor_device_id=vendor_device_id,
+                compatibility_profile_id=compatibility_profile_id,
+            )
+        )
+
+    if pins != sorted(pins, key=lambda pin: pin.transport_ref):
+        raise ValueError("device identity pins are not in canonical order")
+    if revision != len(pins):
+        raise ValueError("device identity pin revision does not match immutable pin count")
+    return revision, tuple(pins)
+
+
+def _exact_mapping(value: object, keys: set[str], *, label: str) -> Mapping[str, object]:
+    if not isinstance(value, dict) or set(value) != keys:
+        raise ValueError(f"{label} does not match the secret-free schema")
+    return cast(Mapping[str, object], value)
+
+
+def _required_string(value: object, *, field_name: str) -> str:
+    if not isinstance(value, str) or not value:
+        raise ValueError(f"{field_name} must be a non-empty string")
+    return value
+
+
+def _validate_identifier(value: str, *, field_name: str) -> None:
+    if _SAFE_IDENTIFIER.fullmatch(value) is None:
+        raise ValueError(f"{field_name} is outside the secret-free identifier schema")
+
+
+def _validate_vendor_device_id(value: str) -> None:
+    try:
+        encoded = value.encode("ascii")
+    except UnicodeEncodeError as exc:
+        raise ValueError("vendor_device_id must use printable ASCII") from exc
+    if not encoded or len(encoded) > _MAX_VENDOR_DEVICE_ID_BYTES:
+        raise ValueError("vendor_device_id is outside the bounded identity schema")
+    if any(byte <= 0x20 or byte > 0x7E for byte in encoded):
+        raise ValueError("vendor_device_id must use printable ASCII without spaces")
+
+
+def _positive_revision(value: object) -> int:
+    if (
+        isinstance(value, bool)
+        or not isinstance(value, int)
+        or value < 1
+        or value > DEVICE_IDENTITY_PIN_MAX_REVISION
+    ):
+        raise ValueError("device identity pin revision must be a bounded positive integer")
+    return value

src/k1link/device_plugins/xgrids_k1/host_diagnostics.py

@@ -0,0 +1,429 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Final, Literal
+
+from bleak.exc import (
+    BleakBluetoothNotAvailableError,
+    BleakBluetoothNotAvailableReason,
+)
+
+HostDiagnosticBoundary = Literal[
+    "corebluetooth",
+    "corewlan",
+    "keychain",
+    "route",
+    "tcp",
+    "mqtt",
+    "filesystem",
+]
+HostDiagnosticDomain = HostDiagnosticBoundary
+HostDiagnosticCode = Literal[
+    "host.bluetooth.permission-denied",
+    "host.bluetooth.adapter-powered-off",
+    "host.bluetooth.adapter-unavailable",
+    "host.bluetooth.runtime-unavailable",
+    "host.bluetooth.operation-timeout",
+    "host.wifi.permission-denied",
+    "host.wifi.adapter-powered-off",
+    "host.wifi.interface-unavailable",
+    "host.wifi.ssid-unavailable",
+    "host.wifi.operation-timeout",
+    "host.wifi.association-failed",
+    "host.keychain.interaction-required",
+    "host.keychain.permission-denied",
+    "host.keychain.unavailable",
+    "host.route.unavailable",
+    "host.tcp.connection-refused",
+    "host.tcp.connection-timeout",
+    "host.tcp.endpoint-unavailable",
+    "host.mqtt.connection-timeout",
+    "host.mqtt.connection-refused",
+    "host.mqtt.transport-unavailable",
+    "host.filesystem.permission-denied",
+    "host.filesystem.ledger-unavailable",
+]
+HostDiagnosticImpact = Literal[
+    "discovery",
+    "host-network",
+    "control",
+    "durable-safety",
+]
+HostDiagnosticAction = Literal[
+    "grant-bluetooth-permission",
+    "power-on-bluetooth",
+    "restore-bluetooth-adapter",
+    "grant-wifi-permission",
+    "power-on-wifi",
+    "restore-wifi-interface",
+    "unlock-or-authorize-keychain",
+    "review-keychain-access",
+    "join-expected-network",
+    "inspect-host-route",
+    "verify-broker-endpoint",
+    "inspect-local-storage",
+    "restart-local-service",
+    "explicit-retry",
+]
+
+
+@dataclass(frozen=True, slots=True)
+class HostFailureDiagnostic:
+    """Secret-free operator diagnostic for one host-side failure boundary."""
+
+    code: HostDiagnosticCode
+    domain: HostDiagnosticDomain
+    impact: HostDiagnosticImpact
+    operator_action: HostDiagnosticAction
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "schema_version": "missioncore.host-failure-diagnostic/v1",
+            "code": self.code,
+            "domain": self.domain,
+            "impact": self.impact,
+            "operator_action": self.operator_action,
+            "automatic_retry": False,
+            "redacted": True,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class _DiagnosticSpec:
+    code: HostDiagnosticCode
+    domain: HostDiagnosticDomain
+    impact: HostDiagnosticImpact
+    operator_action: HostDiagnosticAction
+
+    def diagnostic(self) -> HostFailureDiagnostic:
+        return HostFailureDiagnostic(
+            code=self.code,
+            domain=self.domain,
+            impact=self.impact,
+            operator_action=self.operator_action,
+        )
+
+
+_BLUETOOTH_PERMISSION = _DiagnosticSpec(
+    "host.bluetooth.permission-denied",
+    "corebluetooth",
+    "discovery",
+    "grant-bluetooth-permission",
+)
+_BLUETOOTH_POWERED_OFF = _DiagnosticSpec(
+    "host.bluetooth.adapter-powered-off",
+    "corebluetooth",
+    "discovery",
+    "power-on-bluetooth",
+)
+_BLUETOOTH_UNAVAILABLE = _DiagnosticSpec(
+    "host.bluetooth.adapter-unavailable",
+    "corebluetooth",
+    "discovery",
+    "restore-bluetooth-adapter",
+)
+_BLUETOOTH_RUNTIME_UNAVAILABLE = _DiagnosticSpec(
+    "host.bluetooth.runtime-unavailable",
+    "corebluetooth",
+    "discovery",
+    "restart-local-service",
+)
+_BLUETOOTH_TIMEOUT = _DiagnosticSpec(
+    "host.bluetooth.operation-timeout",
+    "corebluetooth",
+    "discovery",
+    "explicit-retry",
+)
+_WIFI_PERMISSION = _DiagnosticSpec(
+    "host.wifi.permission-denied",
+    "corewlan",
+    "host-network",
+    "grant-wifi-permission",
+)
+_WIFI_POWERED_OFF = _DiagnosticSpec(
+    "host.wifi.adapter-powered-off",
+    "corewlan",
+    "host-network",
+    "power-on-wifi",
+)
+_WIFI_UNAVAILABLE = _DiagnosticSpec(
+    "host.wifi.interface-unavailable",
+    "corewlan",
+    "host-network",
+    "restore-wifi-interface",
+)
+_WIFI_SSID_UNAVAILABLE = _DiagnosticSpec(
+    "host.wifi.ssid-unavailable",
+    "corewlan",
+    "host-network",
+    "join-expected-network",
+)
+_WIFI_TIMEOUT = _DiagnosticSpec(
+    "host.wifi.operation-timeout",
+    "corewlan",
+    "host-network",
+    "explicit-retry",
+)
+_WIFI_ASSOCIATION_FAILED = _DiagnosticSpec(
+    "host.wifi.association-failed",
+    "corewlan",
+    "host-network",
+    "join-expected-network",
+)
+_KEYCHAIN_INTERACTION_REQUIRED = _DiagnosticSpec(
+    "host.keychain.interaction-required",
+    "keychain",
+    "control",
+    "unlock-or-authorize-keychain",
+)
+_KEYCHAIN_PERMISSION = _DiagnosticSpec(
+    "host.keychain.permission-denied",
+    "keychain",
+    "control",
+    "review-keychain-access",
+)
+_KEYCHAIN_UNAVAILABLE = _DiagnosticSpec(
+    "host.keychain.unavailable",
+    "keychain",
+    "control",
+    "unlock-or-authorize-keychain",
+)
+_ROUTE_UNAVAILABLE = _DiagnosticSpec(
+    "host.route.unavailable",
+    "route",
+    "host-network",
+    "inspect-host-route",
+)
+_TCP_REFUSED = _DiagnosticSpec(
+    "host.tcp.connection-refused",
+    "tcp",
+    "control",
+    "verify-broker-endpoint",
+)
+_TCP_TIMEOUT = _DiagnosticSpec(
+    "host.tcp.connection-timeout",
+    "tcp",
+    "control",
+    "verify-broker-endpoint",
+)
+_TCP_UNAVAILABLE = _DiagnosticSpec(
+    "host.tcp.endpoint-unavailable",
+    "tcp",
+    "control",
+    "verify-broker-endpoint",
+)
+_MQTT_TIMEOUT = _DiagnosticSpec(
+    "host.mqtt.connection-timeout",
+    "mqtt",
+    "control",
+    "verify-broker-endpoint",
+)
+_MQTT_REFUSED = _DiagnosticSpec(
+    "host.mqtt.connection-refused",
+    "mqtt",
+    "control",
+    "verify-broker-endpoint",
+)
+_MQTT_UNAVAILABLE = _DiagnosticSpec(
+    "host.mqtt.transport-unavailable",
+    "mqtt",
+    "control",
+    "verify-broker-endpoint",
+)
+_FILESYSTEM_PERMISSION = _DiagnosticSpec(
+    "host.filesystem.permission-denied",
+    "filesystem",
+    "durable-safety",
+    "inspect-local-storage",
+)
+_LEDGER_UNAVAILABLE = _DiagnosticSpec(
+    "host.filesystem.ledger-unavailable",
+    "filesystem",
+    "durable-safety",
+    "inspect-local-storage",
+)
+
+
+_REASON_SPECS: Final[dict[str, _DiagnosticSpec]] = {
+    "ble-permission-denied": _BLUETOOTH_PERMISSION,
+    "ble-adapter-powered-off": _BLUETOOTH_POWERED_OFF,
+    "ble-adapter-unavailable": _BLUETOOTH_UNAVAILABLE,
+    "ble-runtime-owner-loop-conflict": _BLUETOOTH_RUNTIME_UNAVAILABLE,
+    "ble-runtime-restart-required": _BLUETOOTH_RUNTIME_UNAVAILABLE,
+    "connection-verify-runtime-loop-unavailable": _BLUETOOTH_RUNTIME_UNAVAILABLE,
+    "ble-discovery-timeout": _BLUETOOTH_TIMEOUT,
+    "ble-status-read-timeout": _BLUETOOTH_TIMEOUT,
+    "ble-provisioning-timeout": _BLUETOOTH_TIMEOUT,
+    "ble-ap-enable-timeout": _BLUETOOTH_TIMEOUT,
+    "corewlan-permission-denied": _WIFI_PERMISSION,
+    "corewlan-authorization-denied": _WIFI_PERMISSION,
+    "wifi-interface-inactive": _WIFI_POWERED_OFF,
+    "wifi-interface-unavailable": _WIFI_UNAVAILABLE,
+    "network-not-found": _WIFI_SSID_UNAVAILABLE,
+    "host-wifi-operation-timeout": _WIFI_TIMEOUT,
+    "corewlan-error": _WIFI_ASSOCIATION_FAILED,
+    "keychain-authorization-required": _KEYCHAIN_INTERACTION_REQUIRED,
+    "keychain-authorization-denied": _KEYCHAIN_PERMISSION,
+    "keychain-authorization-cancelled": _KEYCHAIN_PERMISSION,
+    "keychain-access-failed": _KEYCHAIN_UNAVAILABLE,
+    "application_authority_unavailable": _KEYCHAIN_UNAVAILABLE,
+    "host-route-unavailable": _ROUTE_UNAVAILABLE,
+    "host-path-unavailable": _ROUTE_UNAVAILABLE,
+    "host-route-interface-unavailable": _ROUTE_UNAVAILABLE,
+    "association-identity-unavailable": _ROUTE_UNAVAILABLE,
+    "host-path-observation-stale": _ROUTE_UNAVAILABLE,
+    "host-path-epoch-changed": _ROUTE_UNAVAILABLE,
+    "host-path-probe-error": _ROUTE_UNAVAILABLE,
+    "host-path-recheck-error": _ROUTE_UNAVAILABLE,
+    "connection-monitor-start-failed": _ROUTE_UNAVAILABLE,
+    "tcp-connection-refused": _TCP_REFUSED,
+    "tcp-connection-timeout": _TCP_TIMEOUT,
+    "tcp-endpoint-unreachable": _TCP_UNAVAILABLE,
+    "tcp-route-lost": _TCP_UNAVAILABLE,
+    "tcp-probe-error": _TCP_UNAVAILABLE,
+    "endpoint-unreachable": _TCP_UNAVAILABLE,
+    "endpoint-observation-stale": _TCP_UNAVAILABLE,
+    "quick_connect_endpoint_unreachable": _TCP_UNAVAILABLE,
+    "connection_lease_endpoint_unreachable_after_provision": _TCP_UNAVAILABLE,
+    "connection_lease_recovered_endpoint_unreachable": _TCP_UNAVAILABLE,
+    "mqtt_connection_timeout": _MQTT_TIMEOUT,
+    "mqtt_connect_call_failed": _MQTT_UNAVAILABLE,
+    "mqtt_connect_rejected": _MQTT_REFUSED,
+    "mqtt_broker_rejected_connection": _MQTT_REFUSED,
+    "mqtt_network_loop_failed": _MQTT_UNAVAILABLE,
+    "mqtt_connection_ended": _MQTT_UNAVAILABLE,
+    "mqtt_client_unavailable": _MQTT_UNAVAILABLE,
+    "mqtt_transport_failure": _MQTT_UNAVAILABLE,
+    "mqtt-control-loop-lost": _MQTT_UNAVAILABLE,
+    "control-proof-observation-stale": _MQTT_UNAVAILABLE,
+    "network-mutation-ledger-error": _LEDGER_UNAVAILABLE,
+    "network-mutation-ledger-corrupt": _LEDGER_UNAVAILABLE,
+    "network-provisioning-idempotency-error": _LEDGER_UNAVAILABLE,
+    "network-provisioning-idempotency-corrupt": _LEDGER_UNAVAILABLE,
+    "physical-command-ledger-error": _LEDGER_UNAVAILABLE,
+    "physical-command-ledger-corrupt": _LEDGER_UNAVAILABLE,
+    "semantic-topology-store-error": _LEDGER_UNAVAILABLE,
+    "semantic-topology-store-corrupt": _LEDGER_UNAVAILABLE,
+    "device-identity-pin-store-error": _LEDGER_UNAVAILABLE,
+    "device-identity-pin-store-corrupt": _LEDGER_UNAVAILABLE,
+    "application-control-process-lease-error": _LEDGER_UNAVAILABLE,
+    "application-control-process-lease-unavailable": _LEDGER_UNAVAILABLE,
+}
+
+
+def host_diagnostic_for_reason(reason_code: object) -> HostFailureDiagnostic | None:
+    """Map only reviewed reason codes; never reflect arbitrary input."""
+
+    if not isinstance(reason_code, str):
+        return None
+    spec = _REASON_SPECS.get(reason_code)
+    return spec.diagnostic() if spec is not None else None
+
+
+def host_diagnostic_for_exception(
+    exc: BaseException,
+    *,
+    boundary: HostDiagnosticBoundary | None = None,
+) -> HostFailureDiagnostic | None:
+    """Classify a host exception into a redacted diagnostic whitelist."""
+
+    explicit = host_diagnostic_for_reason(getattr(exc, "reason_code", None))
+    if explicit is not None:
+        return explicit
+    if isinstance(exc, ConnectionRefusedError):
+        if boundary == "tcp":
+            return _TCP_REFUSED.diagnostic()
+        if boundary == "mqtt":
+            return _MQTT_REFUSED.diagnostic()
+        return None
+    if isinstance(exc, PermissionError):
+        if boundary == "corebluetooth":
+            return _BLUETOOTH_PERMISSION.diagnostic()
+        if boundary == "corewlan":
+            return _WIFI_PERMISSION.diagnostic()
+        if boundary == "keychain":
+            return _KEYCHAIN_PERMISSION.diagnostic()
+        if boundary == "filesystem":
+            return _FILESYSTEM_PERMISSION.diagnostic()
+        return None
+    if isinstance(exc, TimeoutError):
+        if boundary == "corebluetooth":
+            return _BLUETOOTH_TIMEOUT.diagnostic()
+        if boundary == "corewlan":
+            return _WIFI_TIMEOUT.diagnostic()
+        if boundary == "tcp":
+            return _TCP_TIMEOUT.diagnostic()
+        if boundary == "mqtt":
+            return _MQTT_TIMEOUT.diagnostic()
+        return None
+    if boundary == "filesystem" and isinstance(exc, OSError):
+        return _LEDGER_UNAVAILABLE.diagnostic()
+
+    if boundary == "corebluetooth" and isinstance(
+        exc,
+        BleakBluetoothNotAvailableError,
+    ):
+        spec = {
+            BleakBluetoothNotAvailableReason.POWERED_OFF: _BLUETOOTH_POWERED_OFF,
+            BleakBluetoothNotAvailableReason.DENIED_BY_USER: _BLUETOOTH_PERMISSION,
+            BleakBluetoothNotAvailableReason.DENIED_BY_SYSTEM: _BLUETOOTH_PERMISSION,
+            BleakBluetoothNotAvailableReason.DENIED_BY_UNKNOWN: _BLUETOOTH_PERMISSION,
+            BleakBluetoothNotAvailableReason.NO_BLUETOOTH: _BLUETOOTH_UNAVAILABLE,
+            BleakBluetoothNotAvailableReason.NO_BLE_CENTRAL_ROLE: _BLUETOOTH_UNAVAILABLE,
+            BleakBluetoothNotAvailableReason.UNKNOWN: _BLUETOOTH_UNAVAILABLE,
+        }[exc.reason]
+        return spec.diagnostic()
+
+    # Older CoreBluetooth surfaces still expose several failures only as a
+    # human string. Match a narrow, reviewed vocabulary and export none of it.
+    message = str(exc).casefold()
+    if boundary == "corebluetooth":
+        if any(
+            fragment in message
+            for fragment in (
+                "not authorized",
+                "permission denied",
+                "access denied",
+                "bluetooth permission",
+            )
+        ):
+            return _BLUETOOTH_PERMISSION.diagnostic()
+        if any(
+            fragment in message
+            for fragment in ("powered off", "power off", "bluetooth is off", "poweredoff")
+        ):
+            return _BLUETOOTH_POWERED_OFF.diagnostic()
+        if any(
+            fragment in message
+            for fragment in (
+                "adapter unavailable",
+                "bluetooth unavailable",
+                "no bluetooth adapter",
+            )
+        ):
+            return _BLUETOOTH_UNAVAILABLE.diagnostic()
+    if boundary == "keychain" or "authorityloaderror" in type(exc).__name__.casefold():
+        return _KEYCHAIN_UNAVAILABLE.diagnostic()
+    return None
+
+
+def host_diagnostics_for_reasons(*reason_codes: object) -> tuple[HostFailureDiagnostic, ...]:
+    diagnostics: list[HostFailureDiagnostic] = []
+    seen: set[str] = set()
+    for reason_code in reason_codes:
+        diagnostic = host_diagnostic_for_reason(reason_code)
+        if diagnostic is None or diagnostic.code in seen:
+            continue
+        seen.add(diagnostic.code)
+        diagnostics.append(diagnostic)
+    return tuple(diagnostics)
+
+
+__all__ = [
+    "HostDiagnosticBoundary",
+    "HostDiagnosticCode",
+    "HostFailureDiagnostic",
+    "host_diagnostic_for_exception",
+    "host_diagnostic_for_reason",
+    "host_diagnostics_for_reasons",
+]

src/k1link/device_plugins/xgrids_k1/legacy_api.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 import asyncio
 from typing import Any
 
-from fastapi import APIRouter, HTTPException, WebSocket, WebSocketDisconnect
+from fastapi import APIRouter, Header, HTTPException, WebSocket, WebSocketDisconnect
 
 from k1link.device_plugins.xgrids_k1.facade import (
     ACTION_DISCOVERY_SCAN,
@@ -36,23 +36,41 @@ def build_xgrids_k1_legacy_router(runtime: DevicePluginRuntimeTransport) -> APIR
         return await invoke_device_plugin_runtime(runtime, ACTION_STATE_READ, {})
 
     @router.post("/api/ble/scan", deprecated=True)
-    async def scan_ble(request: BleScanRequest) -> dict[str, Any]:
+    async def scan_ble(
+        request: BleScanRequest,
+        expected_snapshot_runtime_id: str = Header(
+            ...,
+            alias="X-Mission-Core-Snapshot-Runtime-Id",
+        ),
+    ) -> dict[str, Any]:
         try:
             return await invoke_device_plugin_runtime(
                 runtime,
                 ACTION_DISCOVERY_SCAN,
-                request.model_dump(),
+                {
+                    **request.model_dump(),
+                    "expected_snapshot_runtime_id": expected_snapshot_runtime_id,
+                },
             )
         except (PluginExecutionError, ValueError) as exc:
             raise HTTPException(status_code=502, detail=f"Ошибка поиска BLE: {exc}") from exc
 
     @router.post("/api/connect", deprecated=True)
-    async def connect(request: ConnectRequest) -> dict[str, Any]:
+    async def connect(
+        request: ConnectRequest,
+        expected_snapshot_runtime_id: str = Header(
+            ...,
+            alias="X-Mission-Core-Snapshot-Runtime-Id",
+        ),
+    ) -> dict[str, Any]:
         try:
             return await invoke_device_plugin_runtime(
                 runtime,
                 ACTION_NETWORK_PROVISION,
-                request.model_dump(),
+                {
+                    **request.model_dump(),
+                    "expected_snapshot_runtime_id": expected_snapshot_runtime_id,
+                },
             )
         except (PluginExecutionError, ValueError) as exc:
             raise HTTPException(

src/k1link/device_plugins/xgrids_k1/live_perception_shadow.py

@@ -13,16 +13,42 @@ from uuid import uuid4
 
 from fastapi import APIRouter, WebSocket, WebSocketDisconnect
 
-from k1link.compute.live_perception import LivePerceptionIngress
+from k1link.compute.live_perception import (
+    LivePerceptionIngress,
+    LivePerceptionResultFrame,
+    decode_live_perception_result,
+)
 
 TOKEN_BYTES = 32
 TOKEN_FILE_NAME = "shadow-worker.token"
 
 
+def build_live_perception_result_receiver(
+    ingress: LivePerceptionIngress,
+    publish_frame: Callable[[LivePerceptionResultFrame], bool],
+) -> Callable[[bytes], bool]:
+    """Bind worker results to the exact active ingress session atomically."""
+
+    def receive(encoded: bytes) -> bool:
+        frame = decode_live_perception_result(encoded)
+        return ingress.admit_result(
+            session_id=frame.session_id,
+            session_generation=frame.session_generation,
+            receiver=lambda: publish_frame(frame),
+        )
+
+    return receive
+
+
 def ensure_live_shadow_token(repository_root: Path) -> tuple[Path, str]:
     """Load or create the private bearer used only through the SSH tunnel."""
 
-    token_root = repository_root.resolve() / ".runtime" / "live-perception"
+    configured_data_root = os.environ.get("MISSIONCORE_DATA_DIR", "").strip()
+    token_root = (
+        Path(configured_data_root).expanduser().resolve() / "live-perception"
+        if configured_data_root
+        else repository_root.resolve() / ".runtime" / "live-perception"
+    )
     token_root.mkdir(mode=0o700, parents=True, exist_ok=True)
     with suppress(OSError):
         token_root.chmod(0o700)
@@ -59,9 +85,7 @@ def build_live_perception_shadow_router(
 
     router = APIRouter(include_in_schema=False)
 
-    @router.websocket(
-        f"/api/v1/device-plugins/{plugin_id}/live-perception-shadow"
-    )
+    @router.websocket(f"/api/v1/device-plugins/{plugin_id}/live-perception-shadow")
     async def live_perception_shadow(websocket: WebSocket) -> None:
         authorization = websocket.headers.get("authorization", "")
         supplied = authorization.removeprefix("Bearer ")
@@ -103,13 +127,19 @@ def build_live_perception_shadow_router(
                         )
                         return
                     try:
-                        await asyncio.to_thread(result_receiver, result)
+                        accepted = await asyncio.to_thread(result_receiver, result)
                     except (RuntimeError, ValueError):
                         await websocket.close(
                             code=1008,
                             reason="Shadow result contract is invalid",
                         )
                         return
+                    if not accepted:
+                        await websocket.close(
+                            code=1008,
+                            reason="Shadow result session is stale",
+                        )
+                        return
                     client_event = asyncio.create_task(websocket.receive())
                 if ingress_event in completed:
                     event = ingress_event.result()

src/k1link/device_plugins/xgrids_k1/mqtt/capture.py

@@ -8,10 +8,11 @@ import os
 import re
 import stat
 import struct
+import threading
 import time
 from collections.abc import Callable, Iterator
 from contextlib import suppress
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import IO, Literal, TypedDict
 
@@ -29,6 +30,19 @@ REPORT_TOPICS: tuple[str, ...] = (
     "DeviceStatus",
 )
 
+# A fresh subscription proves only broker transport.  Recovery of the live
+# scene requires one non-retained point-cloud report from that exact MQTT
+# client to establish a candidate sequence fence. Pose, DeviceStatus and
+# heartbeat remain useful transport/control evidence, but none proves that the
+# visible acquisition cloud has resumed. Only the downstream post-publish
+# observer can confirm that candidate.
+RECOVERY_POINT_CLOUD_TOPICS = frozenset(
+    {
+        "RealtimePointcloud",
+        "lixel/application/report/lio_pcl",
+    }
+)
+
 DEFAULT_MAX_MESSAGE_BYTES = 64 * 1024 * 1024
 MAX_CONFIGURABLE_MESSAGE_BYTES = 256 * 1024 * 1024
 MAX_TOPIC_BYTES = 65_535
@@ -40,6 +54,8 @@ GROUP_COMMIT_MAX_BYTES = 4 * 1024 * 1024
 GROUP_COMMIT_MAX_MESSAGES = 32
 CAPTURE_CLOCK_FILENAME = "mqtt.timeline.json"
 CAPTURE_CLOCK_ORIGIN_FILENAME = "mqtt.timeline.origin.json"
+RECOVERY_GAPS_FILENAME = "mqtt.recovery.jsonl"
+RECOVERY_GAP_SCHEMA_VERSION = 1
 CAPTURE_CLOCK_ORIGIN_SCHEMA_VERSION = 1
 CAPTURE_CLOCK_SEALED_PATTERN = re.compile(r"^mqtt\.timeline\.session-[a-f0-9]{64}\.json$")
 CAPTURE_CLOCK_SCHEMA_VERSION = 1
@@ -61,14 +77,18 @@ StopReason = Literal[
     "message_too_large",
     "connection_failed",
     "connection_lost",
+    "recovery_standby",
     "subscription_failed",
     "capture_error",
 ]
+RecoveryDecision = Literal["retry", "resume", "standby", "fault", "blocked"]
+RecoveryGapOutcome = Literal["recovered", "standby", "fault", "blocked", "interrupted"]
 
 
 class ArtifactPaths(TypedDict):
     raw: str
     metadata_jsonl: str
+    recovery_gaps_jsonl: str
     capture_clock_origin: str
     capture_clock: str
     summary: str
@@ -77,6 +97,7 @@ class ArtifactPaths(TypedDict):
 class ArtifactHashes(TypedDict):
     raw_sha256: str
     metadata_jsonl_sha256: str
+    recovery_gaps_jsonl_sha256: str
     capture_clock_origin_sha256: str
     capture_clock_sha256: str
 
@@ -87,6 +108,17 @@ class RawFormat(TypedDict):
     frame_layout: str
 
 
+class RecoveryGapSummary(TypedDict):
+    gap_index: int
+    started_at_utc: str
+    started_monotonic_ns: int
+    ended_at_utc: str
+    ended_monotonic_ns: int
+    duration_seconds: float
+    recovery_attempt: int
+    outcome: RecoveryGapOutcome
+
+
 class CaptureSummary(TypedDict):
     schema_version: int
     created_at_utc: str
@@ -99,6 +131,11 @@ class CaptureSummary(TypedDict):
     subscription_qos: int
     clean_session: bool
     reconnect_enabled: bool
+    recovery_attempts: int
+    successful_recoveries: int
+    recovery_point_cloud_candidates: int
+    recovery_blocked: bool
+    recovery_gaps: list[RecoveryGapSummary]
     publishing_enabled: bool
     subscriptions: list[str]
     requested_duration_seconds: float | None
@@ -195,6 +232,13 @@ class _CaptureState:
     subscription_mid: int | None = None
     stop_reason: StopReason = "capture_error"
     error: str | None = None
+    connection_lost: bool = False
+    connection_lost_message: str | None = None
+    recovery_attempts: int = 0
+    successful_recoveries: int = 0
+    recovery_point_cloud_candidates: int = 0
+    recovery_blocked: bool = False
+    recovery_gaps: list[RecoveryGapSummary] = field(default_factory=list)
 
 
 class _CaptureWriter:
@@ -202,6 +246,7 @@ class _CaptureWriter:
         self.out_dir = out_dir.expanduser().resolve()
         self.raw_path = self.out_dir / "mqtt.raw.k1mqtt"
         self.metadata_path = self.out_dir / "mqtt.metadata.jsonl"
+        self.recovery_gaps_path = self.out_dir / RECOVERY_GAPS_FILENAME
         self.capture_clock_origin_path = self.out_dir / CAPTURE_CLOCK_ORIGIN_FILENAME
         self.capture_clock_path = self.out_dir / CAPTURE_CLOCK_FILENAME
         self.summary_path = self.out_dir / "mqtt.summary.json"
@@ -212,6 +257,8 @@ class _CaptureWriter:
         self.topic_counts: dict[str, int] = {}
         self._raw: IO[bytes] | None = None
         self._metadata: IO[str] | None = None
+        self._recovery_gaps: IO[str] | None = None
+        self._recovery_gaps_lock = threading.Lock()
         self._pending_metadata: list[str] = []
         self._pending_raw_bytes = 0
         self._last_commit_monotonic = time.monotonic()
@@ -223,6 +270,7 @@ class _CaptureWriter:
         artifact_paths = (
             self.raw_path,
             self.metadata_path,
+            self.recovery_gaps_path,
             self.capture_clock_origin_path,
             self.capture_clock_path,
             self.summary_path,
@@ -236,6 +284,7 @@ class _CaptureWriter:
             self._raw = _open_binary_exclusive(self.raw_path)
             self._raw.write(RAW_MAGIC)
             self._metadata = _open_text_exclusive(self.metadata_path)
+            self._recovery_gaps = _open_text_exclusive(self.recovery_gaps_path)
             _fsync_directory(self.out_dir)
             # This is the earliest durable point from which the capture can
             # accept evidence. The camera is armed only after this writer is
@@ -364,9 +413,36 @@ class _CaptureWriter:
                 except OSError as exc:
                     if first_error is None:
                         first_error = exc
+        with self._recovery_gaps_lock:
+            recovery_stream = self._recovery_gaps
+            if recovery_stream is not None and not recovery_stream.closed:
+                try:
+                    recovery_stream.flush()
+                    os.fsync(recovery_stream.fileno())
+                except OSError as exc:
+                    if first_error is None:
+                        first_error = exc
+                finally:
+                    try:
+                        recovery_stream.close()
+                    except OSError as exc:
+                        if first_error is None:
+                            first_error = exc
         if first_error is not None:
             raise first_error
 
+    def record_recovery_gap_event(self, record: dict[str, object]) -> None:
+        """Append and fsync one recovery boundary without touching frame metadata."""
+
+        payload = json.dumps(record, ensure_ascii=False, separators=(",", ":")) + "\n"
+        with self._recovery_gaps_lock:
+            stream = self._recovery_gaps
+            if stream is None or stream.closed:
+                raise RuntimeError("recovery gap journal is not open")
+            stream.write(payload)
+            stream.flush()
+            os.fsync(stream.fileno())
+
     def finalize_capture_clock(self) -> CaptureClockEnvelope:
         """Publish the exact capture envelope after every producer is sealed."""
 
@@ -748,6 +824,20 @@ def seal_capture_clock(capture_root: Path) -> CaptureClockEnvelope:
         raise CaptureError(f"could not seal session capture clock: {exc}") from exc
 
 
+def _recovery_backoff_seconds(attempt: int) -> float:
+    """Return the canonical capped delay without evaluating an unbounded power."""
+
+    if attempt <= 1:
+        return 0.5
+    if attempt == 2:
+        return 1.0
+    if attempt == 3:
+        return 2.0
+    if attempt == 4:
+        return 4.0
+    return 5.0
+
+
 def capture_mqtt(
     host: str,
     out_dir: Path,
@@ -759,9 +849,23 @@ def capture_mqtt(
     on_ready: Callable[[], None] | None = None,
     on_message_recorded: Callable[[CapturedMqttMessage], None] | None = None,
     should_stop: Callable[[], bool] | None = None,
+    on_connection_lost: Callable[[str], None] | None = None,
+    consume_connection_recovery_request: Callable[[], str | None] | None = None,
+    recover_connection: Callable[[int], RecoveryDecision] | None = None,
+    on_recovery_point_cloud_candidate: Callable[[int, int], None] | None = None,
+    on_recovery_confirmer_ready: Callable[[Callable[[int], bool]], None] | None = None,
     _client_factory: Callable[[], mqtt.Client] | None = None,
 ) -> CaptureSummary:
-    """Capture the fixed K1 report subscriptions once, without publishing or reconnecting."""
+    """Capture fixed K1 reports with an optional externally fenced reconnect.
+
+    The capture layer never decides that an endpoint is still the same K1.  A
+    caller may supply ``recover_connection`` only when a higher-level owner can
+    revalidate the exact route, DeviceInfo identity and physical SCANNING
+    lineage.  ``consume_connection_recovery_request`` lets that same owner wake
+    this existing capture before the MQTT keepalive notices a short host-path
+    outage.  The callback returns the only admitted next step; this function
+    itself never publishes, scans BLE, changes Wi-Fi, or retries START/STOP.
+    """
     target_ipv4 = validate_private_ipv4(host)
     if not 1 <= port <= 65535:
         raise ValueError("port must be between 1 and 65535")
@@ -774,16 +878,18 @@ def capture_mqtt(
             f"max_message_bytes must be between 1 and {MAX_CONFIGURABLE_MESSAGE_BYTES}"
         )
 
-    client = (
-        _client_factory()
-        if _client_factory is not None
-        else mqtt.Client(
-            callback_api_version=CallbackAPIVersion.VERSION2,
-            clean_session=True,
-            protocol=mqtt.MQTTv311,
-            reconnect_on_failure=False,
+    def make_client() -> mqtt.Client:
+        return (
+            _client_factory()
+            if _client_factory is not None
+            else mqtt.Client(
+                callback_api_version=CallbackAPIVersion.VERSION2,
+                clean_session=True,
+                protocol=mqtt.MQTTv311,
+                reconnect_on_failure=False,
+            )
         )
-    )
+
     writer = _CaptureWriter(out_dir, max_message_bytes)
     try:
         writer.open()
@@ -798,12 +904,119 @@ def capture_mqtt(
     operation_started = time.monotonic()
     capture_started: float | None = None
     failure: CaptureError | None = None
+    disconnect_expected = False
+    recovering = False
+    pending_recovery_attempt: int | None = None
+    pending_recovery_confirmation_attempt: int | None = None
+    active_recovery_gap: tuple[int, str, int] | None = None
+    recovery_lock = threading.Lock()
+    active_client: mqtt.Client | None = None
 
     def fail(reason: StopReason, message: str) -> None:
         if state.error is None:
             state.stop_reason = reason
             state.error = message
 
+    def finish_recovery_gap_locked(
+        outcome: RecoveryGapOutcome,
+        attempt: int,
+    ) -> bool:
+        nonlocal active_recovery_gap, pending_recovery_confirmation_attempt
+
+        gap = active_recovery_gap
+        if gap is None:
+            return False
+        gap_index, started_at_utc, started_monotonic_ns = gap
+        ended_at_utc = utc_now_iso()
+        ended_monotonic_ns = time.monotonic_ns()
+        summary: RecoveryGapSummary = {
+            "gap_index": gap_index,
+            "started_at_utc": started_at_utc,
+            "started_monotonic_ns": started_monotonic_ns,
+            "ended_at_utc": ended_at_utc,
+            "ended_monotonic_ns": ended_monotonic_ns,
+            "duration_seconds": round(
+                max(ended_monotonic_ns - started_monotonic_ns, 0) / 1_000_000_000,
+                9,
+            ),
+            "recovery_attempt": attempt,
+            "outcome": outcome,
+        }
+        try:
+            writer.record_recovery_gap_event(
+                {
+                    "schema_version": RECOVERY_GAP_SCHEMA_VERSION,
+                    "record_type": "recovery_gap_ended",
+                    **summary,
+                }
+            )
+        except (OSError, RuntimeError, ValueError) as exc:
+            fail(
+                "capture_error",
+                f"recovery gap journal failed: {type(exc).__name__}: {exc}",
+            )
+            return False
+        state.recovery_gaps.append(summary)
+        if outcome == "recovered":
+            state.successful_recoveries += 1
+        active_recovery_gap = None
+        pending_recovery_confirmation_attempt = None
+        return True
+
+    def begin_recovery_gap() -> bool:
+        nonlocal active_recovery_gap, pending_recovery_confirmation_attempt
+
+        with recovery_lock:
+            # A raw candidate from a client that has already failed can no
+            # longer prove the still-live scene. Keep the original outage open
+            # and fence that late downstream callback.
+            pending_recovery_confirmation_attempt = None
+            if active_recovery_gap is not None:
+                return True
+            gap_index = len(state.recovery_gaps) + 1
+            started_at_utc = utc_now_iso()
+            started_monotonic_ns = time.monotonic_ns()
+            try:
+                writer.record_recovery_gap_event(
+                    {
+                        "schema_version": RECOVERY_GAP_SCHEMA_VERSION,
+                        "record_type": "recovery_gap_started",
+                        "gap_index": gap_index,
+                        "started_at_utc": started_at_utc,
+                        "started_monotonic_ns": started_monotonic_ns,
+                    }
+                )
+            except (OSError, RuntimeError, ValueError) as exc:
+                fail(
+                    "capture_error",
+                    f"recovery gap journal failed: {type(exc).__name__}: {exc}",
+                )
+                return False
+            active_recovery_gap = (gap_index, started_at_utc, started_monotonic_ns)
+            return True
+
+    def confirm_recovery(attempt: int) -> bool:
+        """Consume one exact post-publication proof for the current open gap."""
+
+        with recovery_lock:
+            if (
+                isinstance(attempt, bool)
+                or attempt < 1
+                or attempt != pending_recovery_confirmation_attempt
+                or active_recovery_gap is None
+            ):
+                return False
+            return finish_recovery_gap_locked("recovered", attempt)
+
+    def fail_transport_handshake(reason: StopReason, message: str) -> None:
+        if recovering:
+            state.connected = False
+            state.subscribed = False
+            state.connection_lost = True
+            state.connection_lost_message = message
+            return
+        fail(reason, message)
+
     def on_connect(
         callback_client: mqtt.Client,
         _userdata: object,
@@ -811,129 +1024,347 @@ def capture_mqtt(
         reason_code: ReasonCode,
         _properties: Properties | None,
     ) -> None:
+        if callback_client is not active_client:
+            return
         if reason_code.is_failure:
-            fail("connection_failed", f"broker rejected connection: {reason_code}")
+            fail_transport_handshake(
+                "connection_failed",
+                f"broker rejected connection: {reason_code}",
+            )
             return
         state.connected = True
         try:
             result, mid = callback_client.subscribe([(topic, 0) for topic in REPORT_TOPICS])
         except (OSError, RuntimeError, ValueError) as exc:
-            fail(
+            fail_transport_handshake(
                 "subscription_failed",
                 f"subscribe failed: {type(exc).__name__}: {exc}",
             )
             return
         if result != mqtt.MQTT_ERR_SUCCESS or mid is None:
-            fail("subscription_failed", f"subscribe failed: {mqtt.error_string(result)}")
+            fail_transport_handshake(
+                "subscription_failed",
+                f"subscribe failed: {mqtt.error_string(result)}",
+            )
             return
         state.subscription_mid = mid
 
     def on_subscribe(
-        _callback_client: mqtt.Client,
+        callback_client: mqtt.Client,
         _userdata: object,
         mid: int,
         reason_codes: list[ReasonCode],
         _properties: Properties | None,
     ) -> None:
+        if callback_client is not active_client:
+            return
         if mid != state.subscription_mid:
-            fail("subscription_failed", f"unexpected SUBACK message id: {mid}")
+            fail_transport_handshake(
+                "subscription_failed",
+                f"unexpected SUBACK message id: {mid}",
+            )
             return
         if len(reason_codes) != len(REPORT_TOPICS) or any(
             reason_code.is_failure for reason_code in reason_codes
         ):
-            fail("subscription_failed", "broker rejected one or more fixed subscriptions")
+            fail_transport_handshake(
+                "subscription_failed",
+                "broker rejected one or more fixed subscriptions",
+            )
             return
         state.subscribed = True
 
     def on_message(
-        _callback_client: mqtt.Client,
+        callback_client: mqtt.Client,
         _userdata: object,
         message: mqtt.MQTTMessage,
     ) -> None:
-        if state.error is not None:
+        nonlocal pending_recovery_attempt, pending_recovery_confirmation_attempt, recovering
+
+        if callback_client is not active_client or state.error is not None:
             return
         try:
             recorded = writer.record(message)
         except MessageTooLargeError as exc:
             fail("message_too_large", str(exc))
+            return
         except (OSError, RuntimeError, ValueError) as exc:
             fail("capture_error", f"artifact write failed: {type(exc).__name__}: {exc}")
             return
+        recovery_candidate: tuple[int, int] | None = None
+        if (
+            pending_recovery_attempt is not None
+            and not recorded.retain
+            and recorded.topic in RECOVERY_POINT_CLOUD_TOPICS
+        ):
+            recovered_attempt = pending_recovery_attempt
+            pending_recovery_attempt = None
+            recovering = False
+            with recovery_lock:
+                if active_recovery_gap is not None:
+                    pending_recovery_confirmation_attempt = recovered_attempt
+                state.recovery_point_cloud_candidates += 1
+            recovery_candidate = (recovered_attempt, recorded.sequence)
+        # Arm the provisional sequence fence after the raw frame is durable but
+        # before the preview queue can publish it on another thread. This is
+        # deliberately not a recovery-success edge: only a later normalized,
+        # non-empty Rerun publication may consume the candidate.
+        if recovery_candidate is not None and on_recovery_point_cloud_candidate is not None:
+            try:
+                on_recovery_point_cloud_candidate(*recovery_candidate)
+            except (OSError, RuntimeError, ValueError) as exc:
+                fail(
+                    "capture_error",
+                    f"recovery candidate callback failed: {type(exc).__name__}: {exc}",
+                )
+                return
         if on_message_recorded is not None:
             try:
                 on_message_recorded(recorded)
             except (OSError, RuntimeError, ValueError) as exc:
                 fail("capture_error", f"preview callback failed: {type(exc).__name__}: {exc}")
+                return
 
     def on_disconnect(
-        _callback_client: mqtt.Client,
+        callback_client: mqtt.Client,
         _userdata: object,
         _flags: mqtt.DisconnectFlags,
         reason_code: ReasonCode,
         _properties: Properties | None,
     ) -> None:
-        if not state.stopping:
-            fail("connection_lost", f"broker connection ended: {reason_code}")
+        if callback_client is not active_client or state.stopping or disconnect_expected:
+            return
+        state.connected = False
+        state.subscribed = False
+        state.connection_lost = True
+        state.connection_lost_message = f"broker connection ended: {reason_code}"
 
-    client.on_connect = on_connect
-    client.on_subscribe = on_subscribe
-    client.on_message = on_message
-    client.on_disconnect = on_disconnect
+    def externally_stopped_or_elapsed() -> bool:
+        now = time.monotonic()
+        if should_stop is not None and should_stop():
+            state.stop_reason = "external_stop"
+            return True
+        if (
+            capture_started is not None
+            and duration_seconds is not None
+            and now - capture_started >= duration_seconds
+        ):
+            state.stop_reason = "duration_elapsed"
+            return True
+        return False
+
+    def consume_owner_recovery_request() -> str | None:
+        if consume_connection_recovery_request is None:
+            return None
+        try:
+            reason = consume_connection_recovery_request()
+        except (OSError, RuntimeError, ValueError) as exc:
+            fail(
+                "capture_error",
+                "connection recovery request failed: "
+                f"{type(exc).__name__}: {exc}",
+            )
+            return None
+        if reason is None:
+            return None
+        if not isinstance(reason, str) or not reason.strip():
+            fail("capture_error", "connection recovery request returned an invalid reason")
+            return None
+        return reason.strip()
+
+    def wait_recovery_backoff(attempt: int) -> bool:
+        delay = _recovery_backoff_seconds(attempt)
+        deadline = time.monotonic() + delay
+        while time.monotonic() < deadline:
+            if externally_stopped_or_elapsed():
+                return False
+            time.sleep(min(0.1, max(deadline - time.monotonic(), 0.0)))
+        return True
+
+    if on_recovery_confirmer_ready is not None:
+        try:
+            on_recovery_confirmer_ready(confirm_recovery)
+        except (OSError, RuntimeError, ValueError) as exc:
+            fail(
+                "capture_error",
+                f"recovery confirmer registration failed: {type(exc).__name__}: {exc}",
+            )
 
-    connect_attempted = False
     try:
-        connect_attempted = True
-        connect_result = client.connect(
-            target_ipv4,
-            port=port,
-            keepalive=KEEPALIVE_SECONDS,
-        )
-        if connect_result != mqtt.MQTT_ERR_SUCCESS:
-            fail("connection_failed", f"connect failed: {mqtt.error_string(connect_result)}")
-
         while state.error is None:
-            now = time.monotonic()
-            if should_stop is not None and should_stop():
-                state.stop_reason = "external_stop"
+            if recovering and externally_stopped_or_elapsed():
                 break
-            if state.subscribed and capture_started is None:
-                capture_started = now
-                if on_ready is not None:
-                    on_ready()
-            if (
-                capture_started is not None
-                and duration_seconds is not None
-                and now - capture_started >= duration_seconds
-            ):
-                state.stop_reason = "duration_elapsed"
-                break
-            if capture_started is None and now - operation_started >= CONNECT_TIMEOUT_SECONDS:
-                fail("connection_failed", "timed out waiting for CONNACK/SUBACK")
-                break
-
-            loop_result = client.loop(timeout=LOOP_INTERVAL_SECONDS)
+            state.connected = False
+            state.subscribed = False
+            state.subscription_mid = None
+            state.connection_lost = False
+            state.connection_lost_message = None
+            client = make_client()
+            active_client = client
+            client.on_connect = on_connect
+            client.on_subscribe = on_subscribe
+            client.on_message = on_message
+            client.on_disconnect = on_disconnect
+            connect_attempted = False
             try:
-                writer.maybe_commit()
-            except OSError as exc:
-                fail("capture_error", f"group commit failed: {type(exc).__name__}: {exc}")
+                connect_attempted = True
+                connect_result = client.connect(
+                    target_ipv4,
+                    port=port,
+                    keepalive=KEEPALIVE_SECONDS,
+                )
+                if connect_result != mqtt.MQTT_ERR_SUCCESS:
+                    message = f"connect failed: {mqtt.error_string(connect_result)}"
+                    if recovering:
+                        state.connection_lost = True
+                        state.connection_lost_message = message
+                    else:
+                        fail("connection_failed", message)
+
+                connect_started = time.monotonic()
+                while state.error is None and not state.connection_lost:
+                    now = time.monotonic()
+                    if externally_stopped_or_elapsed():
+                        break
+                    if state.subscribed and capture_started is None:
+                        capture_started = now
+                        if on_ready is not None:
+                            on_ready()
+                    if not state.subscribed and now - connect_started >= CONNECT_TIMEOUT_SECONDS:
+                        message = "timed out waiting for CONNACK/SUBACK"
+                        if recovering:
+                            state.connection_lost = True
+                            state.connection_lost_message = message
+                        else:
+                            fail("connection_failed", message)
+                        break
+
+                    loop_result = client.loop(timeout=LOOP_INTERVAL_SECONDS)
+                    owner_recovery_reason = consume_owner_recovery_request()
+                    try:
+                        writer.maybe_commit()
+                    except OSError as exc:
+                        fail(
+                            "capture_error",
+                            f"group commit failed: {type(exc).__name__}: {exc}",
+                        )
+                        break
+                    if loop_result != mqtt.MQTT_ERR_SUCCESS and state.error is None:
+                        state.connected = False
+                        state.subscribed = False
+                        state.connection_lost = True
+                        state.connection_lost_message = (
+                            f"MQTT network loop failed: {mqtt.error_string(loop_result)}"
+                        )
+                    if owner_recovery_reason is not None and not state.connection_lost:
+                        state.connected = False
+                        state.subscribed = False
+                        state.connection_lost = True
+                        state.connection_lost_message = (
+                            "guarded recovery requested by the connection owner: "
+                            f"{owner_recovery_reason}"
+                        )
+            except KeyboardInterrupt:
+                state.stop_reason = "keyboard_interrupt"
+            except (OSError, RuntimeError, ValueError) as exc:
+                message = f"MQTT capture failed: {type(exc).__name__}: {exc}"
+                if recovering:
+                    state.connection_lost = True
+                    state.connection_lost_message = message
+                else:
+                    fail("connection_failed", message)
+            finally:
+                disconnect_expected = True
+                active_client = None
+                if connect_attempted:
+                    try:
+                        client.disconnect()
+                    except (OSError, RuntimeError, ValueError) as exc:
+                        if state.error is None and not state.connection_lost:
+                            fail(
+                                "capture_error",
+                                f"disconnect failed: {type(exc).__name__}: {exc}",
+                            )
+                disconnect_expected = False
+
+            if (
+                state.stop_reason
+                in {
+                    "external_stop",
+                    "duration_elapsed",
+                    "keyboard_interrupt",
+                }
+                or state.error is not None
+            ):
                 break
-            if loop_result != mqtt.MQTT_ERR_SUCCESS and state.error is None:
+            if not state.connection_lost:
+                break
+            # A guarded ``resume`` authorizes exactly one fresh MQTT client.
+            # If that client never delivers a fresh point-cloud report, consume the
+            # authorization before inspecting the device again; a later
+            # READY/SCAN_OVER must be able to end recovery without another
+            # data-plane connect.
+            pending_recovery_attempt = None
+            if not begin_recovery_gap():
+                break
+            if recover_connection is None:
                 fail(
                     "connection_lost",
-                    f"MQTT network loop failed: {mqtt.error_string(loop_result)}",
+                    state.connection_lost_message or "MQTT connection was lost",
                 )
-    except KeyboardInterrupt:
-        state.stop_reason = "keyboard_interrupt"
-    except (OSError, RuntimeError, ValueError) as exc:
-        fail("connection_failed", f"MQTT capture failed: {type(exc).__name__}: {exc}")
+                break
+            if not recovering and on_connection_lost is not None:
+                on_connection_lost(state.connection_lost_message or "MQTT connection was lost")
+            recovering = True
+
+            while state.error is None:
+                if externally_stopped_or_elapsed():
+                    break
+                state.recovery_attempts += 1
+                attempt = state.recovery_attempts
+                try:
+                    decision = recover_connection(attempt)
+                except (OSError, RuntimeError, ValueError):
+                    decision = "retry"
+                if decision == "resume":
+                    pending_recovery_attempt = attempt
+                    break
+                if decision == "standby":
+                    state.stop_reason = "recovery_standby"
+                    break
+                if decision == "fault":
+                    fail(
+                        "connection_lost",
+                        "exact active-stream recovery rejected the remote state",
+                    )
+                    break
+                if decision == "blocked":
+                    state.recovery_blocked = True
+                    while not externally_stopped_or_elapsed():
+                        time.sleep(LOOP_INTERVAL_SECONDS)
+                    break
+                if decision != "retry":
+                    fail("capture_error", "invalid active-stream recovery decision")
+                    break
+                if not wait_recovery_backoff(attempt):
+                    break
+
+            if pending_recovery_attempt is not None:
+                continue
+            break
     finally:
         state.stopping = True
-        if connect_attempted:
-            try:
-                client.disconnect()
-            except (OSError, RuntimeError, ValueError) as exc:
-                if state.error is None:
-                    fail("capture_error", f"disconnect failed: {type(exc).__name__}: {exc}")
+        with recovery_lock:
+            if active_recovery_gap is not None:
+                if state.stop_reason == "recovery_standby":
+                    gap_outcome: RecoveryGapOutcome = "standby"
+                elif state.recovery_blocked:
+                    gap_outcome = "blocked"
+                elif state.error is not None:
+                    gap_outcome = "fault"
+                else:
+                    gap_outcome = "interrupted"
+                finish_recovery_gap_locked(gap_outcome, state.recovery_attempts)
         try:
             writer.close()
         except OSError as exc:
@@ -958,6 +1389,7 @@ def capture_mqtt(
         created_at_utc=created_at_utc,
         capture_clock=capture_clock,
         state=state,
+        reconnect_enabled=recover_connection is not None,
     )
     try:
         _write_summary_exclusive(writer.summary_path, summary)
@@ -983,6 +1415,7 @@ def _build_summary(
     created_at_utc: str,
     capture_clock: CaptureClockEnvelope,
     state: _CaptureState,
+    reconnect_enabled: bool,
 ) -> CaptureSummary:
     capture_clock_origin = read_capture_clock_origin(writer.capture_clock_origin_path)
     return {
@@ -996,7 +1429,12 @@ def _build_summary(
         "mqtt_protocol": "3.1.1",
         "subscription_qos": 0,
         "clean_session": True,
-        "reconnect_enabled": False,
+        "reconnect_enabled": reconnect_enabled,
+        "recovery_attempts": state.recovery_attempts,
+        "successful_recoveries": state.successful_recoveries,
+        "recovery_point_cloud_candidates": state.recovery_point_cloud_candidates,
+        "recovery_blocked": state.recovery_blocked,
+        "recovery_gaps": list(state.recovery_gaps),
         "publishing_enabled": False,
         "subscriptions": list(REPORT_TOPICS),
         "requested_duration_seconds": duration_seconds,
@@ -1021,6 +1459,7 @@ def _build_summary(
         "artifacts": {
             "raw": writer.raw_path.name,
             "metadata_jsonl": writer.metadata_path.name,
+            "recovery_gaps_jsonl": writer.recovery_gaps_path.name,
             "capture_clock_origin": writer.capture_clock_origin_path.name,
             "capture_clock": writer.capture_clock_path.name,
             "summary": writer.summary_path.name,
@@ -1028,6 +1467,7 @@ def _build_summary(
         "artifact_hashes": {
             "raw_sha256": _sha256_file(writer.raw_path),
             "metadata_jsonl_sha256": _sha256_file(writer.metadata_path),
+            "recovery_gaps_jsonl_sha256": _sha256_file(writer.recovery_gaps_path),
             "capture_clock_origin_sha256": capture_clock_origin.artifact_sha256,
             "capture_clock_sha256": capture_clock.artifact_sha256,
         },

src/k1link/device_plugins/xgrids_k1/network_mutation_ledger.py

@@ -0,0 +1,958 @@
+from __future__ import annotations
+
+import fcntl
+import ipaddress
+import json
+import os
+import re
+import stat
+import tempfile
+import threading
+from collections.abc import Callable, Iterator, Mapping
+from contextlib import contextmanager
+from dataclasses import dataclass, replace
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import IO, Literal, cast
+
+from k1link.sessions.store import resolve_missioncore_data_dir
+
+NETWORK_MUTATION_LEDGER_SCHEMA = "missioncore.xgrids-k1-network-mutation/v2"
+NETWORK_MUTATION_LEDGER_LEGACY_SCHEMA = "missioncore.xgrids-k1-network-mutation/v1"
+NETWORK_MUTATION_LEDGER_FILENAME = "network-mutation.json"
+NETWORK_MUTATION_LEDGER_LOCK_FILENAME = ".network-mutation.lock"
+NETWORK_MUTATION_LEDGER_MAX_BYTES = 64 * 1024
+
+NetworkConnectionMode = Literal["bridge", "quick-connect", "direct-connect"]
+NetworkMutationStage = Literal["prepared", "dispatching", "observing", "resolved"]
+NetworkMutationWriteMode = Literal["with_response", "without_response"]
+NetworkMutationResolution = Literal[
+    "not-dispatched",
+    "target-observed",
+    "interrupted",
+    "superseded",
+]
+NetworkMutationLedgerStatus = Literal["empty", "unresolved", "resolved", "corrupt"]
+
+_CONNECTION_MODES = frozenset({"bridge", "quick-connect", "direct-connect"})
+_STAGES = frozenset({"prepared", "dispatching", "observing", "resolved"})
+_WRITE_MODES = frozenset({"with_response", "without_response"})
+_RESOLUTIONS = frozenset(
+    {"not-dispatched", "target-observed", "interrupted", "superseded"}
+)
+_SAFE_IDENTIFIER = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:+-]{0,159}$")
+_SAFE_STATUS_MODE = re.compile(r"^[A-Z][A-Z0-9_-]{0,31}$")
+
+
+class NetworkMutationLedgerError(RuntimeError):
+    """Base error for the durable K1 network-mutation fence."""
+
+    reason_code = "network-mutation-ledger-error"
+
+
+class NetworkMutationBlocked(NetworkMutationLedgerError):
+    """A prior durable record prevents admission of another device write."""
+
+    reason_code = "network-mutation-reconciliation-required"
+
+
+class NetworkMutationLedgerCorrupt(NetworkMutationBlocked):
+    """The durable fence cannot be trusted and therefore fails closed."""
+
+    reason_code = "network-mutation-ledger-corrupt"
+
+
+class NetworkMutationTransitionError(NetworkMutationLedgerError):
+    """A caller attempted an invalid ledger state transition."""
+
+    reason_code = "network-mutation-ledger-transition-invalid"
+
+
+@dataclass(frozen=True, slots=True)
+class NetworkStatusEvidence:
+    """The bounded, non-secret subset of one decoded K1 7f02 status."""
+
+    mode: str | None
+    ipv4: str | None
+    status_code: int
+    reserved: int | None
+
+    def __post_init__(self) -> None:
+        if self.mode is not None and _SAFE_STATUS_MODE.fullmatch(self.mode) is None:
+            raise ValueError("network status mode is outside the secret-free schema")
+        if self.ipv4 is not None:
+            try:
+                parsed = ipaddress.ip_address(self.ipv4)
+            except ValueError as exc:
+                raise ValueError("network status address must be an IPv4 address") from exc
+            if not isinstance(parsed, ipaddress.IPv4Address) or str(parsed) != self.ipv4:
+                raise ValueError("network status address must be canonical IPv4")
+        _validate_byte(self.status_code, field_name="status_code")
+        if self.reserved is not None:
+            _validate_byte(self.reserved, field_name="reserved")
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "mode": self.mode,
+            "ipv4": self.ipv4,
+            "status_code": self.status_code,
+            "reserved": self.reserved,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class PreviousConnectionEvidence:
+    """Last admitted topology retained as evidence, never as live reachability."""
+
+    transport_ref: str
+    mode: NetworkConnectionMode
+    ipv4: str | None
+    device_session_id: str | None
+
+    def __post_init__(self) -> None:
+        _validate_identifier(self.transport_ref, field_name="transport_ref")
+        _validate_connection_mode(self.mode)
+        if self.ipv4 is not None:
+            try:
+                parsed = ipaddress.ip_address(self.ipv4)
+            except ValueError as exc:
+                raise ValueError("previous connection address must be an IPv4 address") from exc
+            if not isinstance(parsed, ipaddress.IPv4Address) or str(parsed) != self.ipv4:
+                raise ValueError("previous connection address must be canonical IPv4")
+        if self.device_session_id is not None:
+            _validate_identifier(self.device_session_id, field_name="device_session_id")
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "transport_ref": self.transport_ref,
+            "mode": self.mode,
+            "ipv4": self.ipv4,
+            "device_session_id": self.device_session_id,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class NetworkMutationRecord:
+    schema_version: str
+    revision: int
+    operation_id: str
+    transport_ref: str
+    intended_mode: NetworkConnectionMode
+    stage: NetworkMutationStage
+    write_mode: NetworkMutationWriteMode
+    baseline_status: NetworkStatusEvidence
+    previous_connection: PreviousConnectionEvidence | None
+    write_confirmed: bool | None
+    last_observation: NetworkStatusEvidence | None
+    resolution: NetworkMutationResolution | None
+    created_at_utc: str
+    updated_at_utc: str
+
+    @property
+    def unresolved(self) -> bool:
+        return self.stage != "resolved"
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "schema_version": self.schema_version,
+            "revision": self.revision,
+            "operation_id": self.operation_id,
+            "transport_ref": self.transport_ref,
+            "intended_mode": self.intended_mode,
+            "stage": self.stage,
+            "write_mode": self.write_mode,
+            "baseline_status": self.baseline_status.as_dict(),
+            "previous_connection": (
+                self.previous_connection.as_dict() if self.previous_connection is not None else None
+            ),
+            "write_confirmed": self.write_confirmed,
+            "last_observation": (
+                self.last_observation.as_dict() if self.last_observation is not None else None
+            ),
+            "resolution": self.resolution,
+            "created_at_utc": self.created_at_utc,
+            "updated_at_utc": self.updated_at_utc,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class NetworkMutationLedgerSnapshot:
+    status: NetworkMutationLedgerStatus
+    record: NetworkMutationRecord | None
+    reason_code: str | None
+
+    @property
+    def mutation_allowed(self) -> bool:
+        return self.status in {"empty", "resolved"}
+
+
+class NetworkMutationLedger:
+    """One durable, secret-free fence around K1 network side effects.
+
+    The ledger is intentionally independent from the BLE helpers. Integration
+    writes ``dispatching`` durably immediately before entering
+    ``write_gatt_char``. A process crash can therefore create a conservative
+    false-positive fence, but can never silently authorize a second write.
+    """
+
+    def __init__(
+        self,
+        repository_root: Path,
+        *,
+        clock: Callable[[], datetime] | None = None,
+    ) -> None:
+        data_dir = resolve_missioncore_data_dir(repository_root)
+        self.path = data_dir / "xgrids-k1" / NETWORK_MUTATION_LEDGER_FILENAME
+        self._process_lock_path = data_dir / "xgrids-k1" / NETWORK_MUTATION_LEDGER_LOCK_FILENAME
+        self._data_dir = data_dir
+        self._clock = clock or (lambda: datetime.now(UTC))
+        self._lock = threading.RLock()
+        self._record: NetworkMutationRecord | None = None
+        self._corrupt = False
+        with self._lock, self._process_lock_locked():
+            self._reload_locked()
+
+    def snapshot(self) -> NetworkMutationLedgerSnapshot:
+        with self._lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt:
+                return NetworkMutationLedgerSnapshot(
+                    status="corrupt",
+                    record=None,
+                    reason_code=NetworkMutationLedgerCorrupt.reason_code,
+                )
+            if self._record is None:
+                return NetworkMutationLedgerSnapshot(status="empty", record=None, reason_code=None)
+            return NetworkMutationLedgerSnapshot(
+                status="unresolved" if self._record.unresolved else "resolved",
+                record=self._record,
+                reason_code=(
+                    NetworkMutationBlocked.reason_code if self._record.unresolved else None
+                ),
+            )
+
+    def require_mutation_allowed(self) -> None:
+        with self._lock, self._process_lock_locked():
+            self._reload_locked()
+            self._require_mutation_allowed_locked()
+
+    def prepare(
+        self,
+        *,
+        operation_id: str,
+        transport_ref: str,
+        intended_mode: NetworkConnectionMode,
+        write_mode: NetworkMutationWriteMode,
+        baseline_status: NetworkStatusEvidence,
+        previous_connection: PreviousConnectionEvidence | None = None,
+    ) -> NetworkMutationRecord:
+        """Persist a pre-write record after the live baseline has been read."""
+
+        _validate_identifier(operation_id, field_name="operation_id")
+        _validate_identifier(transport_ref, field_name="transport_ref")
+        _validate_connection_mode(intended_mode)
+        _validate_write_mode(write_mode)
+        if not isinstance(baseline_status, NetworkStatusEvidence):
+            raise TypeError("baseline_status must be NetworkStatusEvidence")
+        if previous_connection is not None and not isinstance(
+            previous_connection, PreviousConnectionEvidence
+        ):
+            raise TypeError("previous_connection must be PreviousConnectionEvidence")
+
+        with self._lock, self._process_lock_locked():
+            self._reload_locked()
+            self._require_mutation_allowed_locked()
+            previous_revision = self._record.revision if self._record is not None else 0
+            now = _nondecreasing_audit_timestamp(
+                self._clock(),
+                floor=(self._record.updated_at_utc if self._record is not None else None),
+            )
+            record = NetworkMutationRecord(
+                schema_version=NETWORK_MUTATION_LEDGER_SCHEMA,
+                revision=previous_revision + 1,
+                operation_id=operation_id,
+                transport_ref=transport_ref,
+                intended_mode=intended_mode,
+                stage="prepared",
+                write_mode=write_mode,
+                baseline_status=baseline_status,
+                previous_connection=previous_connection,
+                write_confirmed=None,
+                last_observation=None,
+                resolution=None,
+                created_at_utc=now,
+                updated_at_utc=now,
+            )
+            self._persist_locked(record)
+            return record
+
+    def mark_dispatching(
+        self,
+        operation_id: str,
+        *,
+        expected_revision: int,
+    ) -> NetworkMutationRecord:
+        """Durably cross the side-effect boundary before the BLE write call."""
+
+        _positive_int(expected_revision, field_name="expected_revision")
+        with self._lock, self._process_lock_locked():
+            current = self._current_operation_locked(operation_id)
+            self._require_expected_revision(current, expected_revision)
+            if current.stage != "prepared":
+                raise NetworkMutationTransitionError(
+                    "network mutation may dispatch only from prepared"
+                )
+            return self._transition_locked(current, stage="dispatching")
+
+    def confirm_dispatching_after_uncertain_return(
+        self,
+        operation_id: str,
+        *,
+        expected_prepared: NetworkMutationRecord,
+    ) -> NetworkMutationRecord:
+        """Re-fsync one exact DISPATCHING commit after its return path failed.
+
+        Reading a replaced file is not sufficient proof that its directory
+        entry reached stable storage. This recovery primitive verifies the
+        complete immutable predecessor/operation tuple under the store lock,
+        then republishes the identical record through the normal file+parent
+        fsync path. It never advances revision or creates dispatch authority
+        for a merely PREPARED row.
+        """
+
+        _validate_identifier(operation_id, field_name="operation_id")
+        if not isinstance(expected_prepared, NetworkMutationRecord):
+            raise TypeError("expected_prepared must be NetworkMutationRecord")
+        if (
+            expected_prepared.operation_id != operation_id
+            or expected_prepared.stage != "prepared"
+            or expected_prepared.write_confirmed is not None
+            or expected_prepared.last_observation is not None
+            or expected_prepared.resolution is not None
+        ):
+            raise NetworkMutationTransitionError(
+                "dispatch confirmation requires the exact PREPARED predecessor"
+            )
+        with self._lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt or self._record is None:
+                raise NetworkMutationLedgerCorrupt(
+                    "uncertain network dispatch record cannot be trusted"
+                )
+            current = self._record
+            if not (
+                current.operation_id == expected_prepared.operation_id
+                and current.transport_ref == expected_prepared.transport_ref
+                and current.intended_mode == expected_prepared.intended_mode
+                and current.write_mode == expected_prepared.write_mode
+                and current.baseline_status == expected_prepared.baseline_status
+                and current.previous_connection == expected_prepared.previous_connection
+                and current.created_at_utc == expected_prepared.created_at_utc
+                and current.stage == "dispatching"
+                and current.revision == expected_prepared.revision + 1
+                and current.write_confirmed is None
+                and current.last_observation is None
+                and current.resolution is None
+            ):
+                raise NetworkMutationTransitionError(
+                    "uncertain network dispatch does not match its predecessor"
+                )
+            self._persist_locked(current)
+            return current
+
+    def mark_observing(
+        self,
+        operation_id: str,
+        *,
+        expected_revision: int,
+        write_confirmed: bool,
+        observation: NetworkStatusEvidence | None = None,
+    ) -> NetworkMutationRecord:
+        """Record transport acknowledgement and bounded post-write evidence."""
+
+        _positive_int(expected_revision, field_name="expected_revision")
+        if not isinstance(write_confirmed, bool):
+            raise TypeError("write_confirmed must be bool")
+        if observation is not None and not isinstance(observation, NetworkStatusEvidence):
+            raise TypeError("observation must be NetworkStatusEvidence")
+        with self._lock, self._process_lock_locked():
+            current = self._current_operation_locked(operation_id)
+            self._require_expected_revision(current, expected_revision)
+            if current.stage not in {"dispatching", "observing"}:
+                raise NetworkMutationTransitionError(
+                    "network mutation may observe only after dispatch"
+                )
+            if current.write_confirmed is True and not write_confirmed:
+                raise NetworkMutationTransitionError(
+                    "network mutation write confirmation cannot regress"
+                )
+            return self._transition_locked(
+                current,
+                stage="observing",
+                write_confirmed=write_confirmed,
+                last_observation=observation,
+            )
+
+    def resolve(
+        self,
+        operation_id: str,
+        *,
+        expected_revision: int,
+        resolution: NetworkMutationResolution,
+        observation: NetworkStatusEvidence | None = None,
+    ) -> NetworkMutationRecord:
+        """Terminalize one mutation without authorizing an automatic retry.
+
+        ``interrupted`` and ``superseded`` are audit outcomes.  They make no
+        claim about the device-side result of a previously dispatched write;
+        they only state that the old host session no longer owns the next
+        explicit operator action.
+        """
+
+        _positive_int(expected_revision, field_name="expected_revision")
+        _validate_resolution(resolution)
+        if observation is not None and not isinstance(observation, NetworkStatusEvidence):
+            raise TypeError("observation must be NetworkStatusEvidence")
+        with self._lock, self._process_lock_locked():
+            current = self._current_operation_locked(operation_id)
+            self._require_expected_revision(current, expected_revision)
+            if current.stage == "resolved":
+                if current.resolution == resolution:
+                    return current
+                raise NetworkMutationTransitionError(
+                    "resolved network mutation cannot change its resolution"
+                )
+            if resolution == "not-dispatched":
+                if current.stage != "prepared":
+                    raise NetworkMutationTransitionError(
+                        "not-dispatched resolution requires prepared stage"
+                    )
+                if observation is not None:
+                    raise NetworkMutationTransitionError(
+                        "not-dispatched resolution cannot attach post-write evidence"
+                    )
+            elif resolution == "target-observed":
+                if current.stage not in {"dispatching", "observing"}:
+                    raise NetworkMutationTransitionError(
+                        "target-observed resolution requires a dispatched mutation"
+                    )
+                observation = observation or current.last_observation
+                if observation is None:
+                    raise NetworkMutationTransitionError(
+                        "target-observed resolution requires bounded status evidence"
+                    )
+            else:
+                if current.stage not in {"dispatching", "observing"}:
+                    raise NetworkMutationTransitionError(
+                        f"{resolution} resolution requires a dispatched mutation"
+                    )
+                # Session termination is not device-state evidence. Retain a
+                # previously captured bounded observation when one exists, but
+                # never require a read-only reconciliation before allowing a
+                # later explicit action.
+                observation = observation or current.last_observation
+            return self._transition_locked(
+                current,
+                stage="resolved",
+                last_observation=observation,
+                resolution=resolution,
+            )
+
+    @contextmanager
+    def _process_lock_locked(self) -> Iterator[None]:
+        """Serialize one complete ledger transaction across local processes.
+
+        Atomic replacement protects the JSON from partial publication, but it
+        does not make the preceding read/check/write sequence atomic. A stable,
+        separately opened lock inode fences that whole sequence so two backend
+        processes cannot both admit a device mutation from the same revision.
+        """
+
+        # The shared Mission Core data root may predate this plugin and is
+        # normalized by the same helper used by atomic publication. The
+        # ledger-specific directory must already be private or fail closed.
+        data_dir_created = _ensure_private_directory(self._data_dir, parents=True)
+        parent_created = _ensure_private_lock_directory(self.path.parent, parents=True)
+        if data_dir_created or parent_created:
+            _fsync_directory(self._data_dir)
+
+        flags = os.O_RDWR | os.O_CREAT | getattr(os, "O_CLOEXEC", 0)
+        flags |= getattr(os, "O_NOFOLLOW", 0)
+        try:
+            descriptor = os.open(self._process_lock_path, flags, 0o600)
+        except OSError as exc:
+            raise NetworkMutationLedgerCorrupt(
+                "network mutation ledger lock cannot be opened safely"
+            ) from exc
+        lock_stream: IO[bytes] | None = None
+        try:
+            metadata = os.fstat(descriptor)
+            if not stat.S_ISREG(metadata.st_mode) or stat.S_IMODE(metadata.st_mode) != 0o600:
+                raise NetworkMutationLedgerCorrupt(
+                    "network mutation ledger lock is not a private regular file"
+                )
+            lock_stream = os.fdopen(descriptor, "r+b", closefd=True)
+            descriptor = -1
+            fcntl.flock(lock_stream.fileno(), fcntl.LOCK_EX)
+            try:
+                yield
+            finally:
+                fcntl.flock(lock_stream.fileno(), fcntl.LOCK_UN)
+        finally:
+            if lock_stream is not None:
+                lock_stream.close()
+            elif descriptor >= 0:
+                os.close(descriptor)
+
+    def _current_operation_locked(self, operation_id: str) -> NetworkMutationRecord:
+        _validate_identifier(operation_id, field_name="operation_id")
+        self._reload_locked()
+        if self._corrupt:
+            raise NetworkMutationLedgerCorrupt(
+                "network mutation ledger is corrupt; device writes remain blocked"
+            )
+        current = self._record
+        if current is None or current.operation_id != operation_id:
+            raise NetworkMutationTransitionError("network mutation operation does not match ledger")
+        return current
+
+    def _require_mutation_allowed_locked(self) -> None:
+        if self._corrupt:
+            raise NetworkMutationLedgerCorrupt(
+                "network mutation ledger is corrupt; device writes remain blocked"
+            )
+        if self._record is not None and self._record.unresolved:
+            raise NetworkMutationBlocked(
+                "previous network mutation is unresolved; another device write is blocked"
+            )
+
+    @staticmethod
+    def _require_expected_revision(
+        record: NetworkMutationRecord,
+        expected_revision: int,
+    ) -> None:
+        if record.revision != expected_revision:
+            raise NetworkMutationTransitionError(
+                "network mutation transition used a stale record revision"
+            )
+
+    def _transition_locked(
+        self,
+        current: NetworkMutationRecord,
+        *,
+        stage: NetworkMutationStage,
+        write_confirmed: bool | None = None,
+        last_observation: NetworkStatusEvidence | None = None,
+        resolution: NetworkMutationResolution | None = None,
+    ) -> NetworkMutationRecord:
+        record = replace(
+            current,
+            revision=current.revision + 1,
+            stage=stage,
+            write_confirmed=(
+                write_confirmed if write_confirmed is not None else current.write_confirmed
+            ),
+            last_observation=(
+                last_observation if last_observation is not None else current.last_observation
+            ),
+            resolution=resolution,
+            updated_at_utc=_nondecreasing_audit_timestamp(
+                self._clock(),
+                floor=current.updated_at_utc,
+            ),
+        )
+        self._persist_locked(record)
+        return record
+
+    def _persist_locked(self, record: NetworkMutationRecord) -> None:
+        _write_private_json_atomic(
+            self.path,
+            record.as_dict(),
+            data_dir=self._data_dir,
+        )
+        self._record = record
+        self._corrupt = False
+
+    def _reload_locked(self) -> None:
+        try:
+            metadata = self.path.lstat()
+        except FileNotFoundError:
+            self._record = None
+            self._corrupt = False
+            return
+        except OSError:
+            self._record = None
+            self._corrupt = True
+            return
+        try:
+            if not stat.S_ISREG(metadata.st_mode) or stat.S_IMODE(metadata.st_mode) != 0o600:
+                raise ValueError("ledger file is not a private regular file")
+            parent_metadata = self.path.parent.lstat()
+            if (
+                not stat.S_ISDIR(parent_metadata.st_mode)
+                or stat.S_IMODE(parent_metadata.st_mode) != 0o700
+            ):
+                raise ValueError("ledger directory is not private")
+            if metadata.st_size > NETWORK_MUTATION_LEDGER_MAX_BYTES:
+                raise ValueError("ledger file exceeds the bounded size")
+            payload = json.loads(
+                self.path.read_text(encoding="utf-8"),
+                object_pairs_hook=_unique_json_object,
+            )
+            legacy_schema = (
+                isinstance(payload, dict)
+                and payload.get("schema_version") == NETWORK_MUTATION_LEDGER_LEGACY_SCHEMA
+            )
+            record = _record_from_mapping(payload)
+            if legacy_schema:
+                # v1 was briefly published with two shapes: the original
+                # previous_connection lacked transport_ref, while the final
+                # in-tree shape already carried it.  Only the latter can be
+                # migrated without inventing which physical K1 owned the
+                # previous topology.  _record_from_mapping deliberately
+                # rejects the ambiguous shape and preserves it fail-closed.
+                _write_private_json_atomic(
+                    self.path,
+                    record.as_dict(),
+                    data_dir=self._data_dir,
+                )
+        except (OSError, UnicodeError, json.JSONDecodeError, TypeError, ValueError):
+            self._record = None
+            self._corrupt = True
+            return
+        self._record = record
+        self._corrupt = False
+
+
+def _write_private_json_atomic(
+    path: Path,
+    payload: Mapping[str, object],
+    *,
+    data_dir: Path,
+) -> None:
+    serialized = (
+        json.dumps(payload, ensure_ascii=False, sort_keys=True, separators=(",", ":")) + "\n"
+    ).encode("utf-8")
+    if len(serialized) > NETWORK_MUTATION_LEDGER_MAX_BYTES:
+        raise ValueError("network mutation ledger exceeds the bounded size")
+
+    data_dir_created = _ensure_private_directory(data_dir, parents=True)
+    parent_created = _ensure_private_directory(path.parent, parents=True)
+    if data_dir_created or parent_created:
+        _fsync_directory(data_dir)
+
+    descriptor, temp_name = tempfile.mkstemp(
+        dir=path.parent,
+        prefix=f".{path.name}.",
+        suffix=".tmp",
+    )
+    temp_path = Path(temp_name)
+    try:
+        os.fchmod(descriptor, 0o600)
+        with os.fdopen(descriptor, "wb") as stream:
+            descriptor = -1
+            stream.write(serialized)
+            stream.flush()
+            os.fsync(stream.fileno())
+        os.replace(temp_path, path)
+        path.chmod(0o600)
+        _fsync_directory(path.parent)
+    finally:
+        if descriptor >= 0:
+            os.close(descriptor)
+        temp_path.unlink(missing_ok=True)
+
+
+def _fsync_directory(path: Path) -> None:
+    flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0)
+    descriptor = os.open(path, flags)
+    try:
+        os.fsync(descriptor)
+    finally:
+        os.close(descriptor)
+
+
+def _ensure_private_directory(path: Path, *, parents: bool) -> bool:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        try:
+            path.mkdir(mode=0o700, parents=parents, exist_ok=False)
+        except FileExistsError:
+            # A peer process may have created the shared private directory
+            # between lstat and mkdir. Validate that winner below instead of
+            # failing a safe concurrent ledger acquisition.
+            metadata = path.lstat()
+        else:
+            path.chmod(0o700)
+            return True
+    if not stat.S_ISDIR(metadata.st_mode):
+        raise NetworkMutationLedgerCorrupt(
+            "network mutation ledger directory is not a private directory"
+        )
+    path.chmod(0o700)
+    return False
+
+
+def _ensure_private_lock_directory(path: Path, *, parents: bool) -> bool:
+    """Create a lock parent privately or reject an unsafe existing parent."""
+
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        try:
+            path.mkdir(mode=0o700, parents=parents, exist_ok=False)
+        except FileExistsError:
+            metadata = path.lstat()
+        else:
+            path.chmod(0o700)
+            return True
+    if not stat.S_ISDIR(metadata.st_mode) or stat.S_IMODE(metadata.st_mode) != 0o700:
+        raise NetworkMutationLedgerCorrupt("network mutation ledger lock directory is not private")
+    return False
+
+
+def _unique_json_object(pairs: list[tuple[str, object]]) -> dict[str, object]:
+    document: dict[str, object] = {}
+    for key, value in pairs:
+        if key in document:
+            raise ValueError("network mutation ledger contains duplicate fields")
+        document[key] = value
+    return document
+
+
+def _record_from_mapping(value: object) -> NetworkMutationRecord:
+    document = _exact_mapping(
+        value,
+        {
+            "schema_version",
+            "revision",
+            "operation_id",
+            "transport_ref",
+            "intended_mode",
+            "stage",
+            "write_mode",
+            "baseline_status",
+            "previous_connection",
+            "write_confirmed",
+            "last_observation",
+            "resolution",
+            "created_at_utc",
+            "updated_at_utc",
+        },
+        label="ledger",
+    )
+    if document["schema_version"] not in {
+        NETWORK_MUTATION_LEDGER_SCHEMA,
+        NETWORK_MUTATION_LEDGER_LEGACY_SCHEMA,
+    }:
+        raise ValueError("unsupported network mutation ledger schema")
+    revision = _positive_int(document["revision"], field_name="revision")
+    operation_id = _required_string(document["operation_id"], field_name="operation_id")
+    transport_ref = _required_string(document["transport_ref"], field_name="transport_ref")
+    _validate_identifier(operation_id, field_name="operation_id")
+    _validate_identifier(transport_ref, field_name="transport_ref")
+
+    intended_mode_raw = _required_string(document["intended_mode"], field_name="intended_mode")
+    _validate_connection_mode(intended_mode_raw)
+    intended_mode = cast(NetworkConnectionMode, intended_mode_raw)
+    stage_raw = _required_string(document["stage"], field_name="stage")
+    if stage_raw not in _STAGES:
+        raise ValueError("unsupported network mutation stage")
+    stage = cast(NetworkMutationStage, stage_raw)
+    write_mode_raw = _required_string(document["write_mode"], field_name="write_mode")
+    _validate_write_mode(write_mode_raw)
+    write_mode = cast(NetworkMutationWriteMode, write_mode_raw)
+    baseline_status = _status_from_mapping(document["baseline_status"])
+
+    previous_raw = document["previous_connection"]
+    previous_connection = (
+        None if previous_raw is None else _previous_connection_from_mapping(previous_raw)
+    )
+    write_confirmed_raw = document["write_confirmed"]
+    if write_confirmed_raw is not None and not isinstance(write_confirmed_raw, bool):
+        raise ValueError("write_confirmed must be bool or null")
+    observation_raw = document["last_observation"]
+    last_observation = None if observation_raw is None else _status_from_mapping(observation_raw)
+    resolution_raw = document["resolution"]
+    if resolution_raw is None:
+        resolution = None
+    else:
+        resolution_string = _required_string(resolution_raw, field_name="resolution")
+        _validate_resolution(resolution_string)
+        resolution = cast(NetworkMutationResolution, resolution_string)
+    created_at = _validated_timestamp(document["created_at_utc"], field_name="created_at_utc")
+    updated_at = _validated_timestamp(document["updated_at_utc"], field_name="updated_at_utc")
+    if updated_at < created_at:
+        raise ValueError("ledger update precedes creation")
+
+    if stage == "prepared":
+        if (
+            write_confirmed_raw is not None
+            or last_observation is not None
+            or resolution is not None
+        ):
+            raise ValueError("prepared ledger contains post-write fields")
+    elif stage in {"dispatching", "observing"}:
+        if resolution is not None:
+            raise ValueError("unresolved ledger contains a resolution")
+        if stage == "dispatching" and (
+            write_confirmed_raw is not None or last_observation is not None
+        ):
+            raise ValueError("dispatching ledger contains observation fields")
+        if stage == "observing" and write_confirmed_raw is None:
+            raise ValueError("observing ledger lacks write acknowledgement status")
+    else:
+        if resolution is None:
+            raise ValueError("resolved ledger lacks a resolution")
+        if resolution == "not-dispatched" and (
+            write_confirmed_raw is not None or last_observation is not None
+        ):
+            raise ValueError("not-dispatched resolution contains post-write fields")
+        if resolution == "target-observed" and last_observation is None:
+            raise ValueError("target-observed resolution lacks status evidence")
+
+    return NetworkMutationRecord(
+        schema_version=NETWORK_MUTATION_LEDGER_SCHEMA,
+        revision=revision,
+        operation_id=operation_id,
+        transport_ref=transport_ref,
+        intended_mode=intended_mode,
+        stage=stage,
+        write_mode=write_mode,
+        baseline_status=baseline_status,
+        previous_connection=previous_connection,
+        write_confirmed=write_confirmed_raw,
+        last_observation=last_observation,
+        resolution=resolution,
+        created_at_utc=_timestamp(created_at),
+        updated_at_utc=_timestamp(updated_at),
+    )
+
+
+def _status_from_mapping(value: object) -> NetworkStatusEvidence:
+    document = _exact_mapping(
+        value,
+        {"mode", "ipv4", "status_code", "reserved"},
+        label="network status",
+    )
+    mode = _optional_string(document["mode"], field_name="mode")
+    ipv4 = _optional_string(document["ipv4"], field_name="ipv4")
+    status_code = _byte(document["status_code"], field_name="status_code")
+    reserved_raw = document["reserved"]
+    reserved = None if reserved_raw is None else _byte(reserved_raw, field_name="reserved")
+    return NetworkStatusEvidence(
+        mode=mode,
+        ipv4=ipv4,
+        status_code=status_code,
+        reserved=reserved,
+    )
+
+
+def _previous_connection_from_mapping(value: object) -> PreviousConnectionEvidence:
+    document = _exact_mapping(
+        value,
+        {"transport_ref", "mode", "ipv4", "device_session_id"},
+        label="previous connection",
+    )
+    mode_raw = _required_string(document["mode"], field_name="mode")
+    _validate_connection_mode(mode_raw)
+    return PreviousConnectionEvidence(
+        transport_ref=_required_string(document["transport_ref"], field_name="transport_ref"),
+        mode=cast(NetworkConnectionMode, mode_raw),
+        ipv4=_optional_string(document["ipv4"], field_name="ipv4"),
+        device_session_id=_optional_string(
+            document["device_session_id"], field_name="device_session_id"
+        ),
+    )
+
+
+def _exact_mapping(value: object, keys: set[str], *, label: str) -> Mapping[str, object]:
+    if not isinstance(value, dict) or set(value) != keys:
+        raise ValueError(f"{label} does not match the secret-free schema")
+    return cast(Mapping[str, object], value)
+
+
+def _validate_identifier(value: str, *, field_name: str) -> None:
+    if _SAFE_IDENTIFIER.fullmatch(value) is None:
+        raise ValueError(f"{field_name} is outside the secret-free identifier schema")
+
+
+def _validate_connection_mode(value: str) -> None:
+    if value not in _CONNECTION_MODES:
+        raise ValueError("unsupported network connection mode")
+
+
+def _validate_write_mode(value: str) -> None:
+    if value not in _WRITE_MODES:
+        raise ValueError("unsupported BLE write mode")
+
+
+def _validate_resolution(value: str) -> None:
+    if value not in _RESOLUTIONS:
+        raise ValueError("unsupported network mutation resolution")
+
+
+def _validate_byte(value: int, *, field_name: str) -> None:
+    if isinstance(value, bool) or not isinstance(value, int) or not 0 <= value <= 255:
+        raise ValueError(f"{field_name} must be an unsigned byte")
+
+
+def _byte(value: object, *, field_name: str) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or not 0 <= value <= 255:
+        raise ValueError(f"{field_name} must be an unsigned byte")
+    return value
+
+
+def _positive_int(value: object, *, field_name: str) -> int:
+    if isinstance(value, bool) or not isinstance(value, int) or value < 1:
+        raise ValueError(f"{field_name} must be a positive integer")
+    return value
+
+
+def _required_string(value: object, *, field_name: str) -> str:
+    if not isinstance(value, str) or not value:
+        raise ValueError(f"{field_name} must be a non-empty string")
+    return value
+
+
+def _optional_string(value: object, *, field_name: str) -> str | None:
+    if value is None:
+        return None
+    return _required_string(value, field_name=field_name)
+
+
+def _validated_timestamp(value: object, *, field_name: str) -> datetime:
+    raw = _required_string(value, field_name=field_name)
+    if not raw.endswith("Z"):
+        raise ValueError(f"{field_name} must be UTC")
+    try:
+        parsed = datetime.fromisoformat(raw.removesuffix("Z") + "+00:00")
+    except ValueError as exc:
+        raise ValueError(f"{field_name} is invalid") from exc
+    if parsed.tzinfo is None or parsed.utcoffset() != UTC.utcoffset(parsed):
+        raise ValueError(f"{field_name} must be UTC")
+    return parsed.astimezone(UTC)
+
+
+def _timestamp(value: datetime) -> str:
+    if value.tzinfo is None or value.utcoffset() is None:
+        raise ValueError("network mutation ledger clock must be timezone-aware")
+    return value.astimezone(UTC).isoformat(timespec="milliseconds").replace("+00:00", "Z")
+
+
+def _nondecreasing_audit_timestamp(value: datetime, *, floor: str | None) -> str:
+    """Canonicalize wall time without using it as transition authority.
+
+    Serialized revision and stage checks order ledger transitions.  A host clock
+    may move backwards (NTP correction, RTC repair, suspend/resume), so the
+    human-readable audit timestamp is clamped to the prior durable value rather
+    than allowing a valid transition to publish a record that fails its own
+    ``updated >= created`` structural check after restart.
+    """
+
+    candidate = _validated_timestamp(_timestamp(value), field_name="ledger clock")
+    if floor is None:
+        return _timestamp(candidate)
+    floor_value = _validated_timestamp(floor, field_name="audit timestamp floor")
+    return _timestamp(max(candidate, floor_value))

src/k1link/device_plugins/xgrids_k1/observation.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import os
 import stat
 import threading
 from collections.abc import Mapping
@@ -41,8 +42,16 @@ XGRIDS_K1_PLUGIN_ID = "nodedc.device.xgrids-lixelkity-k1"
 def build_xgrids_k1_observation(repository_root: Path) -> ObservationRuntimeContribution:
     """Compose every K1 evidence root behind the generic observation ABI."""
 
+    configured_legacy_root = os.environ.get(
+        "MISSIONCORE_LEGACY_SESSIONS_DIR", ""
+    ).strip()
+    legacy_root = (
+        Path(configured_legacy_root).expanduser().resolve()
+        if configured_legacy_root
+        else repository_root.resolve() / "sessions"
+    )
     roots = (
-        ("xgrids-k1.viewer-live.repository", repository_root.resolve() / "sessions"),
+        ("xgrids-k1.viewer-live.repository", legacy_root),
         (
             "xgrids-k1.viewer-live.evidence",
             resolve_missioncore_evidence_dir(repository_root),

src/k1link/device_plugins/xgrids_k1/protocol/application_acceptance.py

@@ -21,6 +21,7 @@ from k1link.device_plugins.xgrids_k1.protocol.application_bootstrap import (
 )
 from k1link.device_plugins.xgrids_k1.protocol.application_mqtt import (
     MODELING_RESPONSE_TOPIC,
+    ApplicationMqttTransportError,
 )
 from k1link.device_plugins.xgrids_k1.protocol.application_publish import (
     OneShotPublishEnvelope,
@@ -65,6 +66,8 @@ class ApplicationBatchExchange(Protocol):
         envelopes: Sequence[OneShotPublishEnvelope],
         *,
         required_response_operation_keys: Collection[str],
+        dispatch_admission_deadline_reached: Callable[[], bool] | None = None,
+        dispatch_admission_commit: Callable[[], None] | None = None,
     ) -> dict[str, bytes]: ...
 
     def maintain_open_for(
@@ -129,8 +132,31 @@ class OperatorDialogueCheckpoint:
             )
 
 
+class PhysicalAcceptancePermitReservation:
+    """One commit right that must still be fresh at dispatch admission."""
+
+    def __init__(self, permit: PhysicalAcceptancePermit, token: object) -> None:
+        self._permit = permit
+        self._token = token
+        self._finished = False
+
+    def commit(self) -> None:
+        if self._finished:
+            raise ApplicationAcceptanceError(
+                "physical acceptance reservation was already finished"
+            )
+        self._permit._commit_reservation(self._token)  # noqa: SLF001
+        self._finished = True
+
+    def release(self) -> None:
+        if self._finished:
+            return
+        self._permit._release_reservation(self._token)  # noqa: SLF001
+        self._finished = True
+
+
 class PhysicalAcceptancePermit:
-    """Short, single-action capability that is consumed before MQTT publish."""
+    """Short, single-action capability committed at physical dispatch admission."""
 
     def __init__(
         self,
@@ -152,20 +178,54 @@ class PhysicalAcceptancePermit:
         self._monotonic = monotonic
         self._expires_at = monotonic() + float(ttl_seconds)
         self._consumed = False
+        self._reservation: object | None = None
 
     @property
     def action(self) -> ModelingAction:
         return self._checklist.action
 
     def consume(self, action: ModelingAction) -> None:
+        reservation = self.reserve(action)
+        reservation.commit()
+
+    def reserve(
+        self,
+        action: ModelingAction,
+    ) -> PhysicalAcceptancePermitReservation:
         with self._lock:
             if self._consumed:
                 raise ApplicationAcceptanceError("physical acceptance permit was already consumed")
+            if self._reservation is not None:
+                raise ApplicationAcceptanceError("physical acceptance permit is already reserved")
             if self._monotonic() >= self._expires_at:
                 raise ApplicationAcceptanceError("physical acceptance permit expired")
             if action is not self._checklist.action:
                 raise ApplicationAcceptanceError("physical acceptance permit action mismatch")
+            token = object()
+            self._reservation = token
+            return PhysicalAcceptancePermitReservation(self, token)
+
+    def _commit_reservation(self, token: object) -> None:
+        with self._lock:
+            if self._consumed or self._reservation is not token:
+                raise ApplicationAcceptanceError(
+                    "physical acceptance reservation is no longer current"
+                )
+            if self._monotonic() >= self._expires_at:
+                self._reservation = None
+                raise ApplicationMqttTransportError(
+                    "physical acceptance permit expired before dispatch admission",
+                    reason_code=(
+                        "physical-acceptance-permit-expired-before-dispatch"
+                    ),
+                )
             self._consumed = True
+            self._reservation = None
+
+    def _release_reservation(self, token: object) -> None:
+        with self._lock:
+            if self._reservation is token and not self._consumed:
+                self._reservation = None
 
     def snapshot(self) -> dict[str, object]:
         with self._lock:
@@ -218,6 +278,7 @@ class PhysicalAcceptanceDialogueExecutor:
         self._command_complete = False
         self._dialogue_stage = "new"
         self._start_complete = False
+        self._active_session_adopted = False
         self._stop_attempted = False
         self._stop_complete = False
         self._active_authority: ApplicationControlAuthority | None = None
@@ -246,13 +307,48 @@ class PhysicalAcceptanceDialogueExecutor:
     ) -> LiveDeviceControlBinding:
         """Emit retained ordinals 1-6 at control-session establishment."""
 
+        binding = self.run_read_only_inspection_stage(orchestrator)
+        return self.complete_connection_stage(orchestrator, expected_binding=binding)
+
+    def run_read_only_inspection_stage(
+        self,
+        orchestrator: ShadowApplicationBootstrapOrchestrator,
+    ) -> LiveDeviceControlBinding:
+        """Emit only ordinal-1 DeviceInfo for an explicit passive Verify.
+
+        ModelingStatus and the time-setting DeviceConfig request belong to the
+        canonical preparation dialogue.  They are deliberately excluded from
+        this stage so a recovery Verify cannot cross a device-mutation edge.
+        A later explicit workspace action may promote this same socket through
+        :meth:`complete_connection_stage`.
+        """
+
         if self._dialogue_stage != "new" or self._bootstrap_complete or self._command_complete:
-            raise ApplicationAcceptanceError("connection stage is not admissible now")
-        for expected_batch in (1, 2):
-            self._exchange_bootstrap_batch(orchestrator, expected_batch=expected_batch)
+            raise ApplicationAcceptanceError("inspection stage is not admissible now")
+        self._exchange_bootstrap_batch(orchestrator, expected_batch=1)
         binding = orchestrator.binding
         if binding is None:
-            raise ApplicationAcceptanceError("connection stage produced no live device binding")
+            raise ApplicationAcceptanceError("inspection stage produced no live device binding")
+        self._prepared_binding = binding
+        self._dialogue_stage = "inspection-ready"
+        return binding
+
+    def complete_connection_stage(
+        self,
+        orchestrator: ShadowApplicationBootstrapOrchestrator,
+        *,
+        expected_binding: LiveDeviceControlBinding,
+    ) -> LiveDeviceControlBinding:
+        """Promote an ordinal-1 inspection into the canonical ordinals 2-6."""
+
+        if self._dialogue_stage != "inspection-ready":
+            raise ApplicationAcceptanceError("connection completion requires inspection-ready")
+        if self._prepared_binding != expected_binding:
+            raise ApplicationAcceptanceError("inspection binding changed before completion")
+        self._exchange_bootstrap_batch(orchestrator, expected_batch=2)
+        binding = orchestrator.binding
+        if binding is None or binding != expected_binding:
+            raise ApplicationAcceptanceError("connection stage changed the inspected identity")
         self._prepared_binding = binding
         self._dialogue_stage = "connection-ready"
         return binding
@@ -261,10 +357,13 @@ class PhysicalAcceptanceDialogueExecutor:
         self,
         event: Literal["workspace-entered", "project-prompt-opened", "start-confirmed"],
         event_observed: Callable[[], bool],
-    ) -> OperatorDialogueCheckpoint:
+        *,
+        reconciled_active_observed: Callable[[], bool] | None = None,
+    ) -> OperatorDialogueCheckpoint | None:
         """Service the original socket until one exact operator UI event occurs."""
 
         expected = {
+            "inspection-ready": "workspace-entered",
             "connection-ready": "workspace-entered",
             "workspace-ready": "project-prompt-opened",
             "project-ready": "start-confirmed",
@@ -276,9 +375,19 @@ class PhysicalAcceptanceDialogueExecutor:
         binding = self._prepared_binding
         if binding is None:
             raise ApplicationAcceptanceError("canonical preparation binding is unavailable")
-        while not (
-            self._transport.pre_start_ready(binding) and event_observed()
-        ):
+        if reconciled_active_observed is not None and event != "workspace-entered":
+            raise ApplicationAcceptanceError(
+                "active recovery may replace only the workspace-entry checkpoint"
+            )
+        while True:
+            if reconciled_active_observed is not None and reconciled_active_observed():
+                if not self._transport.scan_initialization_complete(binding):
+                    raise ApplicationAcceptanceError(
+                        "active recovery requires fresh bound SCANNING state"
+                    )
+                return None
+            if self._transport.pre_start_ready(binding) and event_observed():
+                break
             self._transport.maintain_open_for(
                 CONTROL_NETWORK_PUMP_QUANTUM_SECONDS,
                 allowed_response_topics={MODELING_STATUS_RESPONSE_TOPIC},
@@ -291,16 +400,52 @@ class PhysicalAcceptanceDialogueExecutor:
             owner_token=self._checkpoint_owner,
         )
 
+    def adopt_reconciled_scanning(
+        self,
+        *,
+        authority: ApplicationControlAuthority,
+        binding: LiveDeviceControlBinding,
+    ) -> None:
+        """Adopt externally observed SCANNING without inventing a new START.
+
+        The caller has already committed an exact read-only physical-ledger
+        reconciliation for this control generation.  This method performs no
+        publish; it only gives the existing socket enough local state to wait
+        for one later operator-confirmed STOP.
+        """
+
+        if self._dialogue_stage not in {"inspection-ready", "connection-ready"}:
+            raise ApplicationAcceptanceError(
+                "active recovery requires an inspected pre-START control session"
+            )
+        if self._prepared_binding != binding:
+            raise ApplicationAcceptanceError("active recovery binding changed")
+        if self._command_complete or self._start_complete or self._stop_attempted:
+            raise ApplicationAcceptanceError("active recovery cannot replace a command attempt")
+        if not self._transport.scan_initialization_complete(binding):
+            raise ApplicationAcceptanceError(
+                "active recovery requires the bound K1 to report SCANNING"
+            )
+        self._active_authority = authority
+        self._active_binding = binding
+        self._prepared_binding = None
+        self._active_session_adopted = True
+        self._dialogue_stage = "post-initialization-observed"
+
     def run_workspace_entry_stage(
         self,
         orchestrator: ShadowApplicationBootstrapOrchestrator,
         checkpoint: OperatorDialogueCheckpoint,
+        *,
+        dispatch_guard: Callable[[], None] | None = None,
     ) -> LiveDeviceControlBinding:
         """Emit ordinal 7 only for the observed scan-workspace entry action."""
 
         if self._dialogue_stage != "connection-ready":
             raise ApplicationAcceptanceError("workspace entry requires the connection stage")
         self._consume_checkpoint(checkpoint, expected="workspace-entered")
+        if dispatch_guard is not None:
+            dispatch_guard()
         self._exchange_bootstrap_batch(orchestrator, expected_batch=3)
         binding = orchestrator.binding
         if binding is None:
@@ -312,12 +457,16 @@ class PhysicalAcceptanceDialogueExecutor:
         self,
         orchestrator: ShadowApplicationBootstrapOrchestrator,
         checkpoint: OperatorDialogueCheckpoint,
+        *,
+        dispatch_guard: Callable[[], None] | None = None,
     ) -> LiveDeviceControlBinding:
         """Emit ordinals 8-10 when the operator opens the project-name prompt."""
 
         if self._dialogue_stage != "workspace-ready":
             raise ApplicationAcceptanceError("project prompt requires workspace entry")
         self._consume_checkpoint(checkpoint, expected="project-prompt-opened")
+        if dispatch_guard is not None:
+            dispatch_guard()
         self._exchange_bootstrap_batch(orchestrator, expected_batch=4)
         if not orchestrator.snapshot().bootstrap_complete:
             raise ApplicationAcceptanceError("project prompt did not complete the transcript")
@@ -343,6 +492,7 @@ class PhysicalAcceptanceDialogueExecutor:
         binding: LiveDeviceControlBinding,
         permit: PhysicalAcceptancePermit,
         checkpoint: OperatorDialogueCheckpoint,
+        dispatch_guard: Callable[[], None] | None = None,
     ) -> ModelingResponse:
         """Execute retained operations 11-14 on one continuously serviced socket."""
 
@@ -363,6 +513,8 @@ class PhysicalAcceptanceDialogueExecutor:
         if permit.action is not ModelingAction.START:
             raise ApplicationAcceptanceError("canonical START requires a fresh START permit")
 
+        if dispatch_guard is not None:
+            dispatch_guard()
         permit.consume(ModelingAction.START)
         self._start_permit_snapshot = permit.snapshot()
         self._command_complete = True
@@ -392,6 +544,8 @@ class PhysicalAcceptanceDialogueExecutor:
             ) from exc
 
         immediate = post_start.immediate_modeling_status
+        if dispatch_guard is not None:
+            dispatch_guard()
         self._transport.exchange_batch_once(
             [OneShotPublishEnvelope.from_dialogue_request(immediate)],
             required_response_operation_keys=(),
@@ -414,6 +568,8 @@ class PhysicalAcceptanceDialogueExecutor:
         required_operations = {
             f"dialogue:{request.ordinal}:{request.message_type}" for request in refresh
         }
+        if dispatch_guard is not None:
+            dispatch_guard()
         refresh_responses = self._transport.exchange_batch_once(
             [OneShotPublishEnvelope.from_dialogue_request(request) for request in refresh],
             required_response_operation_keys=required_operations,
@@ -469,7 +625,9 @@ class PhysicalAcceptanceDialogueExecutor:
     ) -> None:
         """Continuously service the original socket until the operator requests STOP."""
 
-        if self._dialogue_stage != "post-initialization-observed" or not self._start_complete:
+        if self._dialogue_stage != "post-initialization-observed" or not (
+            self._start_complete or self._active_session_adopted
+        ):
             raise ApplicationAcceptanceError(
                 "active control ownership requires the complete post-START dialogue"
             )
@@ -491,12 +649,17 @@ class PhysicalAcceptanceDialogueExecutor:
         self,
         command: ShadowModelingCommand,
         permit: PhysicalAcceptancePermit,
+        *,
+        dispatch_guard: Callable[[], None] | None = None,
+        dispatch_admission_deadline_reached: Callable[[], bool] | None = None,
     ) -> ModelingResponse:
         """Emit retained STOP on the same socket and with a separate permit."""
 
         if command.action is not ModelingAction.STOP:
             raise ApplicationAcceptanceError("canonical STOP executor requires STOP")
-        if self._dialogue_stage != "stop-requested" or not self._start_complete:
+        if self._dialogue_stage != "stop-requested" or not (
+            self._start_complete or self._active_session_adopted
+        ):
             raise ApplicationAcceptanceError(
                 "STOP requires continuous ownership from the canonical START session"
             )
@@ -507,21 +670,66 @@ class PhysicalAcceptanceDialogueExecutor:
         if authority is None or binding is None:
             raise ApplicationAcceptanceError("canonical START binding is no longer available")
         self._require_command_identity(command, authority=authority, binding=binding)
-        if not self._transport.scan_initialization_complete(binding):
+        self._require_stop_dispatch_deadline_open(
+            dispatch_admission_deadline_reached
+        )
+        scanning = self._transport.scan_initialization_complete(binding)
+        # The status projection can wait on the transport lock.  Deadline
+        # admission is therefore sampled again before its result can advance
+        # the physical STOP dialogue.
+        self._require_stop_dispatch_deadline_open(
+            dispatch_admission_deadline_reached
+        )
+        if not scanning:
             raise ApplicationAcceptanceError(
                 "canonical STOP requires the bound K1 to still report SCANNING"
             )
         if permit.action is not ModelingAction.STOP:
             raise ApplicationAcceptanceError("canonical STOP requires a separate STOP permit")
 
-        permit.consume(ModelingAction.STOP)
-        self._stop_permit_snapshot = permit.snapshot()
+        if dispatch_guard is not None:
+            dispatch_guard()
+            # Route/control validation is read-only but may block.  It cannot
+            # authorize a publish whose operation deadline elapsed meanwhile.
+            self._require_stop_dispatch_deadline_open(
+                dispatch_admission_deadline_reached
+            )
+        permit_reservation = permit.reserve(ModelingAction.STOP)
         self._stop_attempted = True
         self._dialogue_stage = "stop-attempted"
-        responses = self._transport.exchange_batch_once(
-            [OneShotPublishEnvelope.from_modeling_command(command)],
-            required_response_operation_keys={"modeling:stop"},
-        )
+        # Keep source-compatible test/integration transports on the legacy
+        # call shape unless a real operation deadline was supplied.
+        stop_envelope = OneShotPublishEnvelope.from_modeling_command(command)
+        try:
+            if dispatch_admission_deadline_reached is None:
+                responses = self._transport.exchange_batch_once(
+                    [stop_envelope],
+                    required_response_operation_keys={"modeling:stop"},
+                    dispatch_admission_commit=permit_reservation.commit,
+                )
+            else:
+                responses = self._transport.exchange_batch_once(
+                    [stop_envelope],
+                    required_response_operation_keys={"modeling:stop"},
+                    dispatch_admission_deadline_reached=(
+                        dispatch_admission_deadline_reached
+                    ),
+                    dispatch_admission_commit=permit_reservation.commit,
+                )
+        except ApplicationMqttTransportError as exc:
+            if exc.reason_code in {
+                "physical-command-dispatch-deadline-expired",
+                "physical-acceptance-permit-expired-before-dispatch",
+            }:
+                # Atomic physical admission rejected either the operation
+                # deadline or the still-fresh permit before durable
+                # DISPATCHING. Preserve that stronger zero-attempt fact.
+                permit_reservation.release()
+                self._stop_attempted = False
+                self._dialogue_stage = "stop-requested"
+            raise
+        finally:
+            self._stop_permit_snapshot = permit.snapshot()
         payload = responses["modeling:stop"]
         self._record_response_evidence(
             phase="modeling",
@@ -545,6 +753,19 @@ class PhysicalAcceptanceDialogueExecutor:
         self._dialogue_stage = "stop-acknowledged"
         return response
 
+    @staticmethod
+    def _require_stop_dispatch_deadline_open(
+        dispatch_admission_deadline_reached: Callable[[], bool] | None,
+    ) -> None:
+        if (
+            dispatch_admission_deadline_reached is not None
+            and dispatch_admission_deadline_reached()
+        ):
+            raise ApplicationMqttTransportError(
+                "control command dispatch deadline expired before publish admission",
+                reason_code="physical-command-dispatch-deadline-expired",
+            )
+
     def maintain_post_stop_until_standby(self) -> None:
         """Keep servicing control reports through save and protocol standby.
 
@@ -577,6 +798,7 @@ class PhysicalAcceptanceDialogueExecutor:
             "command_complete": self._command_complete,
             "start_attempted": self._command_complete,
             "start_complete": self._start_complete,
+            "active_session_adopted": self._active_session_adopted,
             "stop_attempted": self._stop_attempted,
             "stop_complete": self._stop_complete,
             "dialogue_stage": self._dialogue_stage,

src/k1link/device_plugins/xgrids_k1/protocol/application_authority.py

@@ -17,11 +17,35 @@ KEYCHAIN_SERVICE = "NODEDC Mission Core XGRIDS K1 OpenAPI"
 KEYCHAIN_ACCOUNT = "lixelgo-application-fw-3.0.2"
 KEYCHAIN_TIMEOUT_SECONDS = 5.0
 KEYCHAIN_INTERACTIVE_TIMEOUT_SECONDS = 300.0
+_ERR_SEC_USER_CANCELED = -128
+_ERR_SEC_AUTH_FAILED = -25293
+_ERR_SEC_INTERACTION_NOT_ALLOWED = -25308
 
 
 class ApplicationAuthorityLoadError(RuntimeError):
     """The private application authority could not be loaded safely."""
 
+    def __init__(
+        self,
+        message: str,
+        *,
+        reason_code: str = "application_authority_unavailable",
+    ) -> None:
+        super().__init__(message)
+        self.reason_code = reason_code
+
+
+def _keychain_authority_reason_code(status: int) -> str:
+    """Reduce an OSStatus to a reviewed, secret-free operator class."""
+
+    if status == _ERR_SEC_INTERACTION_NOT_ALLOWED:
+        return "keychain-authorization-required"
+    if status == _ERR_SEC_AUTH_FAILED:
+        return "keychain-authorization-denied"
+    if status == _ERR_SEC_USER_CANCELED:
+        return "keychain-authorization-cancelled"
+    return "application_authority_unavailable"
+
 
 class CommandRunner(Protocol):
     def __call__(
@@ -97,8 +121,12 @@ def _read_keychain_secret_via_security_framework(*, service: str, account: str)
     if not isinstance(result, tuple) or len(result) != 2:
         raise ApplicationAuthorityLoadError("macOS Keychain authority lookup failed")
     status, secret_data = result
-    if int(status) != 0 or secret_data is None:
-        raise ApplicationAuthorityLoadError("macOS Keychain authority is unavailable")
+    status_code = int(status)
+    if status_code != 0 or secret_data is None:
+        raise ApplicationAuthorityLoadError(
+            "macOS Keychain authority is unavailable",
+            reason_code=_keychain_authority_reason_code(status_code),
+        )
     try:
         return bytes(secret_data)
     except Exception as exc:

src/k1link/device_plugins/xgrids_k1/semantic_topology_store.py

@@ -0,0 +1,667 @@
+from __future__ import annotations
+
+import fcntl
+import ipaddress
+import json
+import os
+import re
+import stat
+import tempfile
+import threading
+from collections.abc import Iterator, Mapping
+from contextlib import contextmanager
+from dataclasses import dataclass
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import IO, Final, Literal, cast
+
+from k1link.sessions.store import resolve_missioncore_data_dir
+
+SEMANTIC_TOPOLOGY_SCHEMA: Final = "missioncore.xgrids-k1-semantic-topology/v1"
+SEMANTIC_TOPOLOGY_FILENAME = "semantic-topology.json"
+SEMANTIC_TOPOLOGY_LOCK_FILENAME = ".semantic-topology.lock"
+SEMANTIC_TOPOLOGY_MAX_BYTES = 16 * 1024
+SEMANTIC_TOPOLOGY_MAX_REVISION = (1 << 63) - 1
+
+TopologyConnectionMode = Literal["bridge", "quick-connect", "direct-connect"]
+TopologyEvidenceSource = Literal["ble-post-write-status", "ble-read-only-status"]
+SemanticTopologyStoreStatus = Literal["empty", "available", "corrupt"]
+
+_CONNECTION_MODES = frozenset({"bridge", "quick-connect", "direct-connect"})
+_EVIDENCE_SOURCES = frozenset({"ble-post-write-status", "ble-read-only-status"})
+_SAFE_IDENTIFIER = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._:+-]{0,159}$")
+_SAFE_FIRMWARE_VERSION = re.compile(r"^[A-Za-z0-9][A-Za-z0-9._+() -]{0,63}$")
+
+
+class SemanticTopologyStoreError(RuntimeError):
+    """Base error for durable, non-authoritative K1 topology evidence."""
+
+    reason_code = "semantic-topology-store-error"
+
+
+class SemanticTopologyStoreCorrupt(SemanticTopologyStoreError):
+    """The on-disk evidence cannot be trusted and must not be adopted."""
+
+    reason_code = "semantic-topology-store-corrupt"
+
+
+class StaleSemanticTopologyObservation(SemanticTopologyStoreError):
+    """An asynchronous writer no longer descends from the durable revision."""
+
+    reason_code = "semantic-topology-observation-stale"
+
+
+@dataclass(frozen=True, slots=True)
+class SemanticTopologyRecord:
+    """Last topology proved by an exact K1 BLE status read.
+
+    This record proves only what the K1 reported at ``observed_at_utc``.  It is
+    deliberately not a host-route, TCP, MQTT, connection, or acquisition lease.
+    A record loaded after restart is configured/offline evidence until all live
+    connection-supervisor gates are proved again.
+    """
+
+    schema_version: Literal["missioncore.xgrids-k1-semantic-topology/v1"]
+    revision: int
+    transport_ref: str
+    connection_mode: TopologyConnectionMode
+    ipv4: str
+    compatibility_profile_id: str
+    firmware_version: str
+    source: TopologyEvidenceSource
+    observed_at_utc: str
+
+    @property
+    def live_connection_authority(self) -> Literal[False]:
+        return False
+
+    def as_dict(self) -> dict[str, object]:
+        return {
+            "schema_version": self.schema_version,
+            "revision": self.revision,
+            "transport_ref": self.transport_ref,
+            "connection_mode": self.connection_mode,
+            "ipv4": self.ipv4,
+            "compatibility_profile_id": self.compatibility_profile_id,
+            "firmware_version": self.firmware_version,
+            "source": self.source,
+            "observed_at_utc": self.observed_at_utc,
+        }
+
+
+@dataclass(frozen=True, slots=True)
+class SemanticTopologySnapshot:
+    status: SemanticTopologyStoreStatus
+    record: SemanticTopologyRecord | None
+    reason_code: str | None
+
+    @property
+    def configured_offline_evidence(self) -> bool:
+        return self.status == "available" and self.record is not None
+
+    @property
+    def live_connection_authority(self) -> Literal[False]:
+        return False
+
+    def as_dict(self) -> dict[str, object]:
+        """Project durable evidence without implying a live connection."""
+
+        return {
+            "schema_version": SEMANTIC_TOPOLOGY_SCHEMA,
+            "status": self.status,
+            "configured_offline_evidence": self.configured_offline_evidence,
+            "live_connection_authority": self.live_connection_authority,
+            "reason_code": self.reason_code,
+            "record": self.record.as_dict() if self.record is not None else None,
+        }
+
+
+class SemanticTopologyStore:
+    """Private, atomic store for the last semantically proved K1 topology.
+
+    The stable flock file serializes the complete reload/check/publish
+    transaction across backend processes.  The store remains independent from
+    the network-mutation ledger: the ledger fences possible writes, while this
+    store retains only a successfully decoded, secret-free status observation.
+    """
+
+    def __init__(self, repository_root: Path) -> None:
+        data_dir = resolve_missioncore_data_dir(repository_root)
+        self.path = data_dir / "xgrids-k1" / SEMANTIC_TOPOLOGY_FILENAME
+        self._lock_path = data_dir / "xgrids-k1" / SEMANTIC_TOPOLOGY_LOCK_FILENAME
+        self._data_dir = data_dir
+        self._thread_lock = threading.RLock()
+        self._record: SemanticTopologyRecord | None = None
+        self._corrupt = False
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+
+    def snapshot(self) -> SemanticTopologySnapshot:
+        """Return configured/offline evidence; never return live authority."""
+
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt:
+                return SemanticTopologySnapshot(
+                    status="corrupt",
+                    record=None,
+                    reason_code=SemanticTopologyStoreCorrupt.reason_code,
+                )
+            if self._record is None:
+                return SemanticTopologySnapshot(status="empty", record=None, reason_code=None)
+            return SemanticTopologySnapshot(
+                status="available",
+                record=self._record,
+                reason_code=None,
+            )
+
+    def commit(
+        self,
+        *,
+        transport_ref: str,
+        connection_mode: TopologyConnectionMode,
+        ipv4: str,
+        compatibility_profile_id: str,
+        firmware_version: str,
+        source: TopologyEvidenceSource,
+        observed_at_utc: str,
+        predecessor_revision: int | None = None,
+    ) -> SemanticTopologyRecord:
+        """Atomically publish one exact status observation.
+
+        Callers must invoke this only after the decoded BLE status has proved
+        the topology and after the selected transport's compatibility profile
+        has been attested.  ``observed_at_utc`` is audit metadata and never
+        orders writes; serialized revision publication does.  An asynchronous
+        caller that needs a stale-writer fence supplies the revision from which
+        its observation descends.  No network mutation is performed here.
+        """
+
+        _validate_identifier(transport_ref, field_name="transport_ref")
+        _validate_connection_mode(connection_mode)
+        canonical_ipv4 = _canonical_ipv4(ipv4)
+        _validate_identifier(
+            compatibility_profile_id,
+            field_name="compatibility_profile_id",
+        )
+        _validate_firmware_version(firmware_version)
+        _validate_source(source)
+        observed_at = _validated_timestamp(observed_at_utc, field_name="observed_at_utc")
+        canonical_observed_at = _timestamp(observed_at)
+        if predecessor_revision is not None:
+            _nonnegative_revision(predecessor_revision, field_name="predecessor_revision")
+
+        with self._thread_lock, self._process_lock_locked():
+            self._reload_locked()
+            if self._corrupt:
+                raise SemanticTopologyStoreCorrupt(
+                    "semantic topology store is corrupt; persisted evidence was not replaced"
+                )
+            current = self._record
+            current_revision = current.revision if current is not None else 0
+            if (
+                predecessor_revision is not None
+                and predecessor_revision != current_revision
+            ):
+                raise StaleSemanticTopologyObservation(
+                    "semantic topology predecessor revision is no longer current"
+                )
+            if current is not None and current.revision >= SEMANTIC_TOPOLOGY_MAX_REVISION:
+                raise SemanticTopologyStoreCorrupt(
+                    "semantic topology revision is exhausted"
+                )
+            revision = 1 if current is None else current.revision + 1
+            record = SemanticTopologyRecord(
+                schema_version=SEMANTIC_TOPOLOGY_SCHEMA,
+                revision=revision,
+                transport_ref=transport_ref,
+                connection_mode=connection_mode,
+                ipv4=canonical_ipv4,
+                compatibility_profile_id=compatibility_profile_id,
+                firmware_version=firmware_version,
+                source=source,
+                observed_at_utc=canonical_observed_at,
+            )
+            self._persist_locked(record)
+            return record
+
+    @contextmanager
+    def _process_lock_locked(self) -> Iterator[None]:
+        data_dir_created = _ensure_private_directory(self._data_dir, parents=True)
+        if data_dir_created:
+            _fsync_directory(self._data_dir.parent)
+        store_dir_created = _ensure_private_directory(self.path.parent, parents=False)
+        if store_dir_created:
+            _fsync_directory(self._data_dir)
+
+        flags = os.O_RDWR | getattr(os, "O_CLOEXEC", 0)
+        flags |= getattr(os, "O_NOFOLLOW", 0)
+        try:
+            descriptor = os.open(self._lock_path, flags)
+            lock_created = False
+        except FileNotFoundError:
+            try:
+                descriptor = os.open(
+                    self._lock_path,
+                    flags | os.O_CREAT | os.O_EXCL,
+                    0o600,
+                )
+                lock_created = True
+            except FileExistsError:
+                # A peer may win the create race. Open and validate exactly
+                # that stable inode instead of introducing a second lock.
+                try:
+                    descriptor = os.open(self._lock_path, flags)
+                    lock_created = False
+                except OSError as exc:
+                    raise SemanticTopologyStoreCorrupt(
+                        "semantic topology lock cannot be opened safely"
+                    ) from exc
+            except OSError as exc:
+                raise SemanticTopologyStoreCorrupt(
+                    "semantic topology lock cannot be created safely"
+                ) from exc
+        except OSError as exc:
+            raise SemanticTopologyStoreCorrupt(
+                "semantic topology lock cannot be opened safely"
+            ) from exc
+        stream: IO[bytes] | None = None
+        try:
+            try:
+                _validate_private_open_file(
+                    descriptor,
+                    self._lock_path,
+                    label="semantic topology lock",
+                    require_empty=True,
+                )
+            except ValueError as exc:
+                raise SemanticTopologyStoreCorrupt(
+                    "semantic topology lock is not a stable private file"
+                ) from exc
+            stream = os.fdopen(descriptor, "r+b", closefd=True)
+            descriptor = -1
+            fcntl.flock(stream.fileno(), fcntl.LOCK_EX)
+            try:
+                try:
+                    _validate_private_open_file(
+                        stream.fileno(),
+                        self._lock_path,
+                        label="semantic topology lock",
+                        require_empty=True,
+                    )
+                except ValueError as exc:
+                    raise SemanticTopologyStoreCorrupt(
+                        "semantic topology lock changed while being acquired"
+                    ) from exc
+                if lock_created:
+                    # The lock file is never replaced. Publish its first
+                    # directory entry before relying on it after a restart.
+                    _fsync_directory(self.path.parent)
+                yield
+            finally:
+                fcntl.flock(stream.fileno(), fcntl.LOCK_UN)
+        finally:
+            if stream is not None:
+                stream.close()
+            elif descriptor >= 0:
+                os.close(descriptor)
+
+    def _persist_locked(self, record: SemanticTopologyRecord) -> None:
+        _write_private_json_atomic(
+            self.path,
+            record.as_dict(),
+            data_dir=self._data_dir,
+        )
+        self._record = record
+        self._corrupt = False
+
+    def _reload_locked(self) -> None:
+        try:
+            payload = _read_private_json(self.path)
+        except FileNotFoundError:
+            self._record = None
+            self._corrupt = False
+            return
+        except (OSError, UnicodeError, json.JSONDecodeError, TypeError, ValueError):
+            self._record = None
+            self._corrupt = True
+            return
+        try:
+            record = _record_from_mapping(payload)
+        except (TypeError, ValueError):
+            self._record = None
+            self._corrupt = True
+            return
+        self._record = record
+        self._corrupt = False
+
+
+def _read_private_json(path: Path) -> object:
+    flags = os.O_RDONLY | getattr(os, "O_CLOEXEC", 0)
+    flags |= getattr(os, "O_NOFOLLOW", 0)
+    try:
+        descriptor = os.open(path, flags)
+    except FileNotFoundError:
+        raise
+    except OSError as exc:
+        raise ValueError("semantic topology file cannot be opened safely") from exc
+    try:
+        metadata = _validate_private_open_file(
+            descriptor,
+            path,
+            label="semantic topology file",
+            require_empty=False,
+        )
+        if metadata.st_size > SEMANTIC_TOPOLOGY_MAX_BYTES:
+            raise ValueError("semantic topology file exceeds the bounded size")
+        chunks: list[bytes] = []
+        remaining = SEMANTIC_TOPOLOGY_MAX_BYTES + 1
+        while remaining > 0:
+            chunk = os.read(descriptor, min(remaining, 64 * 1024))
+            if not chunk:
+                break
+            chunks.append(chunk)
+            remaining -= len(chunk)
+        raw = b"".join(chunks)
+        if len(raw) > SEMANTIC_TOPOLOGY_MAX_BYTES:
+            raise ValueError("semantic topology file exceeds the bounded size")
+    finally:
+        os.close(descriptor)
+    return json.loads(raw.decode("utf-8"), object_pairs_hook=_unique_json_object)
+
+
+def _write_private_json_atomic(
+    path: Path,
+    payload: Mapping[str, object],
+    *,
+    data_dir: Path,
+) -> None:
+    serialized = (
+        json.dumps(
+            payload,
+            ensure_ascii=False,
+            sort_keys=True,
+            separators=(",", ":"),
+            allow_nan=False,
+        )
+        + "\n"
+    ).encode("utf-8")
+    if len(serialized) > SEMANTIC_TOPOLOGY_MAX_BYTES:
+        raise ValueError("semantic topology payload exceeds the bounded size")
+
+    _ensure_private_directory(data_dir, parents=True)
+    _ensure_private_directory(path.parent, parents=False)
+    previous_identity = _existing_private_file_identity(path)
+
+    descriptor, temp_name = tempfile.mkstemp(
+        dir=path.parent,
+        prefix=f".{path.name}.",
+        suffix=".tmp",
+    )
+    temp_path = Path(temp_name)
+    try:
+        os.fchmod(descriptor, 0o600)
+        with os.fdopen(descriptor, "wb") as stream:
+            descriptor = -1
+            stream.write(serialized)
+            stream.flush()
+            os.fsync(stream.fileno())
+        _require_unchanged_existing_path(path, previous_identity)
+        os.replace(temp_path, path)
+        _fsync_directory(path.parent)
+    finally:
+        if descriptor >= 0:
+            os.close(descriptor)
+        temp_path.unlink(missing_ok=True)
+
+
+def _existing_private_file_identity(path: Path) -> tuple[int, int] | None:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        return None
+    _validate_private_metadata(metadata, label="semantic topology file", require_empty=False)
+    return metadata.st_dev, metadata.st_ino
+
+
+def _require_unchanged_existing_path(
+    path: Path,
+    expected: tuple[int, int] | None,
+) -> None:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        if expected is None:
+            return
+        raise ValueError("semantic topology file disappeared during publication") from None
+    _validate_private_metadata(metadata, label="semantic topology file", require_empty=False)
+    observed = (metadata.st_dev, metadata.st_ino)
+    if expected is None or observed != expected:
+        raise ValueError("semantic topology file changed during publication")
+
+
+def _validate_private_open_file(
+    descriptor: int,
+    path: Path,
+    *,
+    label: str,
+    require_empty: bool,
+) -> os.stat_result:
+    metadata = os.fstat(descriptor)
+    _validate_private_metadata(metadata, label=label, require_empty=require_empty)
+    _validate_open_path_identity(descriptor, path, metadata, label=label)
+    return metadata
+
+
+def _validate_open_path_identity(
+    descriptor: int,
+    path: Path,
+    metadata: os.stat_result,
+    *,
+    label: str,
+) -> None:
+    del descriptor  # metadata already came from this exact descriptor
+    try:
+        path_metadata = path.lstat()
+    except OSError as exc:
+        raise ValueError(f"{label} path cannot be verified") from exc
+    if (path_metadata.st_dev, path_metadata.st_ino) != (metadata.st_dev, metadata.st_ino):
+        raise ValueError(f"{label} path does not reference the opened inode")
+    _validate_private_metadata(path_metadata, label=label, require_empty=False)
+
+
+def _validate_private_metadata(
+    metadata: os.stat_result,
+    *,
+    label: str,
+    require_empty: bool,
+) -> None:
+    if not stat.S_ISREG(metadata.st_mode):
+        raise ValueError(f"{label} is not a regular file")
+    if stat.S_IMODE(metadata.st_mode) != 0o600:
+        raise ValueError(f"{label} is not private")
+    if metadata.st_nlink != 1:
+        raise ValueError(f"{label} has an unsafe hard link")
+    if require_empty and metadata.st_size != 0:
+        raise ValueError(f"{label} must remain empty")
+
+
+def _ensure_private_directory(path: Path, *, parents: bool) -> bool:
+    try:
+        metadata = path.lstat()
+    except FileNotFoundError:
+        try:
+            path.mkdir(mode=0o700, parents=parents, exist_ok=False)
+        except FileExistsError:
+            metadata = path.lstat()
+        else:
+            path.chmod(0o700)
+            return True
+    if not stat.S_ISDIR(metadata.st_mode):
+        raise SemanticTopologyStoreCorrupt(
+            "semantic topology directory is not a regular private directory"
+        )
+    if stat.S_IMODE(metadata.st_mode) != 0o700:
+        raise SemanticTopologyStoreCorrupt(
+            "semantic topology directory permissions are not private"
+        )
+    return False
+
+
+def _fsync_directory(path: Path) -> None:
+    flags = os.O_RDONLY | getattr(os, "O_DIRECTORY", 0)
+    descriptor = os.open(path, flags)
+    try:
+        os.fsync(descriptor)
+    finally:
+        os.close(descriptor)
+
+
+def _unique_json_object(pairs: list[tuple[str, object]]) -> dict[str, object]:
+    document: dict[str, object] = {}
+    for key, value in pairs:
+        if key in document:
+            raise ValueError("semantic topology store contains duplicate fields")
+        document[key] = value
+    return document
+
+
+def _record_from_mapping(value: object) -> SemanticTopologyRecord:
+    document = _exact_mapping(
+        value,
+        {
+            "schema_version",
+            "revision",
+            "transport_ref",
+            "connection_mode",
+            "ipv4",
+            "compatibility_profile_id",
+            "firmware_version",
+            "source",
+            "observed_at_utc",
+        },
+        label="semantic topology",
+    )
+    if document["schema_version"] != SEMANTIC_TOPOLOGY_SCHEMA:
+        raise ValueError("unsupported semantic topology schema")
+    revision = _positive_revision(document["revision"])
+    transport_ref = _required_string(document["transport_ref"], field_name="transport_ref")
+    _validate_identifier(transport_ref, field_name="transport_ref")
+    mode_raw = _required_string(document["connection_mode"], field_name="connection_mode")
+    _validate_connection_mode(mode_raw)
+    source_raw = _required_string(document["source"], field_name="source")
+    _validate_source(source_raw)
+    compatibility_profile_id = _required_string(
+        document["compatibility_profile_id"],
+        field_name="compatibility_profile_id",
+    )
+    _validate_identifier(
+        compatibility_profile_id,
+        field_name="compatibility_profile_id",
+    )
+    firmware_version = _required_string(
+        document["firmware_version"],
+        field_name="firmware_version",
+    )
+    _validate_firmware_version(firmware_version)
+    observed_at = _validated_timestamp(
+        document["observed_at_utc"],
+        field_name="observed_at_utc",
+    )
+    observed_at_utc = _timestamp(observed_at)
+    if document["observed_at_utc"] != observed_at_utc:
+        raise ValueError("observed_at_utc is not canonical")
+    return SemanticTopologyRecord(
+        schema_version=SEMANTIC_TOPOLOGY_SCHEMA,
+        revision=revision,
+        transport_ref=transport_ref,
+        connection_mode=cast(TopologyConnectionMode, mode_raw),
+        ipv4=_canonical_ipv4(_required_string(document["ipv4"], field_name="ipv4")),
+        compatibility_profile_id=compatibility_profile_id,
+        firmware_version=firmware_version,
+        source=cast(TopologyEvidenceSource, source_raw),
+        observed_at_utc=observed_at_utc,
+    )
+
+
+def _exact_mapping(value: object, keys: set[str], *, label: str) -> Mapping[str, object]:
+    if not isinstance(value, dict) or set(value) != keys:
+        raise ValueError(f"{label} does not match the secret-free schema")
+    return cast(Mapping[str, object], value)
+
+
+def _required_string(value: object, *, field_name: str) -> str:
+    if not isinstance(value, str) or not value:
+        raise ValueError(f"{field_name} must be a non-empty string")
+    return value
+
+
+def _validate_identifier(value: str, *, field_name: str) -> None:
+    if not isinstance(value, str) or _SAFE_IDENTIFIER.fullmatch(value) is None:
+        raise ValueError(f"{field_name} is outside the secret-free identifier schema")
+
+
+def _validate_firmware_version(value: str) -> None:
+    if not isinstance(value, str) or _SAFE_FIRMWARE_VERSION.fullmatch(value) is None:
+        raise ValueError("firmware_version is outside the bounded version schema")
+
+
+def _validate_connection_mode(value: str) -> None:
+    if not isinstance(value, str) or value not in _CONNECTION_MODES:
+        raise ValueError("unsupported semantic topology connection mode")
+
+
+def _validate_source(value: str) -> None:
+    if not isinstance(value, str) or value not in _EVIDENCE_SOURCES:
+        raise ValueError("unsupported semantic topology evidence source")
+
+
+def _canonical_ipv4(value: str) -> str:
+    if not isinstance(value, str) or len(value) > 15:
+        raise ValueError("semantic topology address must be canonical IPv4")
+    try:
+        address = ipaddress.ip_address(value)
+    except ValueError as exc:
+        raise ValueError("semantic topology address must be an IPv4 address") from exc
+    if not isinstance(address, ipaddress.IPv4Address) or str(address) != value:
+        raise ValueError("semantic topology address must be canonical IPv4")
+    return value
+
+
+def _positive_revision(value: object) -> int:
+    if (
+        isinstance(value, bool)
+        or not isinstance(value, int)
+        or value < 1
+        or value > SEMANTIC_TOPOLOGY_MAX_REVISION
+    ):
+        raise ValueError("revision must be a bounded positive integer")
+    return value
+
+
+def _nonnegative_revision(value: object, *, field_name: str) -> int:
+    if (
+        isinstance(value, bool)
+        or not isinstance(value, int)
+        or value < 0
+        or value > SEMANTIC_TOPOLOGY_MAX_REVISION
+    ):
+        raise ValueError(f"{field_name} must be a bounded nonnegative integer")
+    return value
+
+
+def _validated_timestamp(value: object, *, field_name: str) -> datetime:
+    raw = _required_string(value, field_name=field_name)
+    if len(raw) > 32 or not raw.endswith("Z"):
+        raise ValueError(f"{field_name} must be UTC")
+    try:
+        parsed = datetime.fromisoformat(raw.removesuffix("Z") + "+00:00")
+    except ValueError as exc:
+        raise ValueError(f"{field_name} is invalid") from exc
+    if parsed.tzinfo is None or parsed.utcoffset() != UTC.utcoffset(parsed):
+        raise ValueError(f"{field_name} must be UTC")
+    return parsed.astimezone(UTC)
+
+
+def _timestamp(value: datetime) -> str:
+    return value.astimezone(UTC).isoformat(timespec="milliseconds").replace("+00:00", "Z")

src/k1link/device_plugins/xgrids_k1/viewer/messages.py

@@ -13,3 +13,8 @@ class StreamMessage:
     received_at_epoch_ns: int
     received_monotonic_ns: int | None = None
     source: str = "replay"
+    # Assigned by VisualizationRuntime at the acquisition boundary.  A raw
+    # replay/capture message may be unbound, but an observer must never treat a
+    # message from an older producer generation as evidence for the current
+    # device session.
+    producer_generation: int | None = None

src/k1link/host_network/__init__.py

@@ -1,10 +1,14 @@
 """Host-owned network adapters used by device plugins."""
 
 from k1link.host_network.wifi import (
+    HostWifiAssociationIdentityProbe,
+    HostWifiAssociationIdentityResult,
     HostWifiCredentialMaterialAvailabilityResult,
     HostWifiCredentialMaterialStoreResult,
     HostWifiProfileError,
+    associate_with_ephemeral_wifi_credentials_once,
     associate_with_wifi_profile_once,
+    bind_route_fingerprint_to_wifi_association,
     check_wifi_credential_material,
     check_wifi_profile,
     ensure_wifi_profile_from_credential_source,
@@ -12,10 +16,14 @@ from k1link.host_network.wifi import (
 )
 
 __all__ = [
+    "HostWifiAssociationIdentityProbe",
+    "HostWifiAssociationIdentityResult",
     "HostWifiCredentialMaterialAvailabilityResult",
     "HostWifiCredentialMaterialStoreResult",
     "HostWifiProfileError",
+    "associate_with_ephemeral_wifi_credentials_once",
     "associate_with_wifi_profile_once",
+    "bind_route_fingerprint_to_wifi_association",
     "check_wifi_credential_material",
     "check_wifi_profile",
     "ensure_wifi_profile_from_credential_source",

src/k1link/host_network/wifi.py

@@ -1,15 +1,17 @@
 from __future__ import annotations
 
 import hashlib
+import hmac
 import json
 import os
-import stat
+import secrets
+import signal
 import subprocess
 import sys
-import tempfile
+import threading
 import time
-from collections.abc import Callable, Iterator
-from contextlib import contextmanager, suppress
+from collections.abc import Callable
+from contextlib import suppress
 from pathlib import Path
 from typing import Any, Literal, TypedDict
 
@@ -35,6 +37,7 @@ class HostWifiProfileAvailabilityResult(TypedDict):
     schema_version: int
     adapter: str
     available: bool
+    credential_source: str | None
 
 
 class HostWifiProfileEnsureResult(TypedDict):
@@ -57,6 +60,24 @@ class HostWifiCredentialMaterialAvailabilityResult(TypedDict):
     available: bool
 
 
+class HostWifiAssociationIdentityResult(TypedDict):
+    """Secret-free continuity evidence for one host network interface.
+
+    ``continuity_token`` is scoped to one probe instance and is safe to fold
+    into another local route fingerprint.  It is not an SSID/BSSID digest that
+    can be compared between Mission Core processes.
+    """
+
+    schema_version: int
+    adapter: str
+    wifi_interface: bool | None
+    association_state: Literal["associated", "not-wifi", "unavailable"]
+    evidence_quality: Literal["ssid+bssid", "bssid-only", "not-wifi", "unavailable"]
+    continuity_proven: bool
+    continuity_token: str
+    reason_code: str | None
+
+
 _ERROR_MESSAGES = {
     "unsupported-platform": (
         "для этой ОС ещё не установлен адаптер системных Wi-Fi-профилей"
@@ -68,22 +89,28 @@ _ERROR_MESSAGES = {
         "credential provider точной версии прошивки не установлен на этом устройстве"
     ),
     "network-not-found": (
-        "точка доступа выбранного K1 не найдена; проверьте режим K1 и питание"
+        "K1 подтвердил режим точки доступа, но macOS не увидела её Wi-Fi-сеть "
+        "за отведённое время"
     ),
     "profile-ssid-mismatch": "сохранённый профиль принадлежит другой точке доступа",
-    "credential-entry-cancelled": "ввод пароля точки доступа K1 отменён оператором",
-    "credential-invalid": "пароль точки доступа K1 имеет недопустимую длину",
+    "profile-credential-source-mismatch": (
+        "сохранённый профиль K1 не связан с точной версией firmware provider"
+    ),
     "host-wifi-operation-timeout": (
         "оператор не завершил системное подключение Wi-Fi за отведённое время"
     ),
-    "host-wifi-helper-build-timeout": (
-        "локальный Wi-Fi helper не успел скомпилироваться за отведённое время"
-    ),
-    "host-wifi-helper-build-failed": "локальный Wi-Fi helper не удалось скомпилировать",
-    "host-wifi-helper-compiler-unavailable": "компилятор локального Wi-Fi helper недоступен",
-    "host-wifi-helper-cache-unavailable": "кэш локального Wi-Fi helper недоступен",
     "host-wifi-helper-missing": "исходный файл локального Wi-Fi helper не найден",
     "host-wifi-helper-unavailable": "локальный Wi-Fi helper недоступен",
+    "keychain-authorization-required": (
+        "локальный профиль K1 требует отдельного разрешения связки ключей"
+    ),
+    "keychain-authorization-denied": "доступ к локальному профилю K1 запрещён",
+    "keychain-authorization-cancelled": "разрешение связки ключей отменено оператором",
+    "keychain-access-failed": "локальный профиль K1 недоступен в связке ключей",
+    "corewlan-authorization-denied": (
+        "macOS не разрешила локальному сервису выполнять системную Wi-Fi-операцию"
+    ),
+    "wifi-interface-inactive": "системный Wi-Fi-интерфейс выключен или неактивен",
     "wifi-interface-unavailable": "системный Wi-Fi-интерфейс недоступен",
     "corewlan-error": "системный Wi-Fi не смог подключиться к точке доступа K1",
 }
@@ -111,7 +138,11 @@ class HostWifiProfileError(RuntimeError):
 
 
 RunProcess = Callable[..., subprocess.CompletedProcess[bytes]]
-DEFAULT_HELPER_BUILD_TIMEOUT_SECONDS = 120.0
+REVIEWED_MACOS_HELPER_COMMAND_PREFIX = ("/usr/bin/xcrun", "swift")
+_ASSOCIATION_INSPECTION_PROFILE_ID = "host-association-inspection.v1"
+_EPHEMERAL_ASSOCIATION_PROFILE_ID = "bridge-operation-memory.v1"
+_ASSOCIATION_IDENTITY_HEX_LENGTH = 64
+MAX_ASSOCIATION_OBSERVATION_CACHE_AGE_SECONDS = 0.75
 
 
 def _validate_profile_id(profile_id: str) -> None:
@@ -124,232 +155,45 @@ def _validate_profile_id(profile_id: str) -> None:
         raise ValueError("host Wi-Fi profile id contains unsupported characters")
 
 
-def _helper_cache_directory(helper_path: Path) -> Path:
-    """Resolve the process-local helper cache without an environment override."""
-
-    source = helper_path.expanduser().resolve()
-    for ancestor in source.parents:
-        if ancestor.name != "plugins":
-            continue
-        try:
-            relative = source.relative_to(ancestor)
-        except ValueError:  # pragma: no cover - guarded by Path.parents
-            continue
-        if relative.parts[:2] == ("xgrids-k1", "macos"):
-            return ancestor.parent / ".runtime" / "mission-core" / "helpers"
-    # Tests and separately packaged adapters still get a stable cache beside
-    # their source tree. The repository layout above is the production path.
-    return source.parent / ".runtime" / "mission-core" / "helpers"
-
-
-def _compiled_macos_helper_path(helper_path: Path) -> Path:
-    source = helper_path.expanduser().resolve()
-    source_sha256 = hashlib.sha256(source.read_bytes()).hexdigest()
-    return _helper_cache_directory(source) / f"{source.stem}-{source_sha256}"
-
-
-def _is_ready_executable(path: Path) -> bool:
-    try:
-        metadata = path.lstat()
-    except OSError:
-        return False
-    return (
-        stat.S_ISREG(metadata.st_mode)
-        and metadata.st_size > 0
-        and metadata.st_mode & 0o111 != 0
-    )
-
-
-@contextmanager
-def _exclusive_helper_build_lock(
-    path: Path,
-    *,
-    timeout_seconds: float,
-) -> Iterator[None]:
-    if timeout_seconds <= 0:
-        raise ValueError("timeout_seconds must be positive")
-    try:
-        import fcntl
-    except ImportError as exc:  # pragma: no cover - production target is macOS
-        raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-
-    flags = (
-        os.O_RDWR
-        | os.O_CREAT
-        | getattr(os, "O_CLOEXEC", 0)
-        | getattr(os, "O_NOFOLLOW", 0)
-    )
-    try:
-        descriptor = os.open(path, flags, 0o600)
-    except OSError as exc:
-        raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-    locked = False
-    try:
-        if not stat.S_ISREG(os.fstat(descriptor).st_mode):
-            raise HostWifiProfileError("host-wifi-helper-cache-unavailable")
-        lock_started = time.monotonic()
-        while True:
-            try:
-                fcntl.flock(descriptor, fcntl.LOCK_EX | fcntl.LOCK_NB)
-                break
-            except BlockingIOError as exc:
-                elapsed_seconds = max(0.0, time.monotonic() - lock_started)
-                if elapsed_seconds >= timeout_seconds:
-                    raise HostWifiProfileError(
-                        "host-wifi-helper-build-timeout",
-                        helper_stage="compile-lock",
-                        helper_elapsed_ms=int(elapsed_seconds * 1000),
-                    ) from exc
-                time.sleep(min(0.05, timeout_seconds - elapsed_seconds))
-            except OSError as exc:
-                raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-        locked = True
-        yield
-    finally:
-        if locked:
-            with suppress(OSError):
-                fcntl.flock(descriptor, fcntl.LOCK_UN)
-        os.close(descriptor)
-
-
-def _fsync_directory(path: Path) -> None:
-    try:
-        descriptor = os.open(path, os.O_RDONLY | getattr(os, "O_DIRECTORY", 0))
-    except OSError:
-        return
-    try:
-        os.fsync(descriptor)
-    finally:
-        os.close(descriptor)
-
-
-def _ensure_compiled_macos_helper(
-    helper_path: Path,
-    *,
-    build_timeout_seconds: float,
-    runner: RunProcess,
-) -> Path:
-    """Build the source-hash-addressed helper once and reuse it thereafter."""
-
-    if build_timeout_seconds <= 0:
-        raise ValueError("build_timeout_seconds must be positive")
-    source = helper_path.expanduser().resolve()
-    if not source.is_file():
-        raise HostWifiProfileError("host-wifi-helper-missing")
-    try:
-        executable = _compiled_macos_helper_path(source)
-        executable.parent.mkdir(mode=0o700, parents=True, exist_ok=True)
-    except OSError as exc:
-        raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-    if _is_ready_executable(executable):
-        return executable
-
-    build_started_ns = time.monotonic_ns()
-    build_deadline = time.monotonic() + build_timeout_seconds
-
-    def build_error(
-        reason_code: str,
-        *,
-        helper_stage: str = "compile",
-    ) -> HostWifiProfileError:
-        elapsed_ms = max(0, (time.monotonic_ns() - build_started_ns) // 1_000_000)
-        return HostWifiProfileError(
-            reason_code,
-            helper_stage=helper_stage,
-            helper_elapsed_ms=elapsed_ms,
-        )
-
-    lock_path = executable.with_name(f".{executable.name}.lock")
-    with _exclusive_helper_build_lock(
-        lock_path,
-        timeout_seconds=max(0.001, build_deadline - time.monotonic()),
-    ):
-        if _is_ready_executable(executable):
-            return executable
-
-        try:
-            descriptor, staging_name = tempfile.mkstemp(
-                prefix=f".{executable.name}.",
-                suffix=".tmp",
-                dir=executable.parent,
-            )
-            os.close(descriptor)
-            staging = Path(staging_name)
-            staging.unlink()
-        except OSError as exc:
-            raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-
-        try:
-            remaining_build_seconds = build_deadline - time.monotonic()
-            if remaining_build_seconds <= 0:
-                raise build_error("host-wifi-helper-build-timeout")
-
-            try:
-                completed = runner(
-                    [
-                        "/usr/bin/xcrun",
-                        "swiftc",
-                        str(source),
-                        "-o",
-                        str(staging),
-                    ],
-                    stdin=subprocess.DEVNULL,
-                    stdout=subprocess.PIPE,
-                    stderr=subprocess.PIPE,
-                    timeout=remaining_build_seconds,
-                    check=False,
-                )
-            except subprocess.TimeoutExpired as exc:
-                raise build_error("host-wifi-helper-build-timeout") from exc
-            except OSError as exc:
-                raise build_error("host-wifi-helper-compiler-unavailable") from exc
-
-            if completed.returncode != 0 or not _is_ready_executable(staging):
-                raise build_error("host-wifi-helper-build-failed")
-            try:
-                staging.chmod(0o700)
-                with staging.open("rb") as stream:
-                    os.fsync(stream.fileno())
-                os.replace(staging, executable)
-                _fsync_directory(executable.parent)
-            except OSError as exc:
-                raise HostWifiProfileError("host-wifi-helper-cache-unavailable") from exc
-        finally:
-            with suppress(OSError):
-                staging.unlink(missing_ok=True)
-    return executable
-
-
 def _run_macos_helper(
     helper_path: Path,
     request: dict[str, object],
     *,
     timeout_seconds: float,
     runner: RunProcess,
-    build_timeout_seconds: float = DEFAULT_HELPER_BUILD_TIMEOUT_SECONDS,
+    process_fence_descriptor_factory: Callable[[], int] | None = None,
 ) -> dict[str, Any]:
     if sys.platform != "darwin":
         raise HostWifiProfileError("unsupported-platform")
+    source = helper_path.expanduser().resolve()
+    if not source.is_file():
+        raise HostWifiProfileError("host-wifi-helper-missing")
     if timeout_seconds <= 0:
         raise ValueError("timeout_seconds must be positive")
 
-    executable = _ensure_compiled_macos_helper(
-        helper_path,
-        build_timeout_seconds=build_timeout_seconds,
-        runner=runner,
-    )
-
     request_bytes = bytearray(
         json.dumps(request, ensure_ascii=False, separators=(",", ":")).encode("utf-8")
     )
     try:
-        completed = runner(
-            [str(executable)],
-            input=request_bytes,
-            stdout=subprocess.PIPE,
-            stderr=subprocess.PIPE,
-            timeout=timeout_seconds,
-            check=False,
+        command = [*REVIEWED_MACOS_HELPER_COMMAND_PREFIX, str(source)]
+        completed = (
+            _run_macos_helper_with_process_fence(
+                command,
+                request_bytes=request_bytes,
+                timeout_seconds=timeout_seconds,
+                process_fence_descriptor_factory=(
+                    process_fence_descriptor_factory
+                ),
+            )
+            if process_fence_descriptor_factory is not None
+            else runner(
+                command,
+                input=request_bytes,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                timeout=timeout_seconds,
+                check=False,
+            )
         )
     except subprocess.TimeoutExpired as exc:
         raise HostWifiProfileError("host-wifi-operation-timeout") from exc
@@ -389,22 +233,422 @@ def _run_macos_helper(
     return response
 
 
+def _run_macos_helper_with_process_fence(
+    command: list[str],
+    *,
+    request_bytes: bytearray,
+    timeout_seconds: float,
+    process_fence_descriptor_factory: Callable[[], int],
+) -> subprocess.CompletedProcess[bytes]:
+    """Run one mutating helper in a fenced, killable POSIX process group."""
+
+    if os.name != "posix":
+        raise HostWifiProfileError("host-wifi-helper-unavailable")
+    inherited_descriptor = process_fence_descriptor_factory()
+    process: subprocess.Popen[bytes] | None = None
+    try:
+        process = subprocess.Popen(
+            command,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            start_new_session=True,
+            pass_fds=(inherited_descriptor,),
+        )
+    finally:
+        # Popen has either inherited the descriptor into the child or failed.
+        # The parent-side duplicate must never extend ownership on its own.
+        os.close(inherited_descriptor)
+    try:
+        stdout, stderr = process.communicate(
+            # subprocess accepts any bytes-like object at runtime. Keep the
+            # mutable buffer so the caller can zero credentials in-place.
+            input=request_bytes,  # type: ignore[arg-type]
+            timeout=timeout_seconds,
+        )
+    except BaseException:
+        _kill_and_reap_process_group(process)
+        raise
+    return subprocess.CompletedProcess(
+        args=command,
+        returncode=process.returncode,
+        stdout=stdout,
+        stderr=stderr,
+    )
+
+
+def _kill_and_reap_process_group(process: subprocess.Popen[bytes]) -> None:
+    """Prove the helper and every descendant are gone before dropping a fence."""
+
+    with suppress(ProcessLookupError, PermissionError):
+        os.killpg(process.pid, signal.SIGKILL)
+    try:
+        process.communicate(timeout=5.0)
+    except subprocess.TimeoutExpired:
+        with suppress(OSError):
+            process.kill()
+        process.wait(timeout=5.0)
+
+
+class HostWifiAssociationIdentityProbe:
+    """Observe macOS Wi-Fi association continuity without exporting its name.
+
+    The Swift helper HMACs its in-process SSID/BSSID evidence with a random key
+    supplied over stdin.  This object retains that key only for its own process
+    lifetime, so equal networks cannot be correlated across service restarts.
+
+    When CoreWLAN cannot expose an association (for example because privacy
+    authorization hides both identifiers), observations for the same
+    interface/failure scope receive one process-scoped fallback token.  The
+    token is not association proof and callers must keep that distinction, but
+    it prevents an unchanged kernel route from being misclassified as a new
+    network on every monitor poll.  Route loss, interface/source changes,
+    recovery of exact BSSID evidence, or a service restart still changes the
+    combined route fingerprint.
+    """
+
+    def __init__(
+        self,
+        helper_path: Path,
+        *,
+        runner: RunProcess = subprocess.run,
+        continuity_key: bytes | None = None,
+        max_cache_age_seconds: float = MAX_ASSOCIATION_OBSERVATION_CACHE_AGE_SECONDS,
+        monotonic_clock: Callable[[], float] = time.monotonic,
+        wall_clock: Callable[[], float] = time.time,
+    ) -> None:
+        key = secrets.token_bytes(32) if continuity_key is None else bytes(continuity_key)
+        if len(key) != 32:
+            raise ValueError("host Wi-Fi continuity key must contain exactly 32 bytes")
+        if not 0 <= max_cache_age_seconds <= MAX_ASSOCIATION_OBSERVATION_CACHE_AGE_SECONDS:
+            raise ValueError(
+                "host Wi-Fi association cache age must be between 0 and 0.75 seconds"
+            )
+        self._helper_path = helper_path
+        self._runner = runner
+        self._continuity_key = key
+        self._max_cache_age_seconds = max_cache_age_seconds
+        self._monotonic_clock = monotonic_clock
+        self._wall_clock = wall_clock
+        self._cached_observations: dict[
+            str,
+            tuple[float, float, HostWifiAssociationIdentityResult],
+        ] = {}
+        self._last_interface_name: str | None = None
+        self._fallback_tokens: dict[str, str] = {}
+        self._lock = threading.Lock()
+
+    def observe(
+        self,
+        interface_name: str | None,
+        *,
+        timeout_seconds: float = 30.0,
+    ) -> HostWifiAssociationIdentityResult:
+        """Return one opaque association token, failing closed when uncertain."""
+
+        if timeout_seconds <= 0:
+            raise ValueError("timeout_seconds must be positive")
+        if interface_name is not None and (
+            not 1 <= len(interface_name) <= 32
+            or not all(
+                character.isascii()
+                and (character.isalnum() or character in "._-")
+                for character in interface_name
+            )
+        ):
+            raise ValueError("host Wi-Fi interface name contains unsupported characters")
+
+        deadline = time.monotonic() + timeout_seconds
+        if not self._lock.acquire(timeout=timeout_seconds):
+            return self._fallback_without_lock(
+                adapter="unavailable",
+                wifi_interface=None,
+                reason_code="host-wifi-operation-timeout",
+                scope=f"interface:{interface_name or 'none'}",
+            )
+        try:
+            if interface_name is None:
+                self._invalidate_for_interface_transition_locked(None)
+                return self._fallback_locked(
+                    adapter="unavailable",
+                    wifi_interface=None,
+                    reason_code="host-route-interface-unavailable",
+                    scope="interface:none",
+                )
+            self._invalidate_for_interface_transition_locked(interface_name)
+            cached = self._cached_observation_locked(interface_name)
+            if cached is not None:
+                return cached
+            remaining_seconds = deadline - time.monotonic()
+            if remaining_seconds <= 0:
+                return self._remember_observation_locked(
+                    interface_name,
+                    self._fallback_locked(
+                        adapter="unavailable",
+                        wifi_interface=None,
+                        reason_code="host-wifi-operation-timeout",
+                        scope=f"interface:{interface_name}",
+                    ),
+                )
+            try:
+                response = _run_macos_helper(
+                    self._helper_path,
+                    {
+                        "action": "inspect-association",
+                        "profile_id": _ASSOCIATION_INSPECTION_PROFILE_ID,
+                        "interface_name": interface_name,
+                        "continuity_key_hex": self._continuity_key.hex(),
+                    },
+                    timeout_seconds=remaining_seconds,
+                    runner=self._runner,
+                )
+            except HostWifiProfileError as exc:
+                return self._remember_observation_locked(
+                    interface_name,
+                    self._fallback_locked(
+                        adapter="unavailable",
+                        wifi_interface=None,
+                        reason_code=exc.reason_code,
+                        scope=f"interface:{interface_name}",
+                    ),
+                )
+
+            adapter = response.get("adapter")
+            wifi_interface = response.get("wifi_interface")
+            association_identity = response.get("association_identity")
+            evidence_quality = response.get("association_evidence")
+            reason_code = response.get("reason_code")
+            if (
+                not isinstance(adapter, str)
+                or not adapter
+                or not isinstance(wifi_interface, bool)
+                or evidence_quality
+                not in {"ssid+bssid", "bssid-only", "not-wifi", "unavailable"}
+                or (reason_code is not None and not isinstance(reason_code, str))
+            ):
+                return self._remember_observation_locked(
+                    interface_name,
+                    self._fallback_locked(
+                        adapter="unavailable",
+                        wifi_interface=None,
+                        reason_code="host-wifi-response-invalid",
+                        scope=f"interface:{interface_name}",
+                    ),
+                )
+
+            continuity_proven = evidence_quality in {
+                "ssid+bssid",
+                "bssid-only",
+                "not-wifi",
+            }
+            if not continuity_proven:
+                return self._remember_observation_locked(
+                    interface_name,
+                    self._fallback_locked(
+                        adapter=adapter,
+                        wifi_interface=wifi_interface,
+                        reason_code=reason_code or "association-identity-unavailable",
+                        scope=f"interface:{interface_name}",
+                    ),
+                )
+            if (
+                not isinstance(association_identity, str)
+                or len(association_identity) != _ASSOCIATION_IDENTITY_HEX_LENGTH
+                or any(character not in "0123456789abcdef" for character in association_identity)
+            ):
+                return self._remember_observation_locked(
+                    interface_name,
+                    self._fallback_locked(
+                        adapter="unavailable",
+                        wifi_interface=wifi_interface,
+                        reason_code="host-wifi-response-invalid",
+                        scope=f"interface:{interface_name}",
+                    ),
+                )
+            return self._remember_observation_locked(
+                interface_name,
+                {
+                    "schema_version": 1,
+                    "adapter": adapter,
+                    "wifi_interface": wifi_interface,
+                    "association_state": (
+                        "not-wifi" if evidence_quality == "not-wifi" else "associated"
+                    ),
+                    "evidence_quality": evidence_quality,
+                    "continuity_proven": True,
+                    "continuity_token": association_identity,
+                    "reason_code": None,
+                },
+            )
+        finally:
+            self._lock.release()
+
+    def _invalidate_for_interface_transition_locked(
+        self,
+        interface_name: str | None,
+    ) -> None:
+        """Make an interface transition an immediate cache barrier."""
+
+        if interface_name == self._last_interface_name:
+            return
+        self._cached_observations.clear()
+        self._last_interface_name = interface_name
+
+    def _cached_observation_locked(
+        self,
+        interface_name: str,
+    ) -> HostWifiAssociationIdentityResult | None:
+        """Reuse only a sub-second read-only association observation.
+
+        Runtime compilation is deliberately not used here: the field-approved
+        credential/CoreWLAN path relies on Apple's signed Swift interpreter
+        identity.  The cache is therefore limited to the read-only association
+        result, never a credential or mutation result.  Both clocks must remain
+        monotonic within the 750 ms safety window; sleep, a wall-clock jump or
+        an interface transition forces a fresh source-runner observation before
+        the caller can publish.
+        """
+
+        cached = self._cached_observations.get(interface_name)
+        if cached is None or self._max_cache_age_seconds == 0:
+            return None
+        observed_monotonic, observed_wall, observation = cached
+        monotonic_age = self._monotonic_clock() - observed_monotonic
+        wall_age = self._wall_clock() - observed_wall
+        if (
+            monotonic_age < 0
+            or wall_age < 0
+            or monotonic_age > self._max_cache_age_seconds
+            or wall_age > self._max_cache_age_seconds
+        ):
+            self._cached_observations.pop(interface_name, None)
+            return None
+        return observation.copy()
+
+    def _remember_observation_locked(
+        self,
+        interface_name: str,
+        observation: HostWifiAssociationIdentityResult,
+    ) -> HostWifiAssociationIdentityResult:
+        if self._max_cache_age_seconds > 0:
+            self._cached_observations[interface_name] = (
+                self._monotonic_clock(),
+                self._wall_clock(),
+                observation.copy(),
+            )
+        return observation.copy()
+
+    def _fallback_locked(
+        self,
+        *,
+        adapter: str,
+        wifi_interface: bool | None,
+        reason_code: str,
+        scope: str,
+    ) -> HostWifiAssociationIdentityResult:
+        token_scope = "\x1f".join(
+            (
+                scope,
+                adapter,
+                "unknown" if wifi_interface is None else str(wifi_interface).lower(),
+                reason_code,
+            )
+        )
+        token = self._fallback_tokens.get(token_scope)
+        if token is None:
+            material = (
+                "mission-core/host-wifi-association-fallback/v2\x1f"
+                f"{token_scope}"
+            ).encode()
+            token = hmac.new(self._continuity_key, material, hashlib.sha256).hexdigest()
+            self._fallback_tokens[token_scope] = token
+        return {
+            "schema_version": 1,
+            "adapter": adapter,
+            "wifi_interface": wifi_interface,
+            "association_state": "unavailable",
+            "evidence_quality": "unavailable",
+            "continuity_proven": False,
+            "continuity_token": token,
+            "reason_code": reason_code,
+        }
+
+    def _fallback_without_lock(
+        self,
+        *,
+        adapter: str,
+        wifi_interface: bool | None,
+        reason_code: str,
+        scope: str,
+    ) -> HostWifiAssociationIdentityResult:
+        """Build the deterministic process-local fallback under lock contention."""
+
+        token_scope = "\x1f".join(
+            (
+                scope,
+                adapter,
+                "unknown" if wifi_interface is None else str(wifi_interface).lower(),
+                reason_code,
+            )
+        )
+        material = (
+            "mission-core/host-wifi-association-fallback/v2\x1f"
+            f"{token_scope}"
+        ).encode()
+        return {
+            "schema_version": 1,
+            "adapter": adapter,
+            "wifi_interface": wifi_interface,
+            "association_state": "unavailable",
+            "evidence_quality": "unavailable",
+            "continuity_proven": False,
+            "continuity_token": hmac.new(
+                self._continuity_key,
+                material,
+                hashlib.sha256,
+            ).hexdigest(),
+            "reason_code": reason_code,
+        }
+
+
+def bind_route_fingerprint_to_wifi_association(
+    route_fingerprint: str,
+    association: HostWifiAssociationIdentityResult,
+) -> str:
+    """Bind route facts to one opaque association observation."""
+
+    if not route_fingerprint:
+        raise ValueError("route fingerprint must be nonblank")
+    token = association["continuity_token"]
+    if (
+        len(token) != _ASSOCIATION_IDENTITY_HEX_LENGTH
+        or any(character not in "0123456789abcdef" for character in token)
+    ):
+        raise ValueError("association continuity token is invalid")
+    material = (
+        "mission-core/host-route+wifi-association/v1\x1f"
+        f"{route_fingerprint}\x1f{token}"
+    )
+    return hashlib.sha256(material.encode()).hexdigest()
+
+
 def associate_with_wifi_profile_once(
     helper_path: Path,
     profile_id: str,
     expected_ssid: str,
     *,
-    scan_timeout_seconds: float = 15.0,
+    scan_timeout_seconds: float = 30.0,
     timeout_seconds: float = 180.0,
     runner: RunProcess = subprocess.run,
+    process_fence_descriptor_factory: Callable[[], int] | None = None,
 ) -> HostWifiProfileAssociationResult:
     """Associate through one device-scoped OS profile without exposing its secret.
 
-    The selected device supplies the expected, operator-visible AP SSID.  The
+    The selected device supplies the expected, operator-visible AP SSID. The
     platform helper performs bounded exact-SSID discovery followed by at most
-    one association. It resolves a device-scoped secret from the OS credential
-    stores, or asks for it through a native secure prompt on first use. The
-    secret never crosses the helper boundary.
+    one association using the already materialized Mission Core profile. It
+    never falls back to another system credential or a post-write password
+    prompt, and the secret never crosses the helper boundary.
     """
 
     _validate_profile_id(profile_id)
@@ -422,6 +666,7 @@ def associate_with_wifi_profile_once(
         },
         timeout_seconds=timeout_seconds,
         runner=runner,
+        process_fence_descriptor_factory=process_fence_descriptor_factory,
     )
     already_associated = response.get("already_associated")
     profile_enrolled = response.get("profile_enrolled")
@@ -439,12 +684,7 @@ def associate_with_wifi_profile_once(
         or type(scan_elapsed_ms) is not int
         or scan_elapsed_ms < 0
         or not isinstance(credential_source, str)
-        or credential_source not in {
-            "mission-core-keychain",
-            "system-wifi-keychain",
-            "native-secure-prompt",
-            "exact-firmware-profile",
-        }
+        or credential_source != "exact-firmware-profile"
     ):
         raise HostWifiProfileError("host-wifi-response-invalid")
     return {
@@ -459,6 +699,78 @@ def associate_with_wifi_profile_once(
     }
 
 
+def associate_with_ephemeral_wifi_credentials_once(
+    helper_path: Path,
+    expected_ssid: str,
+    password: str,
+    *,
+    scan_timeout_seconds: float = 30.0,
+    timeout_seconds: float = 180.0,
+    runner: RunProcess = subprocess.run,
+    process_fence_descriptor_factory: Callable[[], int] | None = None,
+) -> HostWifiProfileAssociationResult:
+    """Join one operator-selected Bridge network without persisting its secret.
+
+    This host-side handoff is used after K1 has already applied the Bridge
+    topology and the Mac no longer has a direct route to it (most notably when
+    switching from the K1 access point back to the shared network).  The
+    password crosses only the helper's stdin, is zeroed from the mutable Python
+    request buffer, and is neither stored in the Mission Core Keychain profile
+    store nor returned in evidence.
+
+    The call performs at most one CoreWLAN association after a bounded exact
+    SSID scan.  It never repeats the preceding BLE/device mutation.
+    """
+
+    if not 1 <= len(expected_ssid.encode("utf-8")) <= 32:
+        raise ValueError("SSID must contain between 1 and 32 UTF-8 bytes")
+    if not 1 <= len(password.encode("utf-8")) <= 64:
+        raise ValueError("Wi-Fi password must contain between 1 and 64 UTF-8 bytes")
+    if not 0 <= scan_timeout_seconds <= 60:
+        raise ValueError("scan_timeout_seconds must be between 0 and 60")
+    response = _run_macos_helper(
+        helper_path,
+        {
+            "action": "associate-ephemeral",
+            "profile_id": _EPHEMERAL_ASSOCIATION_PROFILE_ID,
+            "ssid": expected_ssid,
+            "password": password,
+            "scan_timeout_seconds": scan_timeout_seconds,
+        },
+        timeout_seconds=timeout_seconds,
+        runner=runner,
+        process_fence_descriptor_factory=process_fence_descriptor_factory,
+    )
+    already_associated = response.get("already_associated")
+    profile_enrolled = response.get("profile_enrolled")
+    scan_attempt_count = response.get("scan_attempt_count")
+    scan_elapsed_ms = response.get("scan_elapsed_ms")
+    credential_source = response.get("credential_source")
+    adapter = response.get("adapter")
+    if (
+        not isinstance(already_associated, bool)
+        or profile_enrolled is not False
+        or not isinstance(adapter, str)
+        or not adapter
+        or type(scan_attempt_count) is not int
+        or scan_attempt_count < 1
+        or type(scan_elapsed_ms) is not int
+        or scan_elapsed_ms < 0
+        or credential_source != "operation-memory"
+    ):
+        raise HostWifiProfileError("host-wifi-response-invalid")
+    return {
+        "schema_version": 1,
+        "adapter": adapter,
+        "outcome": "already-associated" if already_associated else "associated",
+        "already_associated": already_associated,
+        "profile_enrolled": False,
+        "scan_attempt_count": scan_attempt_count,
+        "scan_elapsed_ms": scan_elapsed_ms,
+        "credential_source": "operation-memory",
+    }
+
+
 def check_wifi_profile(
     helper_path: Path,
     profile_id: str,
@@ -483,13 +795,27 @@ def check_wifi_profile(
         runner=runner,
     )
     available = response.get("profile_available")
+    credential_source = response.get("credential_source")
     adapter = response.get("adapter")
-    if not isinstance(available, bool) or not isinstance(adapter, str) or not adapter:
+    if (
+        not isinstance(available, bool)
+        or not isinstance(adapter, str)
+        or not adapter
+        or (
+            available
+            and credential_source != "exact-firmware-profile"
+        )
+        or (
+            not available
+            and credential_source is not None
+        )
+    ):
         raise HostWifiProfileError("host-wifi-response-invalid")
     return {
         "schema_version": 1,
         "adapter": adapter,
         "available": available,
+        "credential_source": credential_source,
     }
 
 
@@ -501,6 +827,7 @@ def ensure_wifi_profile_from_credential_source(
     *,
     timeout_seconds: float = 30.0,
     runner: RunProcess = subprocess.run,
+    process_fence_descriptor_factory: Callable[[], int] | None = None,
 ) -> HostWifiProfileEnsureResult:
     """Materialize one device profile from a firmware-scoped secure-store item.
 
@@ -522,6 +849,7 @@ def ensure_wifi_profile_from_credential_source(
         },
         timeout_seconds=timeout_seconds,
         runner=runner,
+        process_fence_descriptor_factory=process_fence_descriptor_factory,
     )
     available = response.get("profile_available")
     profile_enrolled = response.get("profile_enrolled")

src/k1link/sessions/active.py

@@ -30,6 +30,7 @@ class ActiveSessionLease:
     _descriptor: int
     _marker_identity: tuple[int, int]
     _released: bool = False
+    _marker_removed: bool = False
 
     @classmethod
     def acquire(cls, sessions_root: Path, session_root: Path) -> ActiveSessionLease:
@@ -72,18 +73,21 @@ class ActiveSessionLease:
         if self._released:
             return
         marker = self.sessions_root / ACTIVE_SESSION_MARKER
-        try:
+        if not self._marker_removed:
             try:
                 metadata = marker.lstat()
             except FileNotFoundError:
                 metadata = None
             if metadata is not None and (metadata.st_dev, metadata.st_ino) == self._marker_identity:
                 marker.unlink()
-                _fsync_directory(self.sessions_root)
-        finally:
-            _unlock_descriptor(self._descriptor)
-            os.close(self._descriptor)
-            self._released = True
+            self._marker_removed = True
+        # Keep the descriptor/lock owned until marker retirement is durably
+        # synchronized. If unlink or fsync fails, a later cleanup-only retry
+        # resumes this staged release without reopening command authority.
+        _fsync_directory_strict(self.sessions_root)
+        _unlock_descriptor(self._descriptor)
+        os.close(self._descriptor)
+        self._released = True
 
     def __enter__(self) -> ActiveSessionLease:
         return self
@@ -170,3 +174,13 @@ def _fsync_directory(path: Path) -> None:
         os.fsync(descriptor)
     finally:
         os.close(descriptor)
+
+
+def _fsync_directory_strict(path: Path) -> None:
+    """Synchronize a lease transition or fail while retaining its lock."""
+
+    descriptor = os.open(path, os.O_RDONLY | getattr(os, "O_DIRECTORY", 0))
+    try:
+        os.fsync(descriptor)
+    finally:
+        os.close(descriptor)

src/k1link/viewer/rerun_bridge.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import errno
 import logging
 import math
 import socket
@@ -86,8 +87,19 @@ def _select_available_grpc_port(
                 # Rerun binds all local interfaces. Probe the same address class
                 # so a previous recording retained by a late viewer is detected.
                 probe.bind(("0.0.0.0", candidate))
-            except OSError:
-                continue
+            except OSError as exc:
+                if exc.errno == errno.EADDRINUSE:
+                    continue
+                if exc.errno in {errno.EACCES, errno.EPERM}:
+                    raise PermissionError(
+                        exc.errno,
+                        "Permission denied while probing local Rerun gRPC "
+                        f"port {candidate}",
+                    ) from exc
+                raise RuntimeError(
+                    "Could not probe local Rerun gRPC "
+                    f"port {candidate}: {type(exc).__name__}: {exc}"
+                ) from exc
         return candidate
     raise RuntimeError(
         "No local Rerun gRPC port is available in "

src/k1link/web/app.py

@@ -11,7 +11,6 @@ from typing import Any
 from fastapi import FastAPI, HTTPException, Request, WebSocket, WebSocketDisconnect
 from fastapi.exceptions import RequestValidationError
 from fastapi.responses import JSONResponse
-from fastapi.staticfiles import StaticFiles
 from pydantic import ValidationError
 from starlette.middleware.gzip import GZipMiddleware
 
@@ -76,6 +75,7 @@ from k1link.web.e46j_raw_fisheye_realtime_api import (
 )
 from k1link.web.e47_semantic_slam_api import build_e47_semantic_slam_router
 from k1link.web.environment_api import build_environment_router
+from k1link.web.frontend_assets import ControlStationStaticFiles, frontend_build_id
 from k1link.web.l3_pointpillars_visual_api import (
     build_l3_pointpillars_visual_router,
 )
@@ -385,7 +385,7 @@ async def _recording_preparation_reconciler() -> None:
 async def app_lifespan(_: FastAPI) -> AsyncIterator[None]:
     reconciler: asyncio.Task[None] | None = None
     try:
-        configure_scanner_diagnostics(REPOSITORY_ROOT / ".runtime" / "mission-core" / "logs")
+        configure_scanner_diagnostics(session_store.data_dir / "logs")
         session_recording_preparation_manager.start()
         # Recovery is intentionally a one-shot startup phase. The archive
         # helper owns a cross-process lease, while ordinary catalog requests
@@ -480,11 +480,36 @@ async def invoke_device_plugin_action(
     except ValueError as exc:
         raise HTTPException(status_code=400, detail=str(exc)) from exc
     except PluginExecutionError as exc:
-        raise HTTPException(status_code=502, detail=str(exc)) from exc
+        raise HTTPException(status_code=exc.http_status_code, detail=str(exc)) from exc
     except PluginRuntimeUnavailableError as exc:
         raise HTTPException(status_code=503, detail=str(exc)) from exc
 
 
+_CLOSED_WEBSOCKET_SEND_ERRORS = frozenset(
+    {
+        "handler is closed",
+        'Cannot call "send" once a close message has been sent.',
+        (
+            "Unexpected ASGI message 'websocket.send', after sending "
+            "'websocket.close' or response already completed."
+        ),
+    }
+)
+
+
+def _is_closed_websocket_send_error(exc: RuntimeError) -> bool:
+    message = " ".join(str(exc).split())
+    if message in _CLOSED_WEBSOCKET_SEND_ERRORS:
+        return True
+    # uvloop raises this after the browser has already torn down the TCP
+    # transport, before Starlette can translate the failed send into a
+    # WebSocketDisconnect. Match both stable parts so an unrelated RuntimeError
+    # containing only "closed" is never swallowed.
+    return message.startswith("unable to perform operation on ") and message.endswith(
+        "; the handler is closed"
+    )
+
+
 @app.websocket("/api/v1/device-plugins/{plugin_id}/events")
 async def device_plugin_events(websocket: WebSocket, plugin_id: str) -> None:
     await websocket.accept()
@@ -493,10 +518,17 @@ async def device_plugin_events(websocket: WebSocket, plugin_id: str) -> None:
         while True:
             state = await plugin_dispatcher.invoke(plugin_id, STATE_READ_ACTION_ID, {})
             sequence += 1
-            await websocket.send_json({"pluginId": plugin_id, "sequence": sequence, "state": state})
+            try:
+                await websocket.send_json(
+                    {"pluginId": plugin_id, "sequence": sequence, "state": state}
+                )
+            except WebSocketDisconnect:
+                return
+            except RuntimeError as exc:
+                if _is_closed_websocket_send_error(exc):
+                    return
+                raise
             await asyncio.sleep(0.5)
-    except WebSocketDisconnect:
-        return
     except (PluginNotFoundError, PluginActionNotFoundError):
         await websocket.close(code=1008, reason="Device plugin is not available")
     except (PluginExecutionError, PluginRuntimeUnavailableError):
@@ -781,11 +813,7 @@ app.include_router(
 app.include_router(
     build_e47_semantic_slam_router(
         root_provider=lambda: (
-            REPOSITORY_ROOT
-            / ".runtime"
-            / "compute-experiments"
-            / "e47"
-            / "semantic-slam-results"
+            REPOSITORY_ROOT / ".runtime" / "compute-experiments" / "e47" / "semantic-slam-results"
         ),
     )
 )
@@ -1114,9 +1142,17 @@ app.include_router(
         root_provider=lambda: REPOSITORY_ROOT / ".runtime" / "system",
     )
 )
-app.include_router(build_viewer_diagnostics_router())
-
-
 frontend_dist = REPOSITORY_ROOT / "apps" / "control-station" / "dist"
+app.include_router(
+    build_viewer_diagnostics_router(
+        expected_ui_build_id=lambda: frontend_build_id(frontend_dist),
+    )
+)
+
+
 if frontend_dist.is_dir():
-    app.mount("/", StaticFiles(directory=frontend_dist, html=True), name="frontend")
+    app.mount(
+        "/",
+        ControlStationStaticFiles(directory=frontend_dist, html=True),
+        name="frontend",
+    )

src/k1link/web/device_lifecycle.py

@@ -82,6 +82,8 @@ class OperationRecord:
     result: dict[str, Any] | None = None
     error: dict[str, Any] | None = None
     evidence_refs: tuple[str, ...] = ()
+    context: dict[str, Any] = field(default_factory=dict)
+    events: list[dict[str, Any]] = field(default_factory=list)
     # A keyed, non-reversible digest supplied by the service. It is deliberately
     # excluded from API snapshots: callers only need mismatch detection, while
     # the journal must never retain action inputs or secret material.
@@ -108,6 +110,8 @@ class OperationRecord:
             "result": dict(self.result) if self.result is not None else None,
             "error": dict(self.error) if self.error is not None else None,
             "evidence_refs": list(self.evidence_refs),
+            "context": dict(self.context),
+            "events": [dict(event) for event in self.events],
         }
 
 
@@ -144,6 +148,7 @@ class OperationJournal:
         deadline_seconds: float | None = None,
         cancellable: bool = False,
         request_fingerprint: str | None = None,
+        context: Mapping[str, Any] | None = None,
     ) -> tuple[OperationRecord, bool]:
         action = action.strip()
         if not action:
@@ -189,7 +194,9 @@ class OperationJournal:
                 ),
                 cancellable=cancellable,
                 request_fingerprint=request_fingerprint,
+                context=dict(context or {}),
             )
+            self._append_event_locked(record)
             self._records[resolved_id] = record
             self._order.append(resolved_id)
             if idempotency_key is not None:
@@ -225,6 +232,7 @@ class OperationJournal:
             if status in TERMINAL_OPERATION_STATUSES:
                 record.completed_at = self._clock()
                 self._trim_locked()
+            self._append_event_locked(record)
             return record
 
     def request_cancel(self, operation_id: str) -> OperationRecord:
@@ -239,6 +247,7 @@ class OperationJournal:
             record.state_revision += 1
             record.stage_code = "cancellation-requested"
             record.message_code = "operation.cancellation_requested"
+            self._append_event_locked(record)
             return record
 
     def transition_if_pending(
@@ -276,12 +285,32 @@ class OperationJournal:
             if status in TERMINAL_OPERATION_STATUSES:
                 record.completed_at = self._clock()
                 self._trim_locked()
+            self._append_event_locked(record)
             return record
 
     def get(self, operation_id: str) -> OperationRecord:
         with self._lock:
             return self._require_locked(operation_id)
 
+    def deadline_reached(self, operation_id: str | None) -> bool:
+        """Evaluate one operation deadline on the journal-owned server clock.
+
+        Device lifecycle reducers use this instead of browser timers or a
+        second wall-clock source.  Terminal records deliberately retain the
+        same answer so an idempotent local-cleanup retry can continue after the
+        operation outcome itself has already been sealed.
+        """
+
+        if operation_id is None:
+            return False
+        with self._lock:
+            record = self._records.get(operation_id)
+            return bool(
+                record is not None
+                and record.deadline_at is not None
+                and self._clock() >= record.deadline_at
+            )
+
     def latest(self) -> OperationRecord | None:
         with self._lock:
             return self._records[self._order[-1]] if self._order else None
@@ -298,6 +327,32 @@ class OperationJournal:
         except KeyError as exc:
             raise KeyError(f"unknown operation: {operation_id}") from exc
 
+    def _append_event_locked(self, record: OperationRecord) -> None:
+        """Append one bounded, secret-free stage fact for operator diagnosis."""
+
+        error = record.error or {}
+        result = record.result or {}
+        side_effect_status = error.get("side_effect_status", result.get("side_effect_status"))
+        event: dict[str, Any] = {
+            "schema_version": "missioncore.operation-event/v1",
+            "sequence": record.sequence,
+            "status": record.status,
+            "stage_code": record.stage_code,
+            "message_code": record.message_code,
+            "observed_at": _iso(self._clock()),
+            "side_effect_status": (
+                side_effect_status if isinstance(side_effect_status, str) else None
+            ),
+            "error_code": error.get("code") if isinstance(error.get("code"), str) else None,
+            "safe_to_retry": (
+                error.get("safe_to_retry")
+                if isinstance(error.get("safe_to_retry"), bool)
+                else None
+            ),
+            "automatic_retry": False,
+        }
+        record.events.append(event)
+
     def _trim_locked(self) -> None:
         while len(self._order) > self._max_records:
             oldest_id = next(

src/k1link/web/frontend_assets.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+from fastapi.staticfiles import StaticFiles
+from starlette.responses import Response
+from starlette.types import Scope
+
+HTML_NO_STORE = "no-store"
+HASHED_ASSET_IMMUTABLE = "public, max-age=31536000, immutable"
+
+_MODULE_SCRIPT = re.compile(
+    r'<script\b[^>]*\bsrc=["\'](?P<src>/assets/[^"\']+)["\'][^>]*>',
+    re.IGNORECASE,
+)
+_HASHED_ASSET = re.compile(
+    r"^assets/(?:.+)-[A-Za-z0-9_-]{8,}\.[A-Za-z0-9]+$",
+)
+
+
+def frontend_build_id(frontend_root: Path) -> str | None:
+    """Return the exact content-hashed module loaded by the current index."""
+
+    try:
+        index = (frontend_root / "index.html").read_text(encoding="utf-8")
+    except OSError:
+        return None
+    for match in _MODULE_SCRIPT.finditer(index):
+        source = match.group("src")
+        if _HASHED_ASSET.fullmatch(source.removeprefix("/")):
+            return source
+    return None
+
+
+class ControlStationStaticFiles(StaticFiles):
+    """Serve the SPA shell fresh while retaining immutable hashed assets."""
+
+    async def get_response(self, path: str, scope: Scope) -> Response:
+        response = await super().get_response(path, scope)
+        if response.status_code >= 400:
+            return response
+        content_type = response.headers.get("content-type", "").lower()
+        normalized = path.lstrip("/")
+        if content_type.startswith("text/html"):
+            response.headers["Cache-Control"] = HTML_NO_STORE
+        elif _HASHED_ASSET.fullmatch(normalized):
+            response.headers["Cache-Control"] = HASHED_ASSET_IMMUTABLE
+        return response

src/k1link/web/plugin_runtime.py

@@ -68,6 +68,19 @@ class PluginActionNotFoundError(LookupError):
 class PluginExecutionError(RuntimeError):
     """A validated plugin action failed while talking to its device/runtime."""
 
+    def __init__(
+        self,
+        message: str,
+        *,
+        http_status_code: int = 502,
+        reason_code: str | None = None,
+    ) -> None:
+        super().__init__(message)
+        if not 400 <= http_status_code <= 599:
+            raise ValueError("plugin execution HTTP status must be an error response")
+        self.http_status_code = http_status_code
+        self.reason_code = reason_code
+
 
 class PluginRuntimeCompatibilityError(RuntimeError):
     """A plugin runtime cannot satisfy the reviewed manifest/host contract."""
@@ -124,9 +137,7 @@ class InProcessDevicePluginRuntime:
                 raise PluginRuntimeCompatibilityError(
                     "Runtime does not support the manifest host API version"
                 )
-            if frozenset(request.required_action_ids) != frozenset(
-                self.descriptor.action_ids
-            ):
+            if frozenset(request.required_action_ids) != frozenset(self.descriptor.action_ids):
                 raise PluginRuntimeCompatibilityError(
                     "Runtime handshake actions do not match the manifest"
                 )
@@ -225,9 +236,7 @@ class DevicePluginDispatcher:
 
     @property
     def health_snapshots(self) -> dict[str, RuntimeHealthSnapshot]:
-        return {
-            plugin_id: runtime.health() for plugin_id, runtime in self._runtimes.items()
-        }
+        return {plugin_id: runtime.health() for plugin_id, runtime in self._runtimes.items()}
 
     async def invoke(
         self,

src/k1link/web/runtime_diagnostics.py

@@ -55,6 +55,10 @@ _EXTRA_FIELDS: Final = (
     "start_checkpoint_released",
     "stale_session_retired",
     "stream_id",
+    "ui_build_id",
+    "document_instance_id",
+    "viewer_instance_id",
+    "lifecycle_generation",
     "backend_activity_sequence",
     "viewer_range_max_ns",
     "stalled_for_ms",

src/k1link/web/viewer_diagnostics_api.py

@@ -1,9 +1,11 @@
 from __future__ import annotations
 
 import logging
+from collections.abc import Callable
 from typing import Literal
 
 from fastapi import APIRouter, Response
+from fastapi.responses import JSONResponse
 from pydantic import BaseModel, ConfigDict, Field
 
 logger = logging.getLogger("k1link.device_plugins.xgrids_k1.viewer_receiver")
@@ -23,12 +25,31 @@ LiveViewerFailureStage = Literal[
     "receiver-stalled",
 ]
 
+LIVE_VIEWER_DIAGNOSTIC_SCHEMA = "missioncore.live-viewer-diagnostic/v2"
+LIVE_VIEWER_CLIENT_CONTRACT_SCHEMA = "missioncore.live-viewer-client-contract/v1"
+UI_BUILD_HEADER = "X-MissionCore-UI-Build"
+
+_INSTANCE_ID_PATTERN = r"^[0-9a-f]{8}-[0-9a-f]{4}-[1-8][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$"
+_UI_BUILD_ID_PATTERN = r"^(?:development|/assets/[A-Za-z0-9._/-]+)$"
+
 
 class LiveViewerDiagnosticEvent(BaseModel):
     model_config = ConfigDict(extra="forbid", strict=True)
 
-    schema_version: Literal["missioncore.live-viewer-diagnostic/v1"]
+    schema_version: Literal["missioncore.live-viewer-diagnostic/v2"]
     event_code: LiveViewerEventCode
+    ui_build_id: str = Field(min_length=1, max_length=256, pattern=_UI_BUILD_ID_PATTERN)
+    document_instance_id: str = Field(
+        min_length=36,
+        max_length=36,
+        pattern=_INSTANCE_ID_PATTERN,
+    )
+    viewer_instance_id: str = Field(
+        min_length=36,
+        max_length=36,
+        pattern=_INSTANCE_ID_PATTERN,
+    )
+    lifecycle_generation: int = Field(ge=1)
     failure_stage: LiveViewerFailureStage | None = None
     stream_id: str | None = Field(
         default=None,
@@ -42,17 +63,79 @@ class LiveViewerDiagnosticEvent(BaseModel):
     recovery_attempt: int | None = Field(default=None, ge=1, le=3)
 
 
-def build_viewer_diagnostics_router() -> APIRouter:
+def build_viewer_diagnostics_router(
+    *,
+    expected_ui_build_id: Callable[[], str | None] = lambda: "development",
+) -> APIRouter:
     router = APIRouter(prefix="/api/v1/viewer", tags=["viewer"])
 
+    @router.get("/client-contract")
+    def get_live_viewer_client_contract() -> JSONResponse:
+        expected = expected_ui_build_id()
+        if expected is None:
+            return JSONResponse(
+                status_code=503,
+                content={
+                    "schema_version": LIVE_VIEWER_CLIENT_CONTRACT_SCHEMA,
+                    "status": "frontend-build-unavailable",
+                },
+                headers={"Cache-Control": "no-store"},
+            )
+        return JSONResponse(
+            content={
+                "schema_version": LIVE_VIEWER_CLIENT_CONTRACT_SCHEMA,
+                "status": "ready",
+                "ui_build_id": expected,
+                "diagnostic_schema_version": LIVE_VIEWER_DIAGNOSTIC_SCHEMA,
+            },
+            headers={
+                "Cache-Control": "no-store",
+                UI_BUILD_HEADER: expected,
+            },
+        )
+
     @router.post("/live-diagnostics", status_code=204)
     def record_live_diagnostic(event: LiveViewerDiagnosticEvent) -> Response:
+        expected = expected_ui_build_id()
+        if expected is None:
+            return JSONResponse(
+                status_code=503,
+                content={"detail": "frontend-build-unavailable"},
+                headers={"Cache-Control": "no-store"},
+            )
+        # Vite's unhashed development entry intentionally has no deploy
+        # identity and the browser-side monitor is disabled for it. Keep local
+        # diagnostics usable without turning a dist build into a reload loop;
+        # every production document carries the exact hashed entry below.
+        if event.ui_build_id not in {expected, "development"}:
+            return JSONResponse(
+                status_code=409,
+                content={
+                    "detail": "stale-ui-build",
+                    "expected_ui_build_id": expected,
+                },
+                headers={
+                    "Cache-Control": "no-store",
+                    UI_BUILD_HEADER: expected,
+                },
+            )
         logger.info(
-            "Mission Core live Rerun receiver diagnostic: event=%s stream=%s",
+            (
+                "Mission Core live Rerun receiver diagnostic: event=%s stream=%s "
+                "document=%s viewer=%s lifecycle=%s build=%s"
+            ),
             event.event_code,
             event.stream_id,
+            event.document_instance_id,
+            event.viewer_instance_id,
+            event.lifecycle_generation,
+            event.ui_build_id,
             extra={
                 "event_code": event.event_code,
+                "ui_build_id": event.ui_build_id,
+                "document_instance_id": event.document_instance_id,
+                "viewer_instance_id": event.viewer_instance_id,
+                "lifecycle_generation": event.lifecycle_generation,
                 "failure_stage": event.failure_stage,
                 "stream_id": event.stream_id,
                 "backend_activity_sequence": event.backend_activity_sequence,
@@ -61,6 +144,6 @@ def build_viewer_diagnostics_router() -> APIRouter:
                 "recovery_attempt": event.recovery_attempt,
             },
         )
-        return Response(status_code=204)
+        return Response(status_code=204, headers={UI_BUILD_HEADER: expected})
 
     return router