qplotter 1.6.0b3__py3-none-any.whl

qplot/datahandling/trusted_live_service.py

@@ -0,0 +1,1894 @@
+"""Qt-independent Stage 4 broker for one trusted QCoDeS database instance.
+
+The dispatcher is the sole normal query path and serialises all operations for
+one persistent :class:`TrustedLiveReaderSupervisor`.  A separate control
+thread is allowed to call the supervisor's synchronous exact-job cancellation
+while the dispatcher is blocked in ``wait``.  Public cancellation and close
+only mutate bounded broker state and wake that control path, so they return
+promptly and are safe to call from the Qt GUI thread.  Results are consumed
+through request handles; the broker deliberately exposes no callback facility
+whose arbitrary lifetime could hold service retirement open.
+"""
+
+from __future__ import annotations
+
+import heapq
+import logging
+import math
+import os
+import threading
+import time
+from collections.abc import Callable, Generator
+from dataclasses import dataclass, field
+from enum import IntEnum, StrEnum
+from typing import Any, Generic, TypeVar, cast
+
+from qplot.datahandling.file_identity import DatabaseInstance, database_instance
+from qplot.datahandling.plot_progress import PlotProgress
+from qplot.datahandling.trusted_live import (
+    SqliteBindings,
+    TrustedLiveBusyTimeoutError,
+    TrustedLiveCancelledError,
+    TrustedLiveQueryError,
+    TrustedLiveResultLimitError,
+    TrustedLiveSqlRejectedError,
+    TrustedQuery,
+    TrustedQueryResult,
+)
+from qplot.datahandling.trusted_live_queries import (
+    TrustedBootstrapResult,
+    TrustedDatabaseInfo,
+    TrustedDerivedSourceObservation,
+    TrustedMetadataQueryAdapter,
+    TrustedRefreshResult,
+    TrustedRunPage,
+    TrustedRunRecord,
+    TrustedSelectedRunDetail,
+    TrustedSourceRevisionNamespace,
+)
+from qplot.datahandling.trusted_live_supervisor import (
+    TrustedLiveJob,
+    TrustedLiveReaderSupervisor,
+    TrustedLiveSupervisorLiveness,
+)
+
+_LOG = logging.getLogger(__name__)
+_ResultT = TypeVar("_ResultT")
+_NO_RESULT = object()
+_OPERATION_YIELDED = object()
+_DEFAULT_REQUEST_TIMEOUT_SECONDS = 30.0
+_MAX_REQUEST_TIMEOUT_SECONDS = 300.0
+# Large plots can share the broker for many bounded scans. This wall-clock
+# horizon includes time spent yielding to other plots; it is not a SQLite
+# transaction timeout (those remain capped at five seconds).
+_PLOT_CAPTURE_TIMEOUT_SECONDS = 3600.0
+_DEFAULT_QUEUE_CAPACITY = 512
+_MAX_FOREGROUND_DISPATCH_BURST = 8
+
+TRUSTED_LIVE_MODE = "trusted_live"
+SNAPSHOT_FALLBACK_MODE = "snapshot_fallback"
+
+
+class TrustedReadOperation(StrEnum):
+    BOOTSTRAP = "bootstrap"
+    DATABASE_INFO = "database-info"
+    BASIC_PAGE = "basic-page"
+    REFRESH = "refresh"
+    CHEAP_RUN = "cheap-run"
+    EXPENSIVE_RUN = "expensive-run"
+    SELECTED_RUN = "selected-run"
+    DERIVED_SOURCE = "derived-source"
+    PLOT_DATASET = "plot-dataset"
+    PLOT_PREFIX = "plot-prefix"
+
+
+class TrustedReadPriority(IntEnum):
+    BOOTSTRAP = 0
+    REFRESH = 10
+    SELECTED_CHEAP = 20
+    SELECTED_EXPENSIVE = 30
+    VISIBLE_CHEAP = 40
+    VISIBLE_EXPENSIVE = 50
+    REMAINING_CHEAP = 60
+    REMAINING_EXPENSIVE = 70
+
+
+class TrustedReadServiceError(RuntimeError):
+    """Base class for failures introduced by the Stage 4 broker."""
+
+
+class TrustedReadServiceClosedError(TrustedReadServiceError):
+    """A request targeted a closing or closed broker session."""
+
+
+class TrustedReadQueueFullError(TrustedReadServiceError):
+    """The bounded broker queue rejected additional work."""
+
+
+class TrustedReadRequestCancelledError(
+    InterruptedError,
+    TrustedReadServiceError,
+):
+    """One exact public broker request was cancelled."""
+
+
+class TrustedReadRequestDeadlineError(TimeoutError, TrustedReadServiceError):
+    """A broker request expired before it could publish a result."""
+
+
+class TrustedReadSessionFailedError(TrustedReadServiceError):
+    """A terminal helper/session failure invalidated queued work."""
+
+
+@dataclass(frozen=True, slots=True)
+class TrustedReadRequestIdentity:
+    """Immutable identity and scheduling data for one public request."""
+
+    session_generation: int
+    request_id: int
+    database_instance: DatabaseInstance
+    operation: TrustedReadOperation
+    initial_priority: int
+    deadline: float
+
+
+@dataclass(frozen=True, slots=True)
+class TrustedReadServiceLiveness:
+    dispatcher_alive: bool
+    control_alive: bool
+    helper_alive: bool
+    helper_pid: int | None
+    receiver_alive: bool
+    open_supervisor_endpoints: int
+    unreaped_incarnations: int
+    resource_cleanup_pending: bool
+    outstanding_requests: int
+    closing: bool
+    closed: bool
+
+
+@dataclass(slots=True)
+class _RequestState:
+    identity: TrustedReadRequestIdentity
+    priority: int
+    operation_id: int
+    done: threading.Event = field(default_factory=threading.Event)
+    result: object = _NO_RESULT
+    error: BaseException | None = None
+    cancelled: bool = False
+    request: TrustedReadRequest[Any] | None = None
+    slot_released: bool = False
+    progress: PlotProgress | None = None
+
+
+@dataclass(slots=True)
+class _OperationState:
+    operation_id: int
+    kind: TrustedReadOperation
+    payload: tuple[object, ...]
+    coalesce_key: tuple[object, ...]
+    priority: int
+    sequence: int
+    deadline: float
+    subscribers: set[int] = field(default_factory=set)
+    revision: int = 0
+    status: str = "queued"
+    supervisor_job: TrustedLiveJob[Any] | None = None
+    cancel_underlying: bool = False
+    force_next_transaction: bool = False
+    plot_steps: Generator[None, None, Any] | None = None
+    progress: PlotProgress | None = None
+
+
+class TrustedReadRequest(Generic[_ResultT]):
+    """One exact cancellable handle returned by :class:`TrustedLiveReadService`."""
+
+    __slots__ = ("_service", "_state")
+
+    def __init__(
+        self,
+        service: TrustedLiveReadService,
+        state: _RequestState,
+    ) -> None:
+        self._service = service
+        self._state = state
+
+    @property
+    def identity(self) -> TrustedReadRequestIdentity:
+        return self._state.identity
+
+    @property
+    def request_id(self) -> int:
+        return self._state.identity.request_id
+
+    @property
+    def priority(self) -> int:
+        return self._service._request_priority(self._state)
+
+    @property
+    def deadline(self) -> float:
+        return self._state.identity.deadline
+
+    @property
+    def done(self) -> bool:
+        return self._state.done.is_set()
+
+    @property
+    def cancelled(self) -> bool:
+        return self._state.cancelled
+
+    @property
+    def progress(self) -> PlotProgress | None:
+        """Latest immutable observation; polling never waits for a query."""
+        return self._state.progress
+
+    def cancel(self) -> bool:
+        """Mark only this public request; never block on supervisor cleanup."""
+
+        return self._service.cancel(self)
+
+    def promote(self, priority: int | TrustedReadPriority) -> bool:
+        return self._service.promote(self, priority)
+
+    def reprioritize(self, priority: int | TrustedReadPriority) -> bool:
+        """Replace this subscriber's priority, including a demotion."""
+
+        return self._service.reprioritize(self, priority)
+
+    def wait(self, timeout: float | None = None) -> _ResultT:
+        """Wait for publication; callers must keep this off the Qt GUI thread."""
+
+        if timeout is not None:
+            timeout = _finite_duration(timeout, allow_zero=True)
+        if not self._state.done.wait(timeout):
+            raise TimeoutError("The broker request wait timed out.")
+        error = self._state.error
+        if error is not None:
+            raise error
+        if self._state.result is _NO_RESULT:
+            raise TrustedReadServiceError(
+                "The broker request completed without a result or error."
+            )
+        return cast(_ResultT, self._state.result)
+
+
+class _BrokerQueryExecutor:
+    """Expose supervisor operations while publishing the exact active job."""
+
+    def __init__(
+        self,
+        service: TrustedLiveReadService,
+        operation: _OperationState,
+    ) -> None:
+        self._service = service
+        self._operation = operation
+
+    @property
+    def incarnation(self) -> int:
+        return self._service._required_supervisor().incarnation
+
+    def check_cancelled(self):
+        with self._service._condition:
+            if self._service._closing or self._operation.cancel_underlying:
+                raise TrustedReadRequestCancelledError("Plot capture cancelled.")
+        self._service._operation_remaining(self._operation)
+
+    def report_progress(self, phase, completed=0, total=None, *, unit="scan", stage=1):
+        progress = PlotProgress(phase, completed, total, unit, stage)
+        with self._service._condition:
+            self._operation.progress = progress
+            for request_id in self._operation.subscribers:
+                state = self._service._requests.get(request_id)
+                if state is not None and not state.done.is_set():
+                    state.progress = progress
+
+    def query(
+        self,
+        sql: str,
+        bindings: SqliteBindings = None,
+        *,
+        timeout: float | None = None,
+        wait_timeout: float | None = None,
+    ) -> TrustedQueryResult:
+        del timeout, wait_timeout
+        supervisor = self._service._required_supervisor()
+        remaining = self._service._prepare_supervisor_transaction(
+            self._operation,
+            self,
+        )
+        if self._operation.kind in (TrustedReadOperation.PLOT_DATASET, TrustedReadOperation.PLOT_PREFIX):
+            remaining = min(remaining, 5.0)
+        job = supervisor.submit_query(sql, bindings, timeout=remaining)
+        return self._service._wait_supervisor_job(
+            self._operation,
+            job,
+            remaining,
+        )
+
+    def query_batch(
+        self,
+        queries: tuple[TrustedQuery, ...],
+        *,
+        timeout: float | None = None,
+        wait_timeout: float | None = None,
+    ) -> tuple[TrustedQueryResult, ...]:
+        del timeout, wait_timeout
+        supervisor = self._service._required_supervisor()
+        remaining = self._service._prepare_supervisor_transaction(
+            self._operation,
+            self,
+        )
+        if self._operation.kind in (TrustedReadOperation.PLOT_DATASET, TrustedReadOperation.PLOT_PREFIX):
+            remaining = min(remaining, 5.0)
+        job = supervisor.submit_query_batch(queries, timeout=remaining)
+        return self._service._wait_supervisor_job(
+            self._operation,
+            job,
+            remaining,
+        )
+
+    def data_version(
+        self,
+        *,
+        timeout: float | None = None,
+        wait_timeout: float | None = None,
+    ) -> int:
+        del timeout, wait_timeout
+        supervisor = self._service._required_supervisor()
+        remaining = self._service._prepare_supervisor_transaction(
+            self._operation,
+            self,
+        )
+        job = supervisor.submit_data_version(timeout=remaining)
+        return self._service._wait_supervisor_job(
+            self._operation,
+            job,
+            remaining,
+        )
+
+
+def _finite_duration(value: object, *, allow_zero: bool) -> float:
+    if isinstance(value, bool):
+        raise ValueError("A request timeout must be a finite number of seconds.")
+    try:
+        duration = float(cast(Any, value))
+    except (TypeError, ValueError) as error:
+        raise ValueError(
+            "A request timeout must be a finite number of seconds."
+        ) from error
+    if (
+        not math.isfinite(duration)
+        or duration < 0
+        or (not allow_zero and duration == 0)
+        or duration > _MAX_REQUEST_TIMEOUT_SECONDS
+    ):
+        raise ValueError(
+            "A request timeout must be finite, non-negative, and no greater "
+            f"than {_MAX_REQUEST_TIMEOUT_SECONDS:g} seconds."
+        )
+    return duration
+
+
+def _priority(value: object) -> int:
+    if isinstance(value, bool):
+        raise ValueError("priority must be an integer.")
+    try:
+        priority = int(cast(Any, value))
+    except (TypeError, ValueError) as error:
+        raise ValueError("priority must be an integer.") from error
+    if priority < -1_000_000 or priority > 1_000_000:
+        raise ValueError("priority is outside the bounded broker range.")
+    return priority
+
+
+class TrustedLiveReadService:
+    """Own exactly one persistent trusted helper for one database instance."""
+
+    def __init__(
+        self,
+        database_path: str | os.PathLike[str],
+        *,
+        expected_database_instance: DatabaseInstance | None = None,
+        session_generation: int = 1,
+        queue_capacity: int = _DEFAULT_QUEUE_CAPACITY,
+        request_timeout_seconds: float = _DEFAULT_REQUEST_TIMEOUT_SECONDS,
+        supervisor_factory: Callable[..., TrustedLiveReaderSupervisor] | None = None,
+        supervisor_options: dict[str, Any] | None = None,
+    ) -> None:
+        if type(session_generation) is not int or session_generation <= 0:
+            raise ValueError("session_generation must be a positive integer.")
+        if type(queue_capacity) is not int or queue_capacity <= 0:
+            raise ValueError("queue_capacity must be a positive integer.")
+        selected_path = os.fspath(database_path)
+        if expected_database_instance is None:
+            expected_database_instance = database_instance(selected_path)
+        elif not isinstance(expected_database_instance, DatabaseInstance):
+            raise TypeError(
+                "expected_database_instance must be a DatabaseInstance or None."
+            )
+        self._database_path = selected_path
+        self._database_instance = expected_database_instance
+        self._session_generation = session_generation
+        self._capacity = queue_capacity
+        self._request_timeout = _finite_duration(
+            request_timeout_seconds,
+            allow_zero=False,
+        )
+        self._supervisor_factory = (
+            supervisor_factory or TrustedLiveReaderSupervisor.open
+        )
+        self._supervisor_options = dict(supervisor_options or {})
+        self._supervisor: TrustedLiveReaderSupervisor | None = None
+        self._closing_supervisor: TrustedLiveReaderSupervisor | None = None
+        self._retained_supervisors: list[TrustedLiveReaderSupervisor] = []
+        self._adapter: TrustedMetadataQueryAdapter | None = None
+        # Accessed only by the dispatcher; never owned by a GUI dataset.
+        self._plot_prefix_cache: dict[tuple, Any] = {}
+        self._retired_plot_steps: list[Generator[None, None, Any]] = []
+        self._source_revision_namespace = TrustedSourceRevisionNamespace.create()
+        self._accepted_database_instance: DatabaseInstance | None = None
+
+        self._condition = threading.Condition(threading.RLock())
+        self._control_wakeup = threading.Event()
+        self._control_stop = threading.Event()
+        self._dispatcher_done = threading.Event()
+        self._control_done = threading.Event()
+        self._closed_event = threading.Event()
+        self._requests: dict[int, _RequestState] = {}
+        self._request_slots_in_use = 0
+        self._operations: dict[int, _OperationState] = {}
+        self._coalesced: dict[tuple[object, ...], int] = {}
+        self._heap: list[tuple[int, int, int, int]] = []
+        self._next_request_id = 1
+        self._next_operation_id = 1
+        self._next_sequence = 1
+        self._foreground_dispatch_burst = 0
+        self._active_operation: _OperationState | None = None
+        self._control_cancel_job: TrustedLiveJob[Any] | None = None
+        self._closing = False
+        self._closed = False
+        self._fatal_error: BaseException | None = None
+        self._close_error: BaseException | None = None
+
+        self._control_thread = threading.Thread(
+            target=self._control_loop,
+            name=f"qplot-trusted-read-control-{session_generation}",
+            daemon=True,
+        )
+        self._dispatcher_thread = threading.Thread(
+            target=self._dispatcher_loop,
+            name=f"qplot-trusted-read-dispatcher-{session_generation}",
+            daemon=True,
+        )
+        self._control_thread.start()
+        self._dispatcher_thread.start()
+
+    @property
+    def database_instance(self) -> DatabaseInstance:
+        with self._condition:
+            return self._accepted_database_instance or self._database_instance
+
+    @property
+    def session_generation(self) -> int:
+        return self._session_generation
+
+    @property
+    def source_revision_namespace(self) -> TrustedSourceRevisionNamespace:
+        """Restart-unique identity used by all derived observations in this service."""
+
+        return self._source_revision_namespace
+
+    @property
+    def mode(self) -> str:
+        return TRUSTED_LIVE_MODE
+
+    @property
+    def accepted(self) -> bool:
+        """Whether Stage 3 startup accepted and bound the source instance."""
+
+        with self._condition:
+            return self._accepted_database_instance is not None
+
+    @property
+    def closed(self) -> bool:
+        with self._condition:
+            return self._closed
+
+    @property
+    def closing(self) -> bool:
+        """Whether asynchronous retirement has begun but is not yet complete."""
+
+        with self._condition:
+            return self._closing
+
+    @property
+    def close_error(self) -> BaseException | None:
+        with self._condition:
+            return self._close_error
+
+    @property
+    def fatal_error(self) -> BaseException | None:
+        """Return the exact terminal dispatcher/control failure, if any."""
+
+        with self._condition:
+            return self._fatal_error
+
+    def liveness(self) -> TrustedReadServiceLiveness:
+        with self._condition:
+            supervisors = self._owned_supervisors_locked()
+            outstanding_requests = self._request_slots_in_use
+            closing = self._closing
+            closed = self._closed
+        helper_alive = False
+        helper_pid = None
+        receiver_alive = False
+        open_endpoints = 0
+        unreaped_incarnations = 0
+        resource_cleanup_pending = False
+        for supervisor in supervisors:
+            snapshot = self._supervisor_resource_liveness(
+                supervisor,
+                conservative_on_error=False,
+            )
+            helper_alive = helper_alive or snapshot.process_alive
+            if helper_pid is None:
+                helper_pid = snapshot.helper_pid
+            receiver_alive = receiver_alive or snapshot.receiver_alive
+            open_endpoints += snapshot.open_endpoints
+            unreaped_incarnations += int(snapshot.unreaped_incarnation)
+            resource_cleanup_pending = (
+                resource_cleanup_pending or snapshot.resources_owned
+            )
+        return TrustedReadServiceLiveness(
+            dispatcher_alive=self._dispatcher_thread.is_alive(),
+            control_alive=self._control_thread.is_alive(),
+            helper_alive=helper_alive,
+            helper_pid=helper_pid,
+            receiver_alive=receiver_alive,
+            open_supervisor_endpoints=open_endpoints,
+            unreaped_incarnations=unreaped_incarnations,
+            resource_cleanup_pending=resource_cleanup_pending,
+            outstanding_requests=outstanding_requests,
+            closing=closing,
+            closed=closed,
+        )
+
+    def submit_bootstrap(
+        self,
+        *,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedBootstrapResult]:
+        return cast(
+            TrustedReadRequest[TrustedBootstrapResult],
+            self._submit(
+                TrustedReadOperation.BOOTSTRAP,
+                (),
+                TrustedReadPriority.BOOTSTRAP,
+                deadline,
+            ),
+        )
+
+    def submit_refresh(
+        self,
+        accepted_run_id: int | None = None,
+        *,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedRefreshResult]:
+        if accepted_run_id is not None and (
+            type(accepted_run_id) is not int or accepted_run_id < 0
+        ):
+            raise ValueError("accepted_run_id must be a non-negative integer or None.")
+        return cast(
+            TrustedReadRequest[TrustedRefreshResult],
+            self._submit(
+                TrustedReadOperation.REFRESH,
+                (accepted_run_id,),
+                TrustedReadPriority.REFRESH,
+                deadline,
+            ),
+        )
+
+    def submit_database_info(
+        self,
+        *,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedDatabaseInfo]:
+        """Read current diagnostic metadata without a snapshot or GUI wait."""
+
+        return cast(
+            TrustedReadRequest[TrustedDatabaseInfo],
+            self._submit(
+                TrustedReadOperation.DATABASE_INFO,
+                (),
+                TrustedReadPriority.SELECTED_CHEAP,
+                deadline,
+            ),
+        )
+
+    def submit_basic_page(
+        self,
+        after_run_id: int,
+        through_run_id: int,
+        *,
+        priority: int | TrustedReadPriority = TrustedReadPriority.BOOTSTRAP,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedRunPage]:
+        if type(after_run_id) is not int or after_run_id < 0:
+            raise ValueError("after_run_id must be a non-negative integer.")
+        if type(through_run_id) is not int or through_run_id < after_run_id:
+            raise ValueError(
+                "through_run_id must be an integer no smaller than after_run_id."
+            )
+        return cast(
+            TrustedReadRequest[TrustedRunPage],
+            self._submit(
+                TrustedReadOperation.BASIC_PAGE,
+                (after_run_id, through_run_id),
+                priority,
+                deadline,
+            ),
+        )
+
+    def submit_cheap_run(
+        self,
+        run_id: int,
+        *,
+        priority: int | TrustedReadPriority = TrustedReadPriority.REMAINING_CHEAP,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedRunRecord]:
+        return cast(
+            TrustedReadRequest[TrustedRunRecord],
+            self._submit(
+                TrustedReadOperation.CHEAP_RUN,
+                (self._validated_run_id(run_id),),
+                priority,
+                deadline,
+            ),
+        )
+
+    def submit_expensive_run(
+        self,
+        run_id: int,
+        *,
+        priority: int | TrustedReadPriority = (TrustedReadPriority.REMAINING_EXPENSIVE),
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedRunRecord]:
+        return cast(
+            TrustedReadRequest[TrustedRunRecord],
+            self._submit(
+                TrustedReadOperation.EXPENSIVE_RUN,
+                (self._validated_run_id(run_id),),
+                priority,
+                deadline,
+            ),
+        )
+
+    def submit_selected_run(
+        self,
+        run_id: int,
+        *,
+        priority: int | TrustedReadPriority = TrustedReadPriority.SELECTED_EXPENSIVE,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedSelectedRunDetail]:
+        return cast(
+            TrustedReadRequest[TrustedSelectedRunDetail],
+            self._submit(
+                TrustedReadOperation.SELECTED_RUN,
+                (self._validated_run_id(run_id),),
+                priority,
+                deadline,
+            ),
+        )
+
+    def submit_derived_source(
+        self,
+        run_id: int,
+        *,
+        priority: int | TrustedReadPriority = TrustedReadPriority.REMAINING_EXPENSIVE,
+        deadline: float | None = None,
+    ) -> TrustedReadRequest[TrustedDerivedSourceObservation]:
+        """Capture one bounded immutable prefix through the persistent broker."""
+
+        return cast(
+            TrustedReadRequest[TrustedDerivedSourceObservation],
+            self._submit(
+                TrustedReadOperation.DERIVED_SOURCE,
+                (self._validated_run_id(run_id),),
+                priority,
+                deadline,
+            ),
+        )
+
+    def submit_plot_dataset(self, guid, *, deadline=None):
+        if not isinstance(guid, str) or not 0 < len(guid) <= 256:
+            raise ValueError("Invalid plot GUID.")
+        return self._submit(TrustedReadOperation.PLOT_DATASET, (guid,),
+                            TrustedReadPriority.SELECTED_CHEAP, deadline)
+
+    def submit_plot_prefix(self, dataset, *, deadline=None, heatmap_plan=None):
+        if dataset.service is not self:
+            raise ValueError("The plot belongs to another trusted session.")
+        payload = (dataset,) if heatmap_plan is None else (dataset, heatmap_plan)
+        return self._submit(TrustedReadOperation.PLOT_PREFIX, payload,
+                            TrustedReadPriority.SELECTED_EXPENSIVE, deadline)
+
+    @staticmethod
+    def _validated_run_id(run_id: object) -> int:
+        if type(run_id) is not int or run_id <= 0:
+            raise ValueError("run_id must be a positive integer.")
+        return run_id
+
+    def _submit(
+        self,
+        kind: TrustedReadOperation,
+        payload: tuple[object, ...],
+        priority: int | TrustedReadPriority,
+        deadline: float | None,
+    ) -> TrustedReadRequest[Any]:
+        requested_priority = _priority(priority)
+        now = time.monotonic()
+        default_deadline = deadline is None
+        if deadline is None:
+            request_deadline = now + (
+                _PLOT_CAPTURE_TIMEOUT_SECONDS
+                if kind is TrustedReadOperation.PLOT_PREFIX else self._request_timeout
+            )
+        else:
+            if isinstance(deadline, bool):
+                raise ValueError("deadline must be a finite monotonic timestamp.")
+            request_deadline = float(deadline)
+            if not math.isfinite(request_deadline):
+                raise ValueError("deadline must be a finite monotonic timestamp.")
+            if request_deadline - now > _MAX_REQUEST_TIMEOUT_SECONDS:
+                raise ValueError("deadline exceeds the bounded broker request horizon.")
+
+        with self._condition:
+            if self._closing or self._closed:
+                raise TrustedReadServiceClosedError(
+                    "The trusted read service is closing or closed."
+                )
+            if self._request_slots_in_use >= self._capacity:
+                raise TrustedReadQueueFullError(
+                    f"The trusted read queue reached its {self._capacity}-request cap."
+                )
+            request_id = self._next_request_id
+            self._next_request_id += 1
+            # Ordinary duplicate submissions share the first operation's
+            # bounded default deadline; an active supervisor wait is therefore
+            # never extended. Explicit deadlines coalesce only when exactly
+            # equal, preserving every caller's requested bound.
+            coalesce_key = (
+                (kind, payload, "default")
+                if default_deadline
+                else (kind, payload, "explicit", request_deadline)
+            )
+            operation_id = self._coalesced.get(coalesce_key)
+            operation = (
+                self._operations.get(operation_id) if operation_id is not None else None
+            )
+            coalesce_eligible = bool(
+                operation is not None
+                and operation.status in {"queued", "active"}
+                and not operation.cancel_underlying
+                and operation.deadline > now
+            )
+            if not coalesce_eligible:
+                operation_id = self._next_operation_id
+                self._next_operation_id += 1
+                sequence = self._next_sequence
+                self._next_sequence += 1
+                operation = _OperationState(
+                    operation_id,
+                    kind,
+                    payload,
+                    coalesce_key,
+                    requested_priority,
+                    sequence,
+                    request_deadline,
+                )
+                self._operations[operation_id] = operation
+                self._coalesced[coalesce_key] = operation_id
+                heapq.heappush(
+                    self._heap,
+                    (
+                        operation.priority,
+                        operation.sequence,
+                        operation.revision,
+                        operation.operation_id,
+                    ),
+                )
+            else:
+                assert operation is not None
+                if default_deadline:
+                    request_deadline = operation.deadline
+            assert operation is not None
+            identity = TrustedReadRequestIdentity(
+                self._session_generation,
+                request_id,
+                self._database_instance,
+                kind,
+                requested_priority,
+                request_deadline,
+            )
+            state = _RequestState(
+                identity,
+                requested_priority,
+                operation.operation_id,
+            )
+            state.progress = operation.progress
+            request: TrustedReadRequest[Any] = TrustedReadRequest(self, state)
+            state.request = request
+            self._requests[request_id] = state
+            self._request_slots_in_use += 1
+            operation.subscribers.add(request_id)
+            # A coalesced operation is always scheduled from the complete set
+            # of live subscribers.  This is also the single path used after
+            # cancellation and request-level reprioritization.
+            self._recompute_operation_schedule_locked(operation)
+            self._condition.notify()
+            self._control_wakeup.set()
+            return request
+
+    def _request_priority(self, state: _RequestState) -> int:
+        with self._condition:
+            return state.priority
+
+    def promote(
+        self,
+        request: TrustedReadRequest[Any],
+        priority: int | TrustedReadPriority,
+    ) -> bool:
+        requested_priority = _priority(priority)
+        with self._condition:
+            state = self._owned_request_state(request)
+            if state.done.is_set() or requested_priority >= state.priority:
+                return False
+            return self._reprioritize_request_locked(state, requested_priority)
+
+    def reprioritize(
+        self,
+        request: TrustedReadRequest[Any],
+        priority: int | TrustedReadPriority,
+    ) -> bool:
+        """Replace one live subscriber priority and reschedule its operation."""
+
+        requested_priority = _priority(priority)
+        with self._condition:
+            state = self._owned_request_state(request)
+            if state.done.is_set() or requested_priority == state.priority:
+                return False
+            return self._reprioritize_request_locked(state, requested_priority)
+
+    def _reprioritize_request_locked(
+        self,
+        state: _RequestState,
+        requested_priority: int,
+    ) -> bool:
+        operation = self._operations.get(state.operation_id)
+        if operation is None:
+            return False
+        state.priority = requested_priority
+        self._recompute_operation_schedule_locked(operation)
+        self._condition.notify_all()
+        return True
+
+    def cancel(self, request: TrustedReadRequest[Any]) -> bool:
+        with self._condition:
+            state = self._owned_request_state(request)
+            if state.done.is_set():
+                return False
+            state.cancelled = True
+            self._complete_request_locked(
+                state,
+                error=TrustedReadRequestCancelledError(
+                    f"Trusted request {state.identity.request_id} was cancelled."
+                ),
+            )
+            operation = self._operations.get(state.operation_id)
+            if operation is not None:
+                operation.subscribers.discard(state.identity.request_id)
+                if not operation.subscribers:
+                    if operation.status == "queued":
+                        self._discard_operation_locked(operation)
+                    elif operation.status == "active":
+                        operation.cancel_underlying = True
+                else:
+                    self._recompute_operation_schedule_locked(operation)
+            self._condition.notify_all()
+        # Wake the deadline/cancellation control path even for a queued
+        # cancellation so it can recompute its earliest timed wait.
+        self._control_wakeup.set()
+        return True
+
+    def _owned_request_state(
+        self,
+        request: TrustedReadRequest[Any],
+    ) -> _RequestState:
+        if not isinstance(request, TrustedReadRequest) or request._service is not self:
+            raise ValueError("The request does not belong to this trusted service.")
+        return request._state
+
+    def close_async(self) -> bool:
+        """Begin cancellation/retirement and return without waiting or joining."""
+
+        with self._condition:
+            if self._closing or self._closed:
+                return False
+            self._closing = True
+            error = TrustedReadServiceClosedError(
+                "The trusted read service closed before the request completed."
+            )
+            for state in list(self._requests.values()):
+                state.cancelled = True
+                self._complete_request_locked(state, error=error)
+            for operation in list(self._operations.values()):
+                operation.subscribers.clear()
+                if operation.status == "queued":
+                    self._discard_operation_locked(operation)
+                elif operation.status == "active":
+                    operation.cancel_underlying = True
+            self._condition.notify_all()
+        self._control_wakeup.set()
+        return True
+
+    def escalate_cleanup_async(self) -> None:
+        """Repeat retirement signals and perform one zero-wait resource reap.
+
+        This method is safe for a Qt shutdown timer: it never waits for a
+        broker thread or helper.  The dispatcher/control paths retain sole
+        ownership of blocking Stage 3 cancellation and close operations.
+        """
+
+        self.close_async()
+        with self._condition:
+            operation = self._active_operation
+            if operation is not None:
+                operation.cancel_underlying = True
+            self._condition.notify_all()
+        self._control_wakeup.set()
+        self._reap_retained_supervisors_once()
+
+    def close(self, *, timeout: float | None = None) -> None:
+        """Blocking convenience for non-GUI owners and tests."""
+
+        self.close_async()
+        if not self.wait_closed(timeout):
+            raise TimeoutError("The trusted read service did not close in time.")
+        error = self.close_error
+        if error is not None:
+            raise error
+
+    def wait_closed(self, timeout: float | None = None) -> bool:
+        if timeout is not None:
+            timeout = _finite_duration(timeout, allow_zero=True)
+        deadline = None if timeout is None else time.monotonic() + timeout
+        threads = (
+            self._dispatcher_thread,
+            self._control_thread,
+        )
+        current = threading.current_thread()
+        if current in threads:
+            return False
+        for event in (
+            self._dispatcher_done,
+            self._control_done,
+            self._closed_event,
+        ):
+            remaining = (
+                None if deadline is None else max(0.0, deadline - time.monotonic())
+            )
+            if not event.wait(remaining):
+                return False
+        for thread in threads:
+            remaining = (
+                None if deadline is None else max(0.0, deadline - time.monotonic())
+            )
+            thread.join(remaining)
+            if thread.is_alive():
+                return False
+        while self._reap_retained_supervisors_once():
+            if deadline is not None:
+                remaining = deadline - time.monotonic()
+                if remaining <= 0:
+                    return False
+                time.sleep(min(0.01, remaining))
+            else:
+                time.sleep(0.01)
+        return True
+
+    def _dispatcher_loop(self) -> None:
+        try:
+            while True:
+                self._close_retired_plot_steps()
+                with self._condition:
+                    operation = self._next_operation_locked()
+                    if operation is None:
+                        should_stop = self._closing
+                        if not should_stop:
+                            self._condition.wait(timeout=0.1)
+                            continue
+                    else:
+                        should_stop = False
+                        operation.status = "active"
+                        self._active_operation = operation
+                        # Fairness may select an operation directly rather than
+                        # popping its live heap entry.  Keep the physical heap
+                        # exactly bounded to the remaining queued operations.
+                        self._compact_heap_locked()
+                if operation is None:
+                    if should_stop:
+                        break
+                    continue
+
+                result: object = _NO_RESULT
+                error: BaseException | None = None
+                try:
+                    result = self._execute_operation(operation)
+                except BaseException as operation_error:
+                    error = operation_error
+
+                with self._condition:
+                    self._finish_operation_locked(
+                        operation,
+                        result,
+                        error,
+                    )
+        except BaseException as dispatcher_error:
+            with self._condition:
+                self._fatal_error = dispatcher_error
+                self._closing = True
+                self._fail_all_locked(dispatcher_error)
+                self._condition.notify_all()
+        finally:
+            self._close_retired_plot_steps()
+            self._plot_prefix_cache.clear()
+            self._close_supervisor_from_dispatcher()
+            with self._condition:
+                self._closed = True
+                self._closing = False
+                self._condition.notify_all()
+            self._dispatcher_done.set()
+            self._control_stop.set()
+            self._control_wakeup.set()
+            with self._condition:
+                self._condition.notify_all()
+            self._closed_event.set()
+
+    def _next_operation_locked(self) -> _OperationState | None:
+        while True:
+            now = time.monotonic()
+            background = min(
+                (
+                    operation
+                    for operation in self._operations.values()
+                    if operation.status == "queued"
+                    and operation.priority >= TrustedReadPriority.VISIBLE_CHEAP
+                ),
+                key=lambda operation: operation.sequence,
+                default=None,
+            )
+            fairness_forced = (
+                background is not None
+                and self._foreground_dispatch_burst >= _MAX_FOREGROUND_DISPATCH_BURST
+            )
+            if fairness_forced:
+                operation = background
+            else:
+                operation = self._pop_priority_operation_locked()
+            if operation is None:
+                self._foreground_dispatch_burst = 0
+                return None
+            if not self._prepare_queued_operation_locked(
+                operation,
+                now,
+            ):
+                continue
+            if operation.priority >= TrustedReadPriority.VISIBLE_CHEAP:
+                self._foreground_dispatch_burst = 0
+            elif background is not None:
+                self._foreground_dispatch_burst += 1
+            else:
+                self._foreground_dispatch_burst = 0
+            operation.force_next_transaction = fairness_forced
+            return operation
+
+    def _pop_priority_operation_locked(self) -> _OperationState | None:
+        while self._heap:
+            priority, sequence, revision, operation_id = heapq.heappop(self._heap)
+            operation = self._operations.get(operation_id)
+            if (
+                operation is not None
+                and operation.status == "queued"
+                and operation.priority == priority
+                and operation.sequence == sequence
+                and operation.revision == revision
+            ):
+                return operation
+        return None
+
+    def _compact_heap_locked(self) -> None:
+        """Retain exactly one heap entry for every queued operation.
+
+        Promotions and queued cancellation used to leave lazy tombstones in
+        ``_heap``.  Because completed requests immediately release queue
+        capacity, a producer could create an unbounded number of those physical
+        entries behind one blocked active job.  Rebuilding at the bounded
+        operation cardinality keeps backpressure true for the actual storage.
+        """
+
+        self._heap = [
+            (
+                operation.priority,
+                operation.sequence,
+                operation.revision,
+                operation.operation_id,
+            )
+            for operation in self._operations.values()
+            if operation.status == "queued"
+        ]
+        heapq.heapify(self._heap)
+
+    def _recompute_operation_schedule_locked(
+        self,
+        operation: _OperationState,
+    ) -> None:
+        """Recompute a shared operation from only its live subscribers."""
+
+        live_states: list[_RequestState] = []
+        for request_id in tuple(operation.subscribers):
+            state = self._requests.get(request_id)
+            if state is None or state.done.is_set():
+                operation.subscribers.discard(request_id)
+            else:
+                live_states.append(state)
+        if not live_states:
+            if operation.status == "queued":
+                self._discard_operation_locked(operation)
+            elif operation.status == "active":
+                operation.cancel_underlying = True
+                self._control_wakeup.set()
+            return
+
+        priority = min(state.priority for state in live_states)
+        operation.deadline = min(state.identity.deadline for state in live_states)
+        if priority == operation.priority:
+            return
+        operation.priority = priority
+        operation.revision += 1
+        if operation.status == "queued":
+            self._compact_heap_locked()
+
+    def _expire_queued_requests_locked(self, now: float) -> None:
+        """Complete queued subscribers at their own monotonic deadlines."""
+
+        for operation in tuple(self._operations.values()):
+            if operation.status != "queued":
+                continue
+            expired = tuple(
+                request_id
+                for request_id in operation.subscribers
+                if (
+                    (state := self._requests.get(request_id)) is not None
+                    and state.identity.deadline <= now
+                )
+            )
+            for request_id in expired:
+                state = self._requests.get(request_id)
+                if state is None:
+                    operation.subscribers.discard(request_id)
+                    continue
+                operation.subscribers.discard(request_id)
+                self._complete_request_locked(
+                    state,
+                    error=TrustedReadRequestDeadlineError(
+                        f"Trusted request {request_id} expired in the queue."
+                    ),
+                )
+            if expired:
+                self._recompute_operation_schedule_locked(operation)
+        self._condition.notify_all()
+
+    def _next_queued_deadline_locked(self) -> float | None:
+        return min(
+            (
+                state.identity.deadline
+                for operation in self._operations.values()
+                if operation.status == "queued"
+                for request_id in operation.subscribers
+                if (state := self._requests.get(request_id)) is not None
+            ),
+            default=None,
+        )
+
+    def _next_higher_priority_operation_locked(
+        self,
+        current_priority: int,
+    ) -> _OperationState | None:
+        # Captures already time-slice at the outer dispatcher. Nesting their
+        # slow pages between every transaction of a finite metadata job can
+        # consume that job's deadline and retire the shared helper. Cheap
+        # interactive metadata still preempts here; captures run between jobs.
+        candidates = sorted(
+            (operation for operation in self._operations.values()
+             if operation.status == "queued" and operation.priority < current_priority
+             and operation.kind is not TrustedReadOperation.PLOT_PREFIX),
+            key=lambda operation: (operation.priority, operation.sequence),
+        )
+        for operation in candidates:
+            if self._prepare_queued_operation_locked(
+                operation,
+                time.monotonic(),
+            ):
+                return operation
+        return None
+
+    def _prepare_queued_operation_locked(
+        self,
+        operation: _OperationState,
+        now: float,
+    ) -> bool:
+        for request_id in tuple(operation.subscribers):
+            state = self._requests.get(request_id)
+            if state is None:
+                operation.subscribers.discard(request_id)
+                continue
+            if state.identity.deadline <= now:
+                operation.subscribers.discard(request_id)
+                self._complete_request_locked(
+                    state,
+                    error=TrustedReadRequestDeadlineError(
+                        f"Trusted request {request_id} expired in the queue."
+                    ),
+                )
+        if not operation.subscribers:
+            self._discard_operation_locked(operation)
+            return False
+        self._recompute_operation_schedule_locked(operation)
+        return True
+
+    def _execute_operation(self, operation: _OperationState) -> object:
+        startup_remaining = self._operation_remaining(operation)
+        supervisor = self._ensure_supervisor(startup_remaining)
+        with self._condition:
+            if self._closing or operation.cancel_underlying:
+                raise TrustedReadRequestCancelledError(
+                    "The trusted operation was cancelled before execution."
+                )
+            adapter = self._adapter
+            if adapter is None:
+                adapter = TrustedMetadataQueryAdapter(
+                    _BrokerQueryExecutor(self, operation),
+                    self.database_instance.resolved_path,
+                )
+                self._adapter = adapter
+            else:
+                # The facade is operation-bound so exact cancellation always
+                # targets the currently executing broker attempt.
+                adapter.bind_executor(_BrokerQueryExecutor(self, operation))
+        del supervisor
+        if operation.kind in (TrustedReadOperation.PLOT_DATASET, TrustedReadOperation.PLOT_PREFIX):
+            from .trusted_plot import plot_dataset, plot_prefix
+            executor = _BrokerQueryExecutor(self, operation)
+            if operation.kind is TrustedReadOperation.PLOT_DATASET:
+                return plot_dataset(executor, self, operation.payload[0])
+            if operation.plot_steps is None:
+                operation.plot_steps = plot_prefix(executor, *operation.payload)
+            # Amortise tiny pages into a short time slice. Otherwise a small
+            # run pays for another large run's much slower page after *each*
+            # of its own millisecond-scale pages. All yields end a source
+            # transaction; higher-priority jobs still preempt before queries.
+            slice_end = time.monotonic() + 0.05
+            while True:
+                executor.check_cancelled()
+                try:
+                    next(operation.plot_steps)
+                except StopIteration as complete:
+                    operation.plot_steps = None
+                    return complete.value
+                if time.monotonic() >= slice_end:
+                    return _OPERATION_YIELDED
+        if operation.kind is TrustedReadOperation.BOOTSTRAP:
+            return adapter.bootstrap()
+        if operation.kind is TrustedReadOperation.BASIC_PAGE:
+            return adapter.basic_run_page(
+                cast(int, operation.payload[0]),
+                cast(int, operation.payload[1]),
+            )
+        if operation.kind is TrustedReadOperation.REFRESH:
+            return adapter.refresh_new_runs(cast(int | None, operation.payload[0]))
+        if operation.kind is TrustedReadOperation.DATABASE_INFO:
+            return adapter.database_info()
+        run_id = cast(int, operation.payload[0])
+        if operation.kind is TrustedReadOperation.CHEAP_RUN:
+            return adapter.cheap_run(run_id)
+        if operation.kind is TrustedReadOperation.EXPENSIVE_RUN:
+            return adapter.expensive_run(run_id)
+        if operation.kind is TrustedReadOperation.SELECTED_RUN:
+            return adapter.selected_run_detail(run_id)
+        if operation.kind is TrustedReadOperation.DERIVED_SOURCE:
+            return adapter.derived_source_observation(
+                run_id,
+                database_instance=self.database_instance,
+                namespace=self._source_revision_namespace,
+            )
+        raise TrustedReadServiceError(
+            f"Unsupported trusted broker operation {operation.kind!r}."
+        )
+
+    def _ensure_supervisor(
+        self,
+        startup_remaining: float,
+    ) -> TrustedLiveReaderSupervisor:
+        with self._condition:
+            existing = self._supervisor
+        if existing is not None:
+            return existing
+        supervisor_options = dict(self._supervisor_options)
+        startup_key = "startup_timeout_seconds"
+        if startup_key not in supervisor_options:
+            supervisor_options["startup_timeout_seconds"] = startup_remaining
+        else:
+            configured_startup_timeout = supervisor_options[startup_key]
+            try:
+                configured_duration = (
+                    math.nan
+                    if isinstance(configured_startup_timeout, bool)
+                    else float(cast(Any, configured_startup_timeout))
+                )
+            except (TypeError, ValueError):
+                configured_duration = math.nan
+            if (
+                math.isfinite(configured_duration)
+                and configured_duration > 0
+                and startup_remaining < configured_duration
+            ):
+                supervisor_options[startup_key] = startup_remaining
+        supervisor = self._supervisor_factory(
+            self._database_path,
+            expected_database_instance=self._database_instance,
+            **supervisor_options,
+        )
+        with self._condition:
+            if self._supervisor is not None:
+                # This should be unreachable because only the dispatcher opens,
+                # but fail closed without leaking an unexpected second owner.
+                duplicate = supervisor
+                supervisor = self._supervisor
+            else:
+                duplicate = None
+                self._supervisor = supervisor
+                self._accepted_database_instance = supervisor.database_instance
+        if duplicate is not None:
+            try:
+                duplicate.close()
+            finally:
+                self._retain_supervisor_if_needed(duplicate)
+        return supervisor
+
+    def _required_supervisor(self) -> TrustedLiveReaderSupervisor:
+        with self._condition:
+            supervisor = self._supervisor
+        if supervisor is None:
+            raise TrustedReadServiceError(
+                "The trusted supervisor has not completed startup."
+            )
+        return supervisor
+
+    def _operation_remaining(self, operation: _OperationState) -> float:
+        remaining = operation.deadline - time.monotonic()
+        if remaining <= 0:
+            raise TrustedReadRequestDeadlineError(
+                f"Trusted operation {operation.operation_id} reached its deadline."
+            )
+        return min(remaining, _MAX_REQUEST_TIMEOUT_SECONDS)
+
+    def _prepare_supervisor_transaction(
+        self,
+        operation: _OperationState,
+        resume_executor: _BrokerQueryExecutor,
+    ) -> float:
+        """Yield at a job-free boundary to one strictly higher-priority request."""
+
+        with self._condition:
+            if self._active_operation is not operation:
+                raise TrustedReadServiceError(
+                    "The supervisor transaction does not match the active operation."
+                )
+            if self._closing or operation.cancel_underlying:
+                raise TrustedReadRequestCancelledError(
+                    "The trusted operation was cancelled before its next transaction."
+                )
+            if operation.supervisor_job is not None:
+                raise TrustedReadServiceError(
+                    "Cooperative scheduling requires a job-free transaction boundary."
+                )
+            if operation.force_next_transaction:
+                operation.force_next_transaction = False
+                nested = None
+            else:
+                nested = self._next_higher_priority_operation_locked(
+                    operation.priority,
+                )
+            if nested is not None:
+                if (
+                    operation.priority >= TrustedReadPriority.VISIBLE_CHEAP
+                    and nested.priority < TrustedReadPriority.VISIBLE_CHEAP
+                ):
+                    self._foreground_dispatch_burst += 1
+                nested.status = "active"
+                self._active_operation = nested
+                self._compact_heap_locked()
+
+        if nested is not None:
+            result: object = _NO_RESULT
+            nested_error: BaseException | None = None
+            try:
+                result = self._execute_operation(nested)
+            except BaseException as error:
+                nested_error = error
+            with self._condition:
+                self._finish_operation_locked(
+                    nested,
+                    result,
+                    nested_error,
+                )
+                self._active_operation = operation
+                self._condition.notify_all()
+            adapter = self._adapter
+            if adapter is not None:
+                adapter.bind_executor(resume_executor)
+            if nested_error is not None and self._terminal_session_error(nested_error):
+                raise nested_error
+
+        with self._condition:
+            if self._active_operation is not operation:
+                raise TrustedReadServiceError(
+                    "The original operation was not restored after scheduling."
+                )
+            if self._closing or operation.cancel_underlying:
+                raise TrustedReadRequestCancelledError(
+                    "The trusted operation was cancelled before its next transaction."
+                )
+        return self._operation_remaining(operation)
+
+    def _finish_operation_locked(
+        self,
+        operation: _OperationState,
+        result: object,
+        error: BaseException | None,
+    ) -> None:
+        operation.supervisor_job = None
+        if self._active_operation is operation:
+            self._active_operation = None
+        if result is _OPERATION_YIELDED and error is None:
+            if self._closing or operation.cancel_underlying:
+                error = TrustedReadRequestCancelledError("The plot capture was cancelled.")
+            else:
+                # Move to the tail of this priority, without recursive calls.
+                # Other plots get a turn at every page/BLOB boundary.
+                operation.status = "queued"
+                operation.sequence = self._next_sequence
+                self._next_sequence += 1
+                self._compact_heap_locked()
+                self._condition.notify_all()
+                return
+        self._retire_plot_steps_locked(operation)
+        operation.status = "finished"
+        if self._coalesced.get(operation.coalesce_key) == operation.operation_id:
+            self._coalesced.pop(operation.coalesce_key, None)
+        self._operations.pop(operation.operation_id, None)
+        now = time.monotonic()
+        for request_id in tuple(operation.subscribers):
+            state = self._requests.get(request_id)
+            if state is None or state.done.is_set():
+                continue
+            publication_error = error
+            publication_result = result
+            if publication_error is None and now >= state.identity.deadline:
+                publication_result = _NO_RESULT
+                publication_error = TrustedReadRequestDeadlineError(
+                    f"Trusted request {request_id} reached its deadline."
+                )
+            self._complete_request_locked(
+                state,
+                result=publication_result,
+                error=publication_error,
+            )
+        if error is not None and self._terminal_session_error(error):
+            self._fatal_error = error
+            self._closing = True
+            self._fail_queued_locked(error)
+        self._condition.notify_all()
+
+    def _wait_supervisor_job(
+        self,
+        operation: _OperationState,
+        job: TrustedLiveJob[_ResultT],
+        wait_timeout: float,
+    ) -> _ResultT:
+        wake_control = False
+        with self._condition:
+            if self._active_operation is not operation:
+                raise TrustedReadServiceError(
+                    "The supervisor job does not match the active broker operation."
+                )
+            operation.supervisor_job = job
+            wake_control = operation.cancel_underlying or self._closing
+        if wake_control:
+            self._control_wakeup.set()
+        try:
+            return self._required_supervisor().wait(job, timeout=wait_timeout)
+        finally:
+            with self._condition:
+                if operation.supervisor_job is job:
+                    operation.supervisor_job = None
+                if operation.priority >= TrustedReadPriority.VISIBLE_CHEAP:
+                    # A nested yield runs at most one complete foreground
+                    # operation before returning here for a real suspended
+                    # background transaction.  That progress ends the burst.
+                    self._foreground_dispatch_burst = 0
+
+    def _control_loop(self) -> None:
+        fatal_supervisor: TrustedLiveReaderSupervisor | None = None
+        fatal_job: TrustedLiveJob[Any] | None = None
+        try:
+            while True:
+                # Clear before observing broker state: a mutation after this
+                # snapshot leaves the event set and cannot be lost between an
+                # Event.wait() return and a subsequent clear().
+                self._control_wakeup.clear()
+                with self._condition:
+                    self._expire_queued_requests_locked(time.monotonic())
+                    if self._control_stop.is_set():
+                        return
+                    operation = self._active_operation
+                    supervisor = self._supervisor
+                    job = None if operation is None else operation.supervisor_job
+                    should_cancel = bool(
+                        operation is not None
+                        and job is not None
+                        and (operation.cancel_underlying or self._closing)
+                        and job is not self._control_cancel_job
+                    )
+                    if should_cancel:
+                        self._control_cancel_job = job
+                    next_deadline = self._next_queued_deadline_locked()
+                if should_cancel and supervisor is not None and job is not None:
+                    try:
+                        supervisor.cancel(job)
+                    except TrustedLiveCancelledError:
+                        pass
+                    except BaseException as error:
+                        with self._condition:
+                            # The exact job can finish after the unlocked
+                            # snapshot above, and the dispatcher can then close
+                            # its supervisor before cancel() obtains the Stage 3
+                            # lock.  A resulting closed/stale-job exception says
+                            # nothing about the current session.  Terminalise
+                            # only while this service still owns the same live
+                            # supervisor operation and is not already retiring.
+                            still_current = bool(
+                                not self._closing
+                                and not self._closed
+                                and self._supervisor is supervisor
+                                and operation is not None
+                                and self._active_operation is operation
+                                and operation.supervisor_job is job
+                            )
+                            if still_current:
+                                if self._fatal_error is None:
+                                    self._fatal_error = error
+                                self._closing = True
+                                self._fail_all_locked(error)
+                                self._condition.notify_all()
+                    continue
+
+                timeout = (
+                    None
+                    if next_deadline is None
+                    else max(0.0, next_deadline - time.monotonic())
+                )
+                self._control_wakeup.wait(timeout)
+        except BaseException as control_error:
+            with self._condition:
+                if self._fatal_error is None:
+                    self._fatal_error = control_error
+                self._closing = True
+                operation = self._active_operation
+                fatal_supervisor = self._supervisor
+                fatal_job = None if operation is None else operation.supervisor_job
+                self._fail_all_locked(control_error)
+                self._condition.notify_all()
+            # A control-loop fault must publish terminal request outcomes before
+            # attempting bounded Stage 3 cancellation.  Cancellation remains
+            # off the GUI thread and any second failure is diagnostic only.
+            if fatal_supervisor is not None and fatal_job is not None:
+                try:
+                    fatal_supervisor.cancel(fatal_job)
+                except BaseException:
+                    _LOG.exception(
+                        "Trusted read-service control failure cleanup also failed"
+                    )
+        finally:
+            self._control_done.set()
+
+    def _close_supervisor_from_dispatcher(self) -> None:
+        with self._condition:
+            supervisor, self._supervisor = self._supervisor, None
+            self._closing_supervisor = supervisor
+        if supervisor is None:
+            return
+        try:
+            supervisor.close()
+        except BaseException as error:
+            with self._condition:
+                self._close_error = error
+        finally:
+            self._retain_supervisor_if_needed(supervisor)
+            with self._condition:
+                if self._closing_supervisor is supervisor:
+                    self._closing_supervisor = None
+
+    def _owned_supervisors_locked(self) -> tuple[TrustedLiveReaderSupervisor, ...]:
+        supervisors: list[TrustedLiveReaderSupervisor] = []
+        for supervisor in (
+            self._supervisor,
+            self._closing_supervisor,
+            *self._retained_supervisors,
+        ):
+            if supervisor is not None and all(
+                supervisor is not owned for owned in supervisors
+            ):
+                supervisors.append(supervisor)
+        return tuple(supervisors)
+
+    @staticmethod
+    def _supervisor_resource_liveness(
+        supervisor: TrustedLiveReaderSupervisor,
+        *,
+        conservative_on_error: bool = True,
+    ) -> TrustedLiveSupervisorLiveness:
+        probe = getattr(supervisor, "resource_liveness", None)
+        if callable(probe):
+            try:
+                snapshot = probe()
+                if not isinstance(snapshot, TrustedLiveSupervisorLiveness):
+                    raise TypeError("Invalid supervisor resource-liveness snapshot.")
+                return snapshot
+            except BaseException:
+                if not conservative_on_error:
+                    raise
+                # Losing observability must never be mistaken for successful
+                # cleanup: retain the supervisor and report every dimension
+                # conservatively alive.
+                return TrustedLiveSupervisorLiveness(
+                    helper_pid=None,
+                    process_alive=True,
+                    receiver_alive=True,
+                    open_endpoints=1,
+                    active_incarnation=True,
+                    unreaped_incarnation=True,
+                    active_job=True,
+                    closing=True,
+                    closed=False,
+                )
+        if conservative_on_error:
+            try:
+                helper_alive = supervisor.helper_alive
+            except BaseException:
+                helper_alive = True
+            try:
+                helper_pid = supervisor.helper_pid
+            except BaseException:
+                helper_pid = None
+            try:
+                closed = supervisor.closed
+            except BaseException:
+                closed = not helper_alive
+        else:
+            helper_alive = supervisor.helper_alive
+            helper_pid = supervisor.helper_pid
+            closed = supervisor.closed
+        # Compatibility for supervisor test doubles predating the resource
+        # API.  Production supervisors always take the exact path above.
+        return TrustedLiveSupervisorLiveness(
+            helper_pid=helper_pid,
+            process_alive=helper_alive,
+            receiver_alive=False,
+            open_endpoints=0,
+            active_incarnation=helper_alive,
+            unreaped_incarnation=False,
+            active_job=False,
+            closing=False,
+            closed=closed,
+        )
+
+    def _retain_supervisor_if_needed(
+        self,
+        supervisor: TrustedLiveReaderSupervisor,
+    ) -> bool:
+        snapshot = self._supervisor_resource_liveness(supervisor)
+        with self._condition:
+            retained = any(supervisor is item for item in self._retained_supervisors)
+            if snapshot.resources_owned and not retained:
+                self._retained_supervisors.append(supervisor)
+            elif not snapshot.resources_owned and retained:
+                self._retained_supervisors = [
+                    item
+                    for item in self._retained_supervisors
+                    if item is not supervisor
+                ]
+        return snapshot.resources_owned
+
+    def _reap_retained_supervisors_once(self) -> bool:
+        with self._condition:
+            supervisors = tuple(self._retained_supervisors)
+        pending = False
+        for supervisor in supervisors:
+            reap = getattr(supervisor, "reap_closed_resources", None)
+            if callable(reap):
+                try:
+                    snapshot = reap()
+                    if not isinstance(snapshot, TrustedLiveSupervisorLiveness):
+                        raise TypeError("Invalid supervisor resource-reap snapshot.")
+                except BaseException:
+                    snapshot = self._supervisor_resource_liveness(supervisor)
+            else:
+                snapshot = self._supervisor_resource_liveness(supervisor)
+            if snapshot.resources_owned:
+                pending = True
+                continue
+            with self._condition:
+                self._retained_supervisors = [
+                    item
+                    for item in self._retained_supervisors
+                    if item is not supervisor
+                ]
+        return pending
+
+    def _complete_request_locked(
+        self,
+        state: _RequestState,
+        *,
+        result: object = _NO_RESULT,
+        error: BaseException | None = None,
+    ) -> None:
+        if state.done.is_set():
+            return
+        state.result = result
+        state.error = error
+        state.done.set()
+        self._requests.pop(state.identity.request_id, None)
+        state.request = None
+        self._release_request_slot_locked(state)
+
+    def _release_request_slot_locked(self, state: _RequestState) -> None:
+        if state.slot_released:
+            return
+        state.slot_released = True
+        if self._request_slots_in_use <= 0:
+            raise TrustedReadServiceError(
+                "The broker request-capacity accounting became inconsistent."
+            )
+        self._request_slots_in_use -= 1
+
+    def _discard_operation_locked(self, operation: _OperationState) -> None:
+        self._retire_plot_steps_locked(operation)
+        operation.status = "discarded"
+        self._operations.pop(operation.operation_id, None)
+        if self._coalesced.get(operation.coalesce_key) == operation.operation_id:
+            self._coalesced.pop(operation.coalesce_key, None)
+        self._compact_heap_locked()
+
+    def _retire_plot_steps_locked(self, operation: _OperationState) -> None:
+        if operation.plot_steps is not None:
+            self._retired_plot_steps.append(operation.plot_steps)
+            operation.plot_steps = None
+
+    def _close_retired_plot_steps(self) -> None:
+        # Cancellation/expiry may originate on the GUI or control thread.
+        # Private SQLite handles must be closed on their creating dispatcher.
+        with self._condition:
+            retired, self._retired_plot_steps = self._retired_plot_steps, []
+        for steps in retired:
+            try:
+                steps.close()
+            except BaseException as error:
+                # A private-spool cleanup error must not strand the helper or
+                # prevent other suspended captures from releasing resources.
+                with self._condition:
+                    if self._close_error is None:
+                        self._close_error = error
+                _LOG.exception("Trusted plot spool cleanup failed")
+                self.close_async()
+
+    def _fail_queued_locked(
+        self,
+        cause: BaseException,
+    ) -> None:
+        for operation in list(self._operations.values()):
+            if operation.status != "queued":
+                continue
+            for request_id in tuple(operation.subscribers):
+                state = self._requests.get(request_id)
+                if state is None:
+                    continue
+                error = TrustedReadSessionFailedError(
+                    "A terminal trusted-reader failure invalidated this queued "
+                    f"request ({type(cause).__name__}: {cause})."
+                )
+                error.__cause__ = cause
+                self._complete_request_locked(state, error=error)
+            operation.subscribers.clear()
+            self._discard_operation_locked(operation)
+
+    def _fail_all_locked(self, cause: BaseException) -> None:
+        """Terminally complete every request and clear all broker accounting."""
+
+        for state in tuple(self._requests.values()):
+            error = TrustedReadSessionFailedError(
+                "A fatal trusted read-service loop failure invalidated this "
+                f"request ({type(cause).__name__}: {cause})."
+            )
+            error.__cause__ = cause
+            self._complete_request_locked(state, error=error)
+
+        for operation in self._operations.values():
+            self._retire_plot_steps_locked(operation)
+            operation.cancel_underlying = True
+            operation.supervisor_job = None
+            operation.subscribers.clear()
+            operation.status = "failed"
+        self._operations.clear()
+        self._coalesced.clear()
+        self._heap.clear()
+        self._active_operation = None
+        self._foreground_dispatch_burst = 0
+        self._control_cancel_job = None
+        if self._requests or self._request_slots_in_use:
+            raise TrustedReadServiceError(
+                "Fatal broker cleanup left live request-capacity accounting."
+            )
+
+    @staticmethod
+    def _terminal_session_error(error: BaseException) -> bool:
+        # Broker-local cancellation/deadline outcomes do not describe helper
+        # health.  For errors returned by Stage 3, mirror its protocol
+        # taxonomy exactly: these five operation failures are the complete
+        # reusable set and every other outcome retires this broker session.
+        if isinstance(
+            error,
+            (
+                TrustedReadRequestCancelledError,
+                TrustedReadRequestDeadlineError,
+            ),
+        ):
+            return False
+        return not isinstance(
+            error,
+            (
+                TrustedLiveSqlRejectedError,
+                TrustedLiveQueryError,
+                TrustedLiveResultLimitError,
+                TrustedLiveBusyTimeoutError,
+                TrustedLiveCancelledError,
+            ),
+        )
+
+
+__all__ = [
+    "SNAPSHOT_FALLBACK_MODE",
+    "TRUSTED_LIVE_MODE",
+    "TrustedLiveReadService",
+    "TrustedReadOperation",
+    "TrustedReadPriority",
+    "TrustedReadQueueFullError",
+    "TrustedReadRequest",
+    "TrustedReadRequestCancelledError",
+    "TrustedReadRequestDeadlineError",
+    "TrustedReadRequestIdentity",
+    "TrustedReadServiceClosedError",
+    "TrustedReadServiceError",
+    "TrustedReadServiceLiveness",
+    "TrustedReadSessionFailedError",
+]