fix(backend): resolve the nine findings of the backend-block review

The three silent ones first. A 0.9 config that fails to convert used to boot an
empty config that looks exactly like a fresh install: the reason now reaches the
UI (GET /api/status.config_error), the activity log and an ERROR line, and the
.bak parachute is written on the failure branch too, so a later save from the
Advanced tab cannot destroy the original.

"PBS left powered on" was wrong in both directions -- every successful run
against an always-on PBS warned, and a sync route that left its target awake did
not. lease.release() returned one False for four situations, only two of which
cost power; it now returns a ReleaseOutcome that names the reason, which becomes
both the POWEROFF step's detail and RunContext.left_on. The interrupted-run path
keeps a step-derived rule, now paired per device and filtered by managed_power.

422 bodies echoed the whole config, secrets included: a config-level validator
raises at loc=(), so pydantic attached every token, the secret key, the password
hash, the SMTP and bot tokens as the error's input. One helper with
include_input=False now serves all three config-shaped 422 sites.

Also: a redaction placeholder with nothing to resolve against is rejected instead
of silently clearing the credential (a renamed device id, or a create from a
copied body); the ad-hoc "Run verify" asks for outdated_after=0, since None meant
"only never-verified" and skipped exactly the snapshots the button exists for;
the manual power-off holds the single-run lock so it cannot cut a vzdump that
started in the check-then-act gap; _current_run_id is cleared when a run ends, so
a stop landing between two runs cannot hit the wrong one; and the pre-migration
.bak is chmod 0600 like every other secret-bearing file.

Tests: 617 passed, 2 skipped. Every finding was reproduced against the real code
before the fix, and each new test confirmed failing on the pre-fix code.

backend/app/api/_config_edit.py

@@ -21,6 +21,25 @@ from ..core.config_store import ConfigStore
 from ..core.scheduler import Scheduler, validate_cron
 
 
+def validation_error(exc: ValidationError) -> HTTPException:
+    """A 422 carrying pydantic's error list — **without** the values that failed.
+
+    ``include_input=False`` is the whole point: a ``model_validator(mode="after")`` on
+    ``Config`` raises at ``loc=()``, so pydantic attaches the entire validated config as the
+    error's ``input``. Serialised, that ships every API token, ``app.secret_key``, the
+    password hash, the SMTP password and the bot token to the browser's network tab and any
+    proxy log in between. The config-level cross-checks (a route pointing at a device that
+    doesn't exist, an External route onto an unmanaged PBS) are ordinary user mistakes, so
+    this is a body users really do see. ``include_url`` only drops a docs link.
+
+    Every config-shaped 422 goes through here so a new endpoint inherits the safe default.
+    """
+    return HTTPException(
+        status_code=422,  # the literal avoids the deprecated HTTP_422_UNPROCESSABLE_ENTITY
+        detail=jsonable_encoder(exc.errors(include_input=False, include_url=False)),
+    )
+
+
 def save_section(
     store: ConfigStore, scheduler: Scheduler, section: str, value: list[dict[str, Any]]
 ) -> Config:
@@ -34,9 +53,8 @@ def save_section(
     try:
         new_config = Config.model_validate(raw)
     except ValidationError as exc:
-        # 422 mirrors FastAPI's own body-validation responses (the literal avoids the
-        # deprecated HTTP_422_UNPROCESSABLE_ENTITY constant name).
-        raise HTTPException(status_code=422, detail=jsonable_encoder(exc.errors())) from exc
+        # 422 mirrors FastAPI's own body-validation responses.
+        raise validation_error(exc) from exc
     check_route_crons(new_config, store.config)
     try:
         store.replace(new_config)

backend/app/api/config.py

@@ -11,7 +11,6 @@ from typing import Any
 
 import yaml
 from fastapi import APIRouter, Depends, HTTPException, status
-from fastapi.encoders import jsonable_encoder
 from pydantic import BaseModel, ValidationError
 
 from ..config import (
@@ -25,7 +24,7 @@ from ..config import (
 from ..connectors.errors import WolError
 from ..connectors.wol import normalize_mac
 from ..core.config_store import ConfigStore
-from ._config_edit import check_route_crons
+from ._config_edit import check_route_crons, validation_error
 from .deps import Scheduler, get_config_store, get_scheduler, require_auth
 
 router = APIRouter(dependencies=[Depends(require_auth)], tags=["config"])
@@ -66,9 +65,8 @@ def _apply_config(
     try:
         new_config = Config.model_validate(merged)
     except ValidationError as exc:
-        # 422 to mirror FastAPI's own body-validation responses (literal avoids the
-        # deprecated HTTP_422_UNPROCESSABLE_ENTITY constant name).
-        raise HTTPException(status_code=422, detail=jsonable_encoder(exc.errors())) from exc
+        # 422 to mirror FastAPI's own body-validation responses.
+        raise validation_error(exc) from exc
 
     # Reject a newly-set unparseable route cron before it can be persisted (BE-B1): the
     # route would silently never fire, which is the failure mode hardest to notice.

backend/app/api/devices.py

@@ -17,15 +17,14 @@ from __future__ import annotations
 from typing import Any, Literal
 
 from fastapi import APIRouter, Depends, HTTPException, status
-from fastapi.encoders import jsonable_encoder
 from pydantic import BaseModel, ValidationError
 
-from ..config import PbsDevice, PveDevice, redact, restore_secrets_from
+from ..config import PbsDevice, PveDevice, RedactionError, redact, restore_secrets_from
 from ..connectors.errors import ConnectorError, WolError
 from ..core.config_store import ConfigStore
 from ..db.models import RunTrigger
 from ..jobs import AlreadyRunningError
-from ._config_edit import save_section
+from ._config_edit import save_section, validation_error
 from .deps import (
     JobService,
     Scheduler,
@@ -101,7 +100,12 @@ def create_device(
     scheduler: Scheduler = Depends(get_scheduler),
 ) -> dict[str, Any]:
     section, model = _kind(kind)
-    device = _validate(model, body)
+    # Against an empty stored mapping: a *new* device has no secrets to restore, so any
+    # ***REDACTED*** in the body — a "duplicate this device" action, or a body copy-pasted
+    # from GET /api/devices — fails loudly here instead of being stored as the literal
+    # placeholder, which GET then re-masks so the corruption is invisible until a
+    # connection test fails with a 502 that gives no hint why.
+    device = _validate(model, _resolve(body, {}))
     if any(d.id == device.id for d in getattr(store.config, section)):
         raise HTTPException(
             status_code=status.HTTP_409_CONFLICT, detail=f"Device '{device.id}' already exists"
@@ -124,7 +128,7 @@ def update_device(
     index = _find(store, section, device_id)
     # Resolve ***REDACTED*** against *this* device's stored values: the client only ever
     # echoes back a secret it didn't change.
-    device = _validate(model, restore_secrets_from(body, _dump(store, section)[index]))
+    device = _validate(model, _resolve(body, _dump(store, section)[index]))
     devices = _dump(store, section)
     devices[index] = device.model_dump(mode="python")
     save_section(store, scheduler, section, devices)
@@ -238,19 +242,27 @@ def power(
             ) from exc
         return PowerResult(ok=True)
 
-    # Power-off. The lease is the authority on "is anything using this box": a run holding
-    # it would be cut off mid-vzdump.
-    # ponytail: a run could still take the lease in the gap between this check and the SSH
-    # command. Harmless in practice — the lease probes the box, finds it down and wakes it
-    # again — and closing it properly means holding the lease from here, which would let a
-    # failed HTTP request strand it.
-    if job_service.lease.state(pbs_id).holders:
-        raise HTTPException(
-            status_code=status.HTTP_409_CONFLICT,
-            detail=f"A run is using '{pbs_id}'; cannot power it off",
-        )
+    # Power-off, under the single-run lock. `exclusive()` exists for exactly this: a run
+    # holds that lock for its whole life, so while we hold it none can start — and
+    # `power_off_now` has no idle-wait and no refcount check by design ("a click means
+    # now"). Without the lock the check is a check-then-act: a scheduled route starts in
+    # the gap, `_bring_up` probes, finds the box still up, skips the Wake-on-LAN, vzdump
+    # begins, and the SSH poweroff lands mid-backup. The lease's own self-healing can't
+    # help there — the probe has already happened.
+    # ponytail: the lock is held for the seconds an SSH connect takes, so a scheduled run
+    # firing in that window waits its turn instead of being rejected. Fine for a button.
     try:
-        job_service.lease.power_off_now(device)
+        with job_service.exclusive():
+            # Belt and braces: a lease is only ever taken inside the lock we now hold, so
+            # this cannot fire in production — but it names the box when it does.
+            if job_service.lease.state(pbs_id).holders:
+                raise HTTPException(
+                    status_code=status.HTTP_409_CONFLICT,
+                    detail=f"A run is using '{pbs_id}'; cannot power it off",
+                )
+            job_service.lease.power_off_now(device)
+    except AlreadyRunningError as exc:
+        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail=str(exc)) from exc
     except ConnectorError as exc:
         raise HTTPException(status_code=status.HTTP_502_BAD_GATEWAY, detail=str(exc)) from exc
     return PowerResult(ok=True)
@@ -304,10 +316,19 @@ def run_maintenance(
 # --- helpers -----------------------------------------------------------------
 
 
+def _resolve(body: dict[str, Any], stored: dict[str, Any]) -> dict[str, Any]:
+    """Fill in the ***REDACTED*** placeholders the client echoed back, 422 if one can't be
+    resolved. ``stored`` is the device being edited, or ``{}`` for a create."""
+    try:
+        return restore_secrets_from(body, stored)
+    except RedactionError as exc:
+        raise HTTPException(status_code=422, detail=str(exc)) from exc
+
+
 def _validate(model: type[PveDevice] | type[PbsDevice], body: dict[str, Any]):
     """Validate a device body by hand rather than as a typed parameter: the handlers are
     shared between the two kinds, so which model applies is only known at call time."""
     try:
         return model.model_validate(body)
     except ValidationError as exc:
-        raise HTTPException(status_code=422, detail=jsonable_encoder(exc.errors())) from exc
+        raise validation_error(exc) from exc

backend/app/api/status.py

@@ -14,6 +14,7 @@ from fastapi import APIRouter, Depends
 from pydantic import BaseModel
 from sqlalchemy.orm import Session
 
+from .. import config as config_mod
 from ..core.config_store import ConfigStore
 from ..db import get_session
 from . import _probe
@@ -91,6 +92,10 @@ class StatusResponse(BaseModel):
     pves: list[PveState] = []
     pbss: list[PbsState] = []
     last_run: RunSummary | None = None
+    #: Set when the 0.9 -> 1.0 config migration was refused at startup: the config in use
+    #: has no devices and no routes, so the UI must say why instead of looking like a fresh
+    #: install. Rendered as a persistent banner under the header.
+    config_error: str | None = None
 
 
 @router.get("/status", response_model=StatusResponse)
@@ -139,6 +144,8 @@ def get_status(
         pves=[PveState(id=pve.id, online=pve_online.get(pve.id, False)) for pve in config.pves],
         pbss=[_pbs_state(pbs, pbs_probes, job_service) for pbs in config.pbss],
         last_run=RunSummary.of(last) if last else None,
+        # Read live, not captured at import: a reload after the user fixes the file clears it.
+        config_error=config_mod.MIGRATION_ERROR,
     )
 
 

backend/app/config.py

@@ -38,6 +38,13 @@ SECRET_KEYS: frozenset[str] = frozenset(
 
 REDACTED = "***REDACTED***"
 
+#: Why the 0.9 -> 1.0 migration was refused on the last :func:`load_config`, or ``None``.
+#: A failed migration boots on a config with no devices and no routes, which looks exactly
+#: like a fresh install — so the reason has to outlive the log line: ``GET /api/status``
+#: reports it to the UI and startup writes it to the activity log. Process-wide because the
+#: fact is process-wide; every load clears it and decides again.
+MIGRATION_ERROR: str | None = None
+
 
 class _Base(BaseModel):
     # Reject unknown keys so typos in config.yaml surface as clear validation errors.
@@ -460,6 +467,8 @@ def load_config(path: Path | None = None) -> Config:
         raise FileNotFoundError(
             f"Config file not found at {p}. Copy config.example.yaml to config.yaml."
         )
+    global MIGRATION_ERROR
+    MIGRATION_ERROR = None  # re-decided below; a reload after a fix must clear it
     with p.open("r", encoding="utf-8") as fh:
         raw = yaml.safe_load(fh) or {}
     if not isinstance(raw, dict):
@@ -479,19 +488,28 @@ def _migrate_0_9(raw: dict[str, Any], path: Path) -> Config | None:
     """Convert a 0.9 config to the route model, back it up and save it.
 
     Returns ``None`` — meaning "carry on with the config as it is on disk" — if anything at
-    all goes wrong. A failed migration must never stop the app from booting (BE-B1), and it
-    costs nothing to defer: the 0.9 sections are still what the app reads, and the next
-    start will try again.
+    all goes wrong. A failed migration must never stop the app from booting (BE-B1).
+
+    But it is not a harmless deferral any more: in 0.9 the app kept running on the old
+    sections, while in 1.0 nothing reads them, so the fallback boots a config with no
+    devices and no routes — indistinguishable from a fresh install, with every schedule
+    silently gone. Hence :data:`MIGRATION_ERROR` (surfaced by ``GET /api/status`` and the
+    activity log) and the ``.bak``: the user can still rewrite ``config.yaml`` from the
+    Advanced tab while running empty, and without the copy that would destroy the only
+    0.9 config they have.
     """
     try:
         cfg = Config.model_validate(config_migrate.migrate(raw))
     except Exception as exc:  # noqa: BLE001 — any failure here degrades to "don't migrate"
-        log.warning(
-            "config: could not convert %s to the 1.0 route model (%s); starting on the "
-            "existing config and leaving it untouched",
-            path,
-            exc,
+        global MIGRATION_ERROR
+        MIGRATION_ERROR = (
+            f"Could not convert {path} to the 1.0 route model: {exc}. Joulenap started "
+            "with no devices and no routes — nothing is scheduled and no backup will run "
+            "until this is fixed. The file is untouched; a copy is at "
+            f"{path.name}{config_migrate.BACKUP_SUFFIX}."
         )
+        log.error("config: %s", MIGRATION_ERROR)
+        config_migrate.write_backup(path)
         return None
     config_migrate.write_backup(path)
     try:
@@ -661,7 +679,7 @@ def _restore_in_place(node: Any, current: Any) -> Any:
         for key, value in node.items():
             cur = current.get(key) if isinstance(current, dict) else None
             if key in SECRET_KEYS:
-                node[key] = _unmask(value, cur)
+                node[key] = _unmask(value, cur, key)
             else:
                 _restore_in_place(value, cur)
     elif isinstance(node, list):
@@ -687,7 +705,15 @@ def _match(item: Any, current: Any, index: int) -> Any:
     return current[index] if index < len(current) else None
 
 
-def _unmask(value: Any, current: Any) -> Any:
+def _unmask(value: Any, current: Any, key: str = "") -> Any:
+    """Resolve one secret field: ``REDACTED`` -> the stored value, anything else verbatim.
+
+    An unresolvable placeholder is an **error**, never an empty string. ``current`` is None
+    when the incoming item has no stored counterpart — a device whose ``id`` was renamed in
+    the YAML editor, or one being created — and the client is then echoing back a mask over
+    a secret we do not have. Turning that into ``""`` validates fine, returns 200, and
+    leaves the credential gone with nothing on screen to suggest it.
+    """
     if isinstance(value, list):
         # List secrets (custom_urls) are write-only and all-or-nothing to avoid the
         # index-positional corruption of the old per-entry masking:
@@ -701,7 +727,9 @@ def _unmask(value: Any, current: Any) -> Any:
         if not value:
             return []
         if all(v == REDACTED for v in value):
-            return list(current) if isinstance(current, list) else []
+            if not isinstance(current, list):
+                raise _unresolvable(key)
+            return list(current)
         if any(v == REDACTED for v in value):
             raise RedactionError(
                 "custom_urls must be sent in full (all real values) or left unchanged "
@@ -709,5 +737,17 @@ def _unmask(value: Any, current: Any) -> Any:
             )
         return value
     if value == REDACTED:
-        return current if current is not None else ""
+        if current is None:
+            raise _unresolvable(key)
+        return current
     return value
+
+
+def _unresolvable(key: str) -> RedactionError:
+    field = f"'{key}'" if key else "a secret"
+    return RedactionError(
+        f"{field} was sent as {REDACTED} but there is no stored value to restore it from. "
+        "This happens when an entry's id is renamed (the old entry's secrets don't follow "
+        "the new id) or when a new one is created from a copy. Send the real value, or an "
+        "empty string to clear it."
+    )

backend/app/config_migrate.py

@@ -87,16 +87,27 @@ def migrate(raw: dict[str, Any]) -> dict[str, Any]:
 def write_backup(path: Path) -> None:
     """Copy the pre-migration config beside itself as the rollback parachute.
 
+    Written whether the migration succeeded or was refused: a refused one leaves the app
+    running on an empty config, and the next save from the Advanced tab would overwrite
+    the 0.9 original with it.
+
     An existing backup is never overwritten: the first migration wins, so re-running this
     can't lose the true 0.9 original. A backup that can't be written (read-only mount) is
-    logged, not fatal — the migration only *adds* sections, so the old ones stay as the
-    in-file fallback either way.
+    logged, not fatal — the same mount stops the migrated config from being saved too, so
+    the file on disk stays 0.9 either way.
     """
     bak = path.with_name(path.name + BACKUP_SUFFIX)
     if bak.exists():
         return
+    # Local import: config.py imports this module, so a top-level one would be circular.
+    from .config import restrict_secret_file
+
     try:
         shutil.copyfile(path, bak)
+        # copyfile copies contents, not mode bits, so the copy would land 0644 under a
+        # container's default umask — holding every API token, secret_key, password hash,
+        # SMTP password and bot token, indefinitely (an existing .bak is never overwritten).
+        restrict_secret_file(bak)
         log.info("config: saved a pre-migration copy at %s", bak)
     except OSError as exc:
         log.warning("config: could not write %s (%s) — migrating anyway", bak, exc)

backend/app/jobs/lease.py

@@ -18,6 +18,7 @@ import ssl
 import threading
 from collections.abc import Callable
 from dataclasses import dataclass
+from enum import StrEnum
 
 from ..config import PbsDevice
 from ..connectors import net, tls
@@ -32,6 +33,21 @@ class PbsUnreachableError(RuntimeError):
     """The PBS never answered: it could not be woken, or it is an unmanaged box that is off."""
 
 
+class ReleaseOutcome(StrEnum):
+    """What became of a box when a run dropped its lease.
+
+    Only :attr:`LEFT_ON` means "still burning power, and nothing is going to fix that" —
+    the others are boxes Joulenap either put to sleep, never powers, or will power off at
+    the end of the run that still holds them. The value doubles as the POWEROFF step's
+    detail in the run timeline.
+    """
+
+    POWERED_OFF = "powered off"
+    STILL_NEEDED = "left on: still needed by another run"
+    UNMANAGED = "left on: Joulenap does not manage this box's power"
+    LEFT_ON = "left powered on"
+
+
 # --- device-shaped connector calls -------------------------------------------
 # jobs/deps.py has the same four operations bound to the 0.9 single-PBS ``Config``. These
 # take the device a route points at; the Config-shaped originals die with the old cycle.
@@ -180,37 +196,48 @@ class PowerLease:
             held.holders += 1
         return was_awake
 
-    def release(self, pbs: PbsDevice, *, power_off: bool = True) -> bool:
-        """Drop this run's hold and power the box down if it is now safe to. Returns whether
-        it actually powered off.
+    def release(self, pbs: PbsDevice, *, power_off: bool = True) -> ReleaseOutcome:
+        """Drop this run's hold and power the box down if it is now safe to. Returns *why*
+        the box ended up in the state it did.
 
         ``power_off`` is the caller's policy (a failed run leaves the PBS on for inspection,
         a manual run honours the "power off when finished" toggle). The lease adds the
         conditions the caller cannot see: no other holder, and no *queued* route that still
         needs this box.
+
+        Four outcomes rather than a bool because only two of them cost the user power, and
+        the lease is the only place that can tell them apart — a caller re-deriving "was
+        that a real 'left on'?" would have to duplicate every condition below.
         """
         with self._lock:
             held = self._state.get(pbs.id)
             if held is None or not held.holders:
                 log.warning("Release of an unheld lease on PBS %s — ignoring", pbs.id)
-                return False
+                # A bug, not a policy: say "left on" so it is at least visible.
+                return ReleaseOutcome.LEFT_ON
             held.holders -= 1
             remaining = held.holders
             if not remaining:
                 del self._state[pbs.id]
 
+        # Order matters for the *reason*, not for the action — every branch below stops the
+        # power-off equally. The caller's policy is checked last on purpose: "this run
+        # doesn't want to power off" is only the real explanation once the box is one we
+        # could have powered off and nobody else needs it.
         if remaining:
             log.info("PBS %s still held by %d run(s); leaving it on", pbs.id, remaining)
-            return False
-        if not power_off:
-            return False
+            return ReleaseOutcome.STILL_NEEDED
         if not pbs.managed_power:
             # An always-on / cloud-hosted PBS: Joulenap never touches its power.
-            return False
+            return ReleaseOutcome.UNMANAGED
         if pbs.id in self._pending():
             log.info("PBS %s is needed by a queued route; leaving it on", pbs.id)
-            return False
-        return self._power_off(pbs)
+            return ReleaseOutcome.STILL_NEEDED
+        if not power_off:
+            return ReleaseOutcome.LEFT_ON
+        # False here is a busy box we declined to cut off, or a poweroff that errored — both
+        # leave it burning power.
+        return ReleaseOutcome.POWERED_OFF if self._power_off(pbs) else ReleaseOutcome.LEFT_ON
 
     # --- internals -----------------------------------------------------------
 

backend/app/jobs/route_cycle.py

@@ -265,9 +265,11 @@ def run_pbs_maintenance(
         if action == "gc":
             _route_gc_step(pbs, recorder, deps)
         elif action == "verify":
-            # Everything, not just the recently-changed: an ad-hoc verify is a deliberate
-            # "check this box now", and reverify_days is a per-route pacing knob.
-            _route_verify_step(pbs, recorder, deps, outdated_after=None)
+            # 0 means "everything" — see _route_verify_step. An ad-hoc verify is a deliberate
+            # "check this box now" after a restore or a disk scare, so it must re-read the
+            # older snapshots too; None would mean "only never-verified", i.e. skip exactly
+            # the ones the user is worried about. reverify_days is the per-route pacing knob.
+            _route_verify_step(pbs, recorder, deps, outdated_after=0)
         else:
             raise CycleAbort(f"unsupported maintenance action '{action}'")
         return _route_read_datastore(pbs, recorder, deps)

backend/app/jobs/service.py

@@ -31,7 +31,7 @@ from ..db.models import LogLevel, RunKind, RunStatus, RunTrigger, StepName, Step
 from ..db.prune import PruneResult, prune_history
 from ..notify.messages import RunContext
 from .deps import CycleDeps
-from .lease import LeaseDeps, PbsUnreachableError, PowerLease
+from .lease import LeaseDeps, PbsUnreachableError, PowerLease, ReleaseOutcome
 from .recorder import RunRecorder
 from .route_cycle import RUN_KINDS, run_pbs_maintenance, run_route
 
@@ -276,6 +276,11 @@ class JobService:
             finally:
                 with self._state_lock:
                     self._current = None
+                    # Cleared with it, or a stop aimed at the run that just ended would pass
+                    # cancel()'s guard and land on the next one (or be swallowed): _start
+                    # sets the new id only after the DB insert, leaving a window where the
+                    # stale id still matches while another run holds the lock.
+                    self._current_run_id = None
 
     def _execute(self, item: QueuedRun) -> None:
         """Take the run's power leases, do its job, release them, then notify.
@@ -314,9 +319,10 @@ class JobService:
                 # The cycle sets the run's final status itself and hands back what to say
                 # about it (None = cancelled, say nothing).
                 ctx = item.job(config, subject, recorder, self.deps)
-                self._release_all(item, held, recorder)
+                left_on = self._release_all(item, held, recorder)
                 held = []
                 if ctx is not None:
+                    ctx.left_on = left_on
                     self._notify(ctx, recorder)
         finally:
             # Only reached with leases still held when the job raised out of the block.
@@ -364,30 +370,38 @@ class JobService:
 
     def _release_all(
         self, item: QueuedRun, devices: list[PbsDevice], recorder: RunRecorder | None
-    ) -> None:
-        """Drop every lease this run holds, recording each power-off decision.
+    ) -> list[str]:
+        """Drop every lease this run holds, recording each power-off decision. Returns the
+        ids of the boxes genuinely left burning power, for the notification's warning.
 
         Devices are released independently: a sync route holds two leases and each box may
-        have a different answer (another holder, a queued route, unmanaged power).
+        have a different answer (another holder, a queued route, unmanaged power) — which
+        is why the warning cannot be a single fact about the run.
         ``recorder=None`` is the crash path — the leases must still be dropped, but the run
         row is already being finalised, so there is nothing left to record against.
         """
         succeeded = recorder is not None and recorder.run.status == RunStatus.SUCCESS
         power_off = self._power_off_policy(item, succeeded=succeeded)
         multi = len(devices) > 1
+        left_on: list[str] = []
         for device in devices:
             if recorder is None:
-                self.lease.release(device, power_off=power_off)
+                if self.lease.release(device, power_off=power_off) is ReleaseOutcome.LEFT_ON:
+                    left_on.append(device.id)
                 continue
             with recorder.step(StepName.POWEROFF, label=device.id if multi else None) as step:
-                if self.lease.release(device, power_off=power_off):
+                outcome = self.lease.release(device, power_off=power_off)
+                step.detail = str(outcome)
+                if outcome is ReleaseOutcome.POWERED_OFF:
                     continue
                 # Not a failure: leaving the box on is the *correct* outcome after a failed
-                # run, for an unmanaged device, or while another route still needs it. The
-                # notification's "PBS left powered on" warning keys on this step not being
-                # SUCCESS, so the user still hears about it when it costs them power.
+                # run, for an unmanaged device, or while another route still needs it — the
+                # detail says which. Only LEFT_ON costs the user power with nobody left to
+                # fix it, so only that one earns the notification's warning.
                 step.status = StepStatus.SKIPPED
-                step.detail = "left powered on"
+                if outcome is ReleaseOutcome.LEFT_ON:
+                    left_on.append(device.id)
+        return left_on
 
     def _notify(self, ctx: RunContext, recorder: RunRecorder) -> None:
         """Send the result notification. A delivery failure is logged, never fatal — the run

backend/app/main.py

@@ -19,6 +19,7 @@ from fastapi.staticfiles import StaticFiles
 from starlette.middleware.sessions import SessionMiddleware
 
 from . import __version__
+from . import config as config_mod
 from .api import api_router
 from .api import metrics as metrics_api
 from .config import Config
@@ -26,6 +27,7 @@ from .core.config_store import ConfigStore
 from .core.ratelimit import LoginRateLimiter
 from .core.scheduler import Scheduler
 from .db import init_db, session_scope
+from .db.models import LogEvent, LogLevel
 from .db.startup import sweep_orphaned_runs
 from .jobs import JobService
 from .notify import NotificationService
@@ -59,6 +61,10 @@ async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     with session_scope() as session:
         swept = sweep_orphaned_runs(session)
         interrupted_alerts = [build_interrupted_message(store.config, run) for run in swept]
+        # The config was loaded before the DB existed, so a refused 0.9 migration can only
+        # reach the activity log from here. Run-less row: it belongs to no cycle.
+        if config_mod.MIGRATION_ERROR:
+            session.add(LogEvent(level=LogLevel.ERROR, message=config_mod.MIGRATION_ERROR))
     if swept:
         log.warning("Marked %d interrupted run(s) as failed at startup", len(swept))
     service = JobService(store)

backend/app/notify/messages.py

@@ -58,6 +58,11 @@ class RunContext:
     guests: GuestSummary | None = None
     #: When this route next fires, for the "Next scheduled run" line.
     next_at: datetime | None = None
+    #: PBS ids this run left awake and burning power, with nothing queued to shut them
+    #: down. Filled in by ``JobService`` after the leases are released: it is the only
+    #: place that knows *why* a box stayed on, and only some of the reasons cost energy
+    #: (an always-on box, or one another run still holds, cost nothing).
+    left_on: list[str] = field(default_factory=list)
 
 
 # event keys: success | failure | aborted | test
@@ -279,25 +284,52 @@ def _step_is(step: RunStep, name: StepName) -> bool:
     return step.name.split(":", 1)[0] == name.value
 
 
-def _pbs_left_on(run: Run) -> bool:
-    """True if the cycle woke a PBS but never powered it back off — so a box is still
-    burning energy and the user should check it.
+def _step_label(step: RunStep) -> str | None:
+    """The device a step names (``poweroff:pbs-02`` -> ``pbs-02``), or None when unlabelled
+    — a single-device run doesn't repeat which box it means."""
+    _, _, label = step.name.partition(":")
+    return label or None
 
-    The rule: a WAIT step succeeded (a PBS actually came up) **and** no POWEROFF step
-    succeeded. That single condition covers every "left on" case uniformly:
 
-      * success but power-off failed / was skipped (PBS busy) — POWEROFF present, not SUCCESS;
-      * failure after the PBS woke (vzdump/GC/verify errored) — no POWEROFF step at all;
-      * abort after wake (preflight free-space, no guests selected) — no POWEROFF step.
+def _pbs_left_on(config: Config, run: Run) -> bool:
+    """True if a PBS came up and nothing ever powered it back off — a box still burning
+    energy that the user should go and check.
 
-    An abort *before* the box came up (wake/wait timeout) leaves the WAIT step non-SUCCESS, so
-    the PBS is off and this correctly returns False — hence why it keys on WAIT, not on the
-    run status."""
-    woke = any(_step_is(s, StepName.WAIT) and s.status == StepStatus.SUCCESS for s in run.steps)
-    powered_off = any(
-        _step_is(s, StepName.POWEROFF) and s.status == StepStatus.SUCCESS for s in run.steps
-    )
-    return woke and not powered_off
+    Only for a run a restart interrupted: there the run row is all there is, and no
+    POWEROFF step was ever reached. A run that *finished* knows the answer exactly and
+    reports it through ``RunContext.left_on``, because "was it left on?" depends on facts
+    the timeline doesn't carry (whether another queued route still needs the box).
+
+    Two things the steps alone get wrong, both introduced by this same release:
+
+      * an **unmanaged** box (``managed_power: false``) is never Joulenap's to power down,
+        so a run against one must not warn — hence taking ``config``;
+      * a run holding **several** leases needs the WAIT and POWEROFF steps paired *per
+        device*, or one box's successful power-off hides another's that stayed up.
+
+    An abort *before* the box came up leaves the WAIT step non-SUCCESS, so the PBS is off
+    and this correctly returns False — hence why it keys on WAIT, not on the run status.
+    """
+    managed = {p.id for p in config.pbss if p.managed_power}
+    route = next((r for r in config.routes if r.id == run.route_id), None)
+    # An unlabelled step belongs to the run's only box — a route's target (a sync route
+    # labels both sides). An ad-hoc GC/verify records neither a label nor a route, so its
+    # box cannot be named: warn rather than stay silent about a box that may be awake.
+    default = route.target if route else None
+    powered_off = {
+        _step_label(s)
+        for s in run.steps
+        if _step_is(s, StepName.POWEROFF) and s.status == StepStatus.SUCCESS
+    }
+    for step in run.steps:
+        if not _step_is(step, StepName.WAIT) or step.status != StepStatus.SUCCESS:
+            continue
+        pbs_id = _step_label(step) or default
+        if pbs_id is not None and pbs_id not in managed:
+            continue
+        if _step_label(step) not in powered_off:
+            return True
+    return False
 
 
 #: Route kinds, for the body's ``Route:`` line. Localized because the kind is a user-facing
@@ -368,7 +400,9 @@ def build_run_message(ctx: RunContext) -> tuple[str, str]:
     if run.error:
         lines.append(f"{labels['error']}: {run.error}")
 
-    if _pbs_left_on(run):
+    # From the service, not from the steps: a finished run knows exactly which boxes it
+    # left burning power, and only those warrant the warning.
+    if ctx.left_on:
         lines.append(labels["pbs_left_on"])
 
     if next_at is not None:
@@ -432,15 +466,15 @@ def build_interrupted_message(config: Config, run: Run) -> tuple[str, str]:
     """``(title, body)`` for a run that a restart interrupted (swept to FAILURE at startup,
     BE-R2), in the configured language.
 
-    Adds the "PBS left powered on" warning when the box had actually woken before the crash
-    (WAIT succeeded, no POWEROFF) — the whole point of the alert: a normally-off box that a
-    crash left awake and burning power."""
+    Adds the "PBS left powered on" warning when a box Joulenap powers had actually woken
+    before the crash (WAIT succeeded, no matching POWEROFF) — the whole point of the alert:
+    a normally-off box that a crash left awake and burning power."""
     pack = _pack(config.app.language)
     labels = pack["_labels"]
     lines = [pack["interrupted"]["intro"]]
     if run.error:
         lines.append(f"{labels['error']}: {run.error}")
-    if _pbs_left_on(run):
+    if _pbs_left_on(config, run):
         lines.append(labels["pbs_left_on"])
         # This alert has no Duration line (the run's own span would span the whole downtime,
         # not the work), so the one interval worth reporting is how long the box has been

backend/tests/test_api.py

@@ -12,6 +12,7 @@ import pytest
 from fakes import FakeBox, FakePve, UnreachablePve, make_deps
 from fastapi.testclient import TestClient
 
+from app import config
 from app.config import load_config
 from app.connectors.pve import Guest
 from app.db import session_scope
@@ -106,6 +107,14 @@ def test_status_lists_the_next_run_of_every_armed_route(app_ctx):
     assert times == sorted(times)
 
 
+def test_status_reports_a_refused_config_migration(app_ctx, monkeypatch):
+    """An empty config looks exactly like a fresh install, so the UI has to be told why."""
+    client, _app = app_ctx
+    assert client.get("/api/status").json()["config_error"] is None
+    monkeypatch.setattr(config, "MIGRATION_ERROR", "routes.0.options.bwlimit: bad")
+    assert client.get("/api/status").json()["config_error"] == "routes.0.options.bwlimit: bad"
+
+
 def test_status_pill_says_paused_when_the_kill_switch_is_off(app_ctx):
     client, app = app_ctx
     app.state.config_store.update(lambda c: setattr(c.app, "scheduler_enabled", False))

backend/tests/test_api_devices.py

@@ -161,6 +161,45 @@ def test_delete_404s_on_an_unknown_device(app_ctx):
     assert client.delete("/api/devices/pbss/nope").status_code == 404
 
 
+def test_a_validation_failure_does_not_echo_the_config_back(app_ctx):
+    """A config-level cross-check raises at loc=(), so pydantic attaches the *whole*
+    validated config as the error's ``input`` — every API token, the secret key, the
+    password hash, the SMTP and bot tokens — and the 422 shipped it to the browser."""
+    client, app = app_ctx
+    client.post(
+        "/api/routes",
+        json={"id": "watch", "kind": "external", "target": "pbs-02", "schedule": {"time": "03:00"}},
+    )
+    # An External route watches a box Joulenap wakes, so making that box unmanaged is
+    # exactly the cross-check that fires.
+    r = client.put("/api/devices/pbss/pbs-02", json={"id": "pbs-02", "host": "192.0.2.21",
+                                                    "datastore": "offsite",
+                                                    "api_token_id": "root@pam!joulenap",
+                                                    "api_token_secret": REDACTED,
+                                                    "managed_power": False})
+    assert r.status_code == 422
+    assert "managed_power" in r.text  # the reason still reaches the user
+    stored = app.state.config_store.config
+    for secret in (
+        stored.app.secret_key,
+        stored.app.auth.password_hash,
+        *[d.api_token_secret for d in stored.pves],
+        *[d.api_token_secret for d in stored.pbss],
+    ):
+        assert secret and secret not in r.text
+
+
+def test_create_rejects_the_redaction_placeholder(app_ctx):
+    """A body copy-pasted from GET /api/devices carries ***REDACTED*** as the token. Storing
+    it verbatim is invisible — GET re-masks it (it is non-empty) — until a connection test
+    fails with a 502 that gives no hint the stored token is the placeholder text itself."""
+    client, app = app_ctx
+    r = client.post("/api/devices/pbss", json={**NEW_PBS, "api_token_secret": REDACTED})
+    assert r.status_code == 422
+    assert "api_token_secret" in r.text
+    assert [d.id for d in app.state.config_store.config.pbss] == ["pbs-01", "pbs-02"]
+
+
 # --- test ---------------------------------------------------------------------
 
 
@@ -228,6 +267,25 @@ def test_power_off_409s_while_a_run_holds_the_box(app_ctx):
     assert "pbs-01" in r.json()["detail"]
 
 
+def test_power_off_409s_while_a_run_holds_the_single_run_lock(app_ctx):
+    """The lease check alone was a check-then-act: a scheduled route could start in the gap,
+    find the box still up (so no wake), and begin vzdump — then the SSH poweroff, which has
+    no idle-wait and no refcount check, cut it off mid-backup. Holding the same lock a run
+    holds for its whole life is what closes that window."""
+    client, app = app_ctx
+    box = FakeBox()
+    service = _inject(app, box)
+    service._lock.acquire()  # stand in for a run in flight; no lease taken yet
+    try:
+        r = client.post("/api/devices/pbss/pbs-01/power", json={"action": "poweroff"})
+    finally:
+        service._lock.release()
+
+    assert r.status_code == 409
+    assert "in progress" in r.json()["detail"]
+    assert box.poweroffs == []  # the SSH command never ran
+
+
 def test_power_409s_on_an_unmanaged_device(app_ctx):
     client, app = app_ctx
     app.state.config_store.update(
@@ -296,7 +354,10 @@ def test_verify_queues_a_verify_run(app_ctx):
 
     assert pbs.verify_started is True
     # An ad-hoc verify checks everything, rather than pacing itself like a Verify route.
-    assert pbs.verify_args["outdated_after"] is None
+    # "Everything" reaches PBS as ignore-verified: 0. The bug was asking for
+    # outdated_after=None, which means "only never-verified" — skipping precisely the old
+    # snapshots the user clicked the button about.
+    assert pbs.verify_args["ignore_verified"] is False
 
 
 def test_maintenance_works_on_an_always_on_box(app_ctx):

backend/tests/test_config.py

@@ -231,6 +231,31 @@ def test_a_new_device_keeps_the_secret_it_was_sent():
     assert out["pbss"][0]["api_token_secret"] == "brand-new"
 
 
+def test_renaming_a_device_id_rejects_its_unresolvable_placeholder():
+    """The Advanced tab shows ``id: pbs-01`` with ``api_token_secret: ***REDACTED***``. Fix a
+    typo in the id and save: nothing matches by id any more, so the placeholder used to
+    resolve to "" — 200 OK, credential gone, nothing on screen to suggest it."""
+    cfg = Config()
+    cfg.pbss = [PbsDevice(id="pbs-01", api_token_secret="first", managed_power=False)]
+    incoming = {"pbss": [{"id": "pbs01", "api_token_secret": cfgmod.REDACTED}]}
+    with pytest.raises(cfgmod.RedactionError, match="api_token_secret"):
+        restore_secrets(incoming, cfg)
+
+
+def test_a_placeholder_with_nothing_stored_is_rejected():
+    # Same guard one level up: restore_secrets_from({}) is what a *create* resolves against.
+    with pytest.raises(cfgmod.RedactionError):
+        cfgmod.restore_secrets_from({"api_token_secret": cfgmod.REDACTED}, {})
+
+
+def test_an_empty_string_still_clears_a_secret():
+    # The escape hatch the rejection points at: "" means clear, and must keep working.
+    cfg = Config()
+    cfg.pbss = [PbsDevice(id="pbs-01", api_token_secret="first", managed_power=False)]
+    out = restore_secrets({"pbss": [{"id": "pbs-01", "api_token_secret": ""}]}, cfg)
+    assert out["pbss"][0]["api_token_secret"] == ""
+
+
 def test_session_defaults():
     s = Config().app.session
     assert s.https_only is False and s.max_age_days == 14

backend/tests/test_config_migrate.py

@@ -7,12 +7,14 @@ quietly disappearing from a real user's config.
 
 from __future__ import annotations
 
+import os
+import stat
 from pathlib import Path
 
 import pytest
 import yaml
 
-from app import config_migrate
+from app import config, config_migrate
 from app.config import REDACTED, Config, load_config, save_config
 
 FIXTURE = Path(__file__).parent / "fixtures" / "config-0.9.yaml"
@@ -21,6 +23,7 @@ BAK = "config.yaml" + config_migrate.BACKUP_SUFFIX
 
 def _write(tmp_path: Path, **overrides) -> Path:
     """Drop the 0.9 fixture into tmp_path, deep-merging any section overrides."""
+    tmp_path.mkdir(parents=True, exist_ok=True)
     raw = yaml.safe_load(FIXTURE.read_text(encoding="utf-8"))
     for section, values in overrides.items():
         for key, value in values.items():
@@ -163,7 +166,36 @@ def test_a_config_that_cannot_be_converted_is_left_alone(tmp_path: Path, monkeyp
     # dropped at load, so the app boots on defaults rather than failing to start.
     assert cfg.routes == []
     assert path.read_text(encoding="utf-8") == original
-    assert not (tmp_path / BAK).exists()
+    # ...but the parachute still opens: the app is now running on an empty config, and the
+    # next save from the Advanced tab would overwrite the only 0.9 config the user has.
+    assert (tmp_path / BAK).read_text(encoding="utf-8") == original
+
+
+def test_a_refused_migration_is_reported_not_just_logged(tmp_path: Path):
+    # An empty config is indistinguishable from a fresh install, so the reason has to
+    # outlive the log line: GET /api/status and the activity log both read this.
+    path = _write(tmp_path, backup={"bwlimit": -1})
+    cfg = load_config(path)
+    assert (cfg.pves, cfg.pbss, cfg.routes) == ([], [], [])
+    assert config.MIGRATION_ERROR is not None
+    assert "bwlimit" in config.MIGRATION_ERROR
+    assert str(path) in config.MIGRATION_ERROR
+
+
+def test_a_later_good_load_clears_the_reported_error(tmp_path: Path):
+    load_config(_write(tmp_path, backup={"bwlimit": -1}))
+    assert config.MIGRATION_ERROR is not None
+    load_config(_write(tmp_path / "fixed", backup={}))
+    assert config.MIGRATION_ERROR is None
+
+
+@pytest.mark.skipif(os.name == "nt", reason="POSIX permission bits")
+def test_the_backup_is_owner_only(tmp_path: Path):
+    # It holds every API token, secret_key, password hash, SMTP and bot token — and is
+    # never overwritten, so a 0644 copy would sit there forever.
+    path = _write(tmp_path)
+    load_config(path)
+    assert stat.S_IMODE((tmp_path / BAK).stat().st_mode) == 0o600
 
 
 def test_an_unwritable_config_still_boots(tmp_path: Path, monkeypatch):

backend/tests/test_lease.py

@@ -6,7 +6,7 @@ import pytest
 from fakes import FakeBox
 
 from app.config import PbsDevice
-from app.jobs.lease import PbsUnreachableError, PowerLease
+from app.jobs.lease import PbsUnreachableError, PowerLease, ReleaseOutcome
 
 
 def make_pbs(**over) -> PbsDevice:
@@ -84,7 +84,7 @@ def test_unmanaged_pbs_is_only_probed():
     pbs = make_pbs(managed_power=False, mac="")
 
     assert lease.acquire(pbs) is True
-    assert lease.release(pbs) is False  # never powered off
+    assert lease.release(pbs) is ReleaseOutcome.UNMANAGED  # never Joulenap's to power
     assert box.wol == []
     assert box.poweroffs == []
 
@@ -108,7 +108,7 @@ def test_last_holder_powers_the_box_off():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is True
+    assert lease.release(pbs) is ReleaseOutcome.POWERED_OFF
     assert box.poweroffs == ["pbs1"]
     assert lease.state("pbs1").holders == 0
 
@@ -120,9 +120,9 @@ def test_release_with_another_holder_leaves_the_box_on():
 
     lease.acquire(pbs)
     lease.acquire(pbs)
-    assert lease.release(pbs) is False
+    assert lease.release(pbs) is ReleaseOutcome.STILL_NEEDED
     assert box.poweroffs == []
-    assert lease.release(pbs) is True
+    assert lease.release(pbs) is ReleaseOutcome.POWERED_OFF
 
 
 def test_a_queued_route_on_the_same_pbs_keeps_it_on():
@@ -131,7 +131,7 @@ def test_a_queued_route_on_the_same_pbs_keeps_it_on():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is False
+    assert lease.release(pbs) is ReleaseOutcome.STILL_NEEDED
     assert box.poweroffs == []
 
 
@@ -141,7 +141,7 @@ def test_power_off_false_leaves_the_box_on():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs, power_off=False) is False
+    assert lease.release(pbs, power_off=False) is ReleaseOutcome.LEFT_ON
     assert box.poweroffs == []
 
 
@@ -151,7 +151,7 @@ def test_a_busy_pbs_is_not_interrupted():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is False
+    assert lease.release(pbs) is ReleaseOutcome.LEFT_ON
     assert box.poweroffs == []
 
 
@@ -162,7 +162,7 @@ def test_a_failing_idle_check_fails_open():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is True
+    assert lease.release(pbs) is ReleaseOutcome.POWERED_OFF
     assert box.poweroffs == ["pbs1"]
 
 
@@ -172,7 +172,7 @@ def test_the_idle_check_is_skipped_when_the_wait_is_zero():
     pbs = make_pbs(poweroff_task_wait=0)
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is True
+    assert lease.release(pbs) is ReleaseOutcome.POWERED_OFF
 
 
 def test_a_failed_power_off_is_swallowed():
@@ -182,7 +182,7 @@ def test_a_failed_power_off_is_swallowed():
     pbs = make_pbs()
 
     lease.acquire(pbs)
-    assert lease.release(pbs) is False
+    assert lease.release(pbs) is ReleaseOutcome.LEFT_ON
     assert box.poweroffs == []
 
 
@@ -190,7 +190,7 @@ def test_releasing_an_unheld_lease_is_ignored():
     box = FakeBox()
     lease = PowerLease(box.deps())
 
-    assert lease.release(make_pbs()) is False
+    assert lease.release(make_pbs()) is ReleaseOutcome.LEFT_ON
     assert lease.state("pbs1").holders == 0
     assert box.poweroffs == []
 

backend/tests/test_notify.py

@@ -187,7 +187,7 @@ def _route(route_id: str = "nightly", name: str = "Nightly", **overrides) -> Rou
     )
 
 
-def _msg(config, run, datastore=None, guests=None, next_at=None, route=None):
+def _msg(config, run, datastore=None, guests=None, next_at=None, route=None, left_on=()):
     """``build_run_message`` with the old positional tail, so these tests stay readable.
 
     The production seam is a single :class:`RunContext`; spelling that out at ~25 call sites
@@ -201,6 +201,7 @@ def _msg(config, run, datastore=None, guests=None, next_at=None, route=None):
             datastore=datastore,
             guests=guests,
             next_at=next_at,
+            left_on=list(left_on),
         )
     )
 
@@ -271,6 +272,7 @@ def test_run_message_field_order():
         ds,
         GuestSummary(total=2, ok=1, failed=["web01"]),
         datetime(2026, 6, 29, 4, 0, tzinfo=UTC),
+        left_on=["pbs-01"],
     )
     assert [line.split(":")[0] for line in body.splitlines()] == [
         "Trigger",
@@ -359,51 +361,23 @@ def _woke() -> RunStep:
     return RunStep(name=StepName.WAIT, status=StepStatus.SUCCESS)
 
 
-def test_run_message_flags_pbs_left_on_when_poweroff_failed():
-    run = _run(RunStatus.SUCCESS)
-    run.steps = [_woke(), RunStep(name=StepName.POWEROFF, status=StepStatus.FAILURE)]
-    _title, body = _msg(Config(), run)
-    assert "left powered on" in body
-
-
-def test_run_message_flags_pbs_left_on_when_poweroff_skipped():
-    run = _run(RunStatus.SUCCESS)
-    run.steps = [_woke(), RunStep(name=StepName.POWEROFF, status=StepStatus.SKIPPED)]
-    _title, body = _msg(Config(), run)
-    assert "left powered on" in body
-
-
-def test_run_message_no_pbs_line_when_poweroff_succeeded():
-    run = _run(RunStatus.SUCCESS)
-    run.steps = [_woke(), RunStep(name=StepName.POWEROFF, status=StepStatus.SUCCESS)]
-    _title, body = _msg(Config(), run)
-    assert "left powered on" not in body
-
-
-def test_run_message_flags_pbs_left_on_when_backup_fails_after_wake():
-    # Failure after the PBS woke: no POWEROFF step at all, box is left on for inspection.
+def test_run_message_flags_the_boxes_the_run_left_awake():
+    # A finished run doesn't guess from its own timeline: JobService hands it the ids of the
+    # boxes still burning power (RunContext.left_on), because only the lease knows whether
+    # "not powered off" meant an always-on box, one another run still holds, or a real
+    # left-on. Which reason applies is decided (and tested) in tests/test_queue.py.
     run = _run(RunStatus.FAILURE, error="vzdump failed")
     run.steps = [_woke(), RunStep(name=StepName.BACKUP, status=StepStatus.FAILURE)]
-    _title, body = _msg(Config(), run)
+    _title, body = _msg(Config(), run, left_on=["pbs-01"])
     assert "left powered on" in body
 
 
-def test_run_message_flags_pbs_left_on_when_aborted_after_wake():
-    # An abort after wake (e.g. free-space preflight) also leaves the box on.
-    run = _run(RunStatus.ABORTED, error="datastore too full")
-    run.steps = [_woke(), RunStep(name=StepName.PRECHECK, status=StepStatus.FAILURE)]
-    _title, body = _msg(Config(), run)
-    assert "left powered on" in body
-
-
-def test_run_message_no_pbs_line_when_wait_timed_out():
-    # Aborted before the PBS came up (WAIT failed): the box never turned on, so no warning.
-    run = _run(RunStatus.ABORTED, error="PBS not reachable")
-    run.steps = [
-        RunStep(name=StepName.WAKE, status=StepStatus.SUCCESS),
-        RunStep(name=StepName.WAIT, status=StepStatus.FAILURE),
-    ]
-    _title, body = _msg(Config(), run)
+def test_run_message_no_pbs_line_when_nothing_was_left_awake():
+    # A SKIPPED power-off is not evidence on its own — an always-on PBS records exactly
+    # this on every successful run, and used to be warned about every night.
+    run = _run(RunStatus.SUCCESS)
+    run.steps = [_woke(), RunStep(name=StepName.POWEROFF, status=StepStatus.SKIPPED)]
+    _title, body = _msg(Config(), run, left_on=[])
     assert "left powered on" not in body
 
 
@@ -501,6 +475,49 @@ def test_interrupted_message_flags_pbs_left_on_when_it_had_woken():
     assert "left powered on" in body
 
 
+def test_interrupted_message_ignores_a_box_joulenap_never_powers():
+    # A crash cannot leave an always-on PBS "burning power" — it was on before and after.
+    cfg = Config.model_validate(
+        {
+            "pbss": [{"id": "pbs-01", "host": "192.0.2.20", "managed_power": False}],
+            "routes": [{"id": "nightly", "kind": "verify", "target": "pbs-01"}],
+        }
+    )
+    run = _run(RunStatus.FAILURE, error="Interrupted")
+    run.route_id = "nightly"
+    run.steps = [
+        RunStep(name=StepName.WAIT, status=StepStatus.SUCCESS),
+        RunStep(name=StepName.BACKUP, status=StepStatus.FAILURE),
+    ]
+    _title, body = build_interrupted_message(cfg, run)
+    assert "left powered on" not in body
+
+
+def test_interrupted_message_pairs_wake_and_power_off_per_device():
+    """A sync route holds two boxes. One powering off must not hide the other staying up —
+    the old rule ORed both steps across the whole run and went silent."""
+    cfg = Config.model_validate(
+        {
+            "pbss": [
+                {"id": "pbs-01", "host": "192.0.2.20", "mac": "00:11:22:33:44:55"},
+                {"id": "pbs-02", "host": "192.0.2.21", "mac": "00:11:22:33:44:66"},
+            ],
+            "routes": [
+                {"id": "off", "kind": "sync", "source_pbs": "pbs-01", "target": "pbs-02"}
+            ],
+        }
+    )
+    run = _run(RunStatus.FAILURE, error="Interrupted")
+    run.route_id = "off"
+    run.steps = [
+        RunStep(name="wait:pbs-01", status=StepStatus.SUCCESS),
+        RunStep(name="wait:pbs-02", status=StepStatus.SUCCESS),
+        RunStep(name="poweroff:pbs-02", status=StepStatus.SUCCESS),
+    ]
+    _title, body = build_interrupted_message(cfg, run)
+    assert "left powered on" in body  # pbs-01 is still up
+
+
 def test_interrupted_message_no_pbs_line_when_it_never_woke():
     # Crashed during WAIT (PBS never came up): no "left on" warning.
     run = _run(RunStatus.FAILURE, error="Interrupted")

backend/tests/test_queue.py

@@ -12,8 +12,10 @@ from sqlalchemy import select
 from app.config import PbsDevice, PveDevice, Route, RouteSource
 from app.core.config_store import ConfigStore
 from app.db import session_scope
-from app.db.models import Run, RunKind, RunStatus, RunTrigger
+from app.db.models import Run, RunKind, RunStatus, RunTrigger, StepName, StepStatus
+from app.jobs.lease import ReleaseOutcome
 from app.jobs.service import AlreadyQueuedError, JobService, QueuedRun
+from app.notify.messages import RunContext
 
 
 def make_service(box: FakeBox | None = None) -> tuple[JobService, FakeBox]:
@@ -284,6 +286,101 @@ def test_an_unreachable_pbs_fails_the_run_without_running_the_job(temp_config, t
         assert "not reachable" in (run.error or "")
 
 
+# --- what the run reports as left powered on ---------------------------------
+#
+# "PBS left powered on" is the notification's energy warning, so it must fire exactly when
+# a box is still awake with nothing left to shut it down. The lease is the only place that
+# can tell the reasons apart, and it hands its verdict to the message through
+# RunContext.left_on (tests/test_notify.py covers the rendering).
+
+
+def notifying_job(status=RunStatus.SUCCESS):
+    """A job that finishes with ``status`` and asks to be notified about it."""
+
+    def job(config, subject, recorder, _deps):
+        recorder.finish(status)
+        return RunContext(config=config, run=recorder.run)
+
+    return job
+
+
+def sent(service: JobService) -> list[RunContext]:
+    seen: list[RunContext] = []
+    service.deps.notify = seen.append
+    return seen
+
+
+def test_an_always_on_pbs_is_never_reported_as_left_on(temp_config, temp_db):
+    # The common false positive: a managed_power: false box records a SKIPPED power-off on
+    # *every* successful run, so the old step-based rule warned about it every single night.
+    service, box = make_service()
+    service._store.config.pbss[0].managed_power = False
+    seen = sent(service)
+
+    enqueue(service, "r1", notifying_job(), trigger=RunTrigger.SCHEDULED)
+    drain(service)
+
+    assert box.poweroffs == []
+    assert seen[0].left_on == []
+
+
+def test_a_failed_run_reports_the_box_it_left_awake(temp_config, temp_db):
+    service, box = make_service()
+    seen = sent(service)
+
+    enqueue(service, "r1", notifying_job(RunStatus.FAILURE), trigger=RunTrigger.SCHEDULED)
+    drain(service)
+
+    assert box.poweroffs == []  # left up for inspection
+    assert seen[0].left_on == ["pbs1"]
+
+
+def test_a_sync_route_reports_only_the_box_that_stayed_up(temp_config, temp_db):
+    # The old false negative: one box powering off hid the other one staying awake, because
+    # the rule ORed the steps across the whole run instead of pairing them per device.
+    service, box = make_service()
+    service._store.config.pbss[1].managed_power = False  # target pbs2 is always on
+    seen = sent(service)
+
+    enqueue(service, "sync", notifying_job(RunStatus.FAILURE), trigger=RunTrigger.SCHEDULED)
+    drain(service)
+
+    assert box.poweroffs == []
+    assert seen[0].left_on == ["pbs1"]  # not pbs2: that one is never ours to power down
+
+
+def test_a_box_a_queued_route_still_needs_is_not_reported_as_left_on(temp_config, temp_db):
+    # It stays awake on purpose, and the run that follows will close it.
+    service, box = make_service()
+    seen = sent(service)
+    gate = Gate()
+
+    enqueue(service, "r1", gate.job, trigger=RunTrigger.SCHEDULED)
+    assert gate.started.wait(timeout=5)
+    enqueue(service, "r2", notifying_job(), trigger=RunTrigger.SCHEDULED)
+    gate.release.set()
+    drain(service)
+
+    assert box.poweroffs == ["pbs1"]  # only r2, the last holder, closed it
+    assert seen[0].left_on == []
+
+
+def test_the_power_off_step_records_why_the_box_stayed_up(temp_config, temp_db):
+    # The timeline says which of the three "not powered off" reasons applied, so the run
+    # detail view doesn't just show a bare SKIPPED.
+    service, _box = make_service()
+    service._store.config.pbss[0].managed_power = False
+
+    enqueue(service, "r1", ok_job, trigger=RunTrigger.SCHEDULED)
+    drain(service)
+
+    with session_scope() as session:
+        run = session.scalars(select(Run)).one()
+        step = next(s for s in run.steps if s.name == StepName.POWEROFF)
+        assert step.status == StepStatus.SKIPPED
+        assert step.detail == ReleaseOutcome.UNMANAGED
+
+
 def test_a_run_on_another_pbs_does_not_hold_the_first(temp_config, temp_db):
     service, box = make_service()
 

backend/tests/test_service.py

@@ -275,6 +275,23 @@ def test_cancel_is_refused_when_nothing_is_running(temp_config, temp_db):
     assert service.cancel(run_id) is False
 
 
+def test_a_finished_run_stops_being_the_cancellable_one(temp_config, temp_db):
+    """``cancel`` takes a run id precisely so a click landing between two runs can't hit the
+    wrong one — but the id was only ever *assigned*, never cleared, so it kept naming the
+    finished run. ``_start`` sets the new id after acquiring the lock and doing the DB
+    insert, so in that window ``is_running`` is already True while the stale id still
+    matches: the stop would be swallowed, or set the cancel flag on a run the user never
+    targeted, which then aborts on its first poll."""
+    service, _box = _service()
+    service.run_route("nightly")
+    _drain(service)
+    with session_scope() as session:
+        run_id = session.scalars(select(Run)).one().id
+
+    assert service._current_run_id is None
+    assert service.cancel(run_id) is False
+
+
 def test_a_stale_cancel_does_not_kill_the_next_run(temp_config, temp_db):
     # Cancel arrives moments before the run ends on its own; the flag must not leak into
     # the run that starts next.