diff --git a/README.md b/README.md
index 550cd21..671effc 100644
--- a/README.md
+++ b/README.md
@@ -270,6 +270,12 @@ If reconnects become persistent (more than a handful per minute), something's ac
 
 Deregistering a device in SmartThings causes a reset of its network settings as soon as it accesses Samsung's servers, dropping it off Wi-Fi until it's re-onboarded through the SmartThings app. As such, consider keeping devices registered even if egress-blocked, to avoid them resetting upon brief internet access.
 
+### Restarting while an appliance is powered off
+
+If Home Assistant restarts while an appliance is unplugged or switched off at the wall, its device and entities still load — restored from the last successful discovery, showing `unavailable` until the appliance answers again. Automations and dashboards keep referring to entities that exist, and the integration retries in the background, so the device comes back on its own within a poll cycle of being powered on. Entities read `unavailable` rather than their last known values on purpose: the integration can't verify what a disconnected appliance is doing, and recorded history is kept by the recorder either way.
+
+This only applies to an appliance the integration has reached at least once. A brand-new device that has never answered has nothing to restore from, so setting it up still requires it to be reachable.
+
 ### Multi-subdevice ("2-in-1") air conditioner systems
 
 Some Samsung installs run more than one indoor subdevice off a single outdoor unit, all reachable over the *one* IP/DTLS session your config entry connects to (a floor-standing + wall-mounted 2-in-1 is a common shape). The integration discovers any sibling subdevices automatically, once, right after the first successful poll — there's nothing to configure. Each discovered subdevice gets its own HA device (linked to the main one via "via device") and its own `climate` card, so it lands in its own room in the dashboard instead of being invisible or mixed into the master's state.
diff --git a/custom_components/localthings/__init__.py b/custom_components/localthings/__init__.py
index ce47f79..ed5417f 100644
--- a/custom_components/localthings/__init__.py
+++ b/custom_components/localthings/__init__.py
@@ -13,12 +13,13 @@ from homeassistant.const import (
 )
 from homeassistant.const import EVENT_HOMEASSISTANT_STOP
 from homeassistant.core import Event, HomeAssistant, callback
+from homeassistant.exceptions import ConfigEntryNotReady
 from homeassistant.helpers import device_registry as dr
 from homeassistant.helpers import entity_registry as er
 from homeassistant.helpers.typing import ConfigType
 
 from .const import CONF_DEVICE_TYPE, CONF_HOST, CONF_PORT, CONF_SERIAL, DOMAIN, PLATFORMS
-from .coordinator import LocalThingsCoordinator
+from .coordinator import LocalThingsCoordinator, snapshot_store
 from .registry.identity import resolve_serial
 from .services import async_setup_services
 
@@ -259,13 +260,30 @@ async def async_migrate_entry(hass: HomeAssistant, entry: ConfigEntry) -> bool:
 async def async_setup_entry(hass: HomeAssistant, entry: ConfigEntry) -> bool:
     hass.data.setdefault(DOMAIN, {})
     coordinator = LocalThingsCoordinator(hass, entry)
+
+    # Before the first refresh, so the coordinator keeps rescheduling even
+    # when that refresh fails and leaves nothing subscribed: the base class
+    # only re-arms its timer while it has listeners, and an offline load can
+    # legitimately have zero live entities (every one of them disabled, say).
+    # Without this the entry loads once and never polls again (issue #295).
+    entry.async_on_unload(coordinator.async_add_listener(lambda: None))
+
     try:
         await coordinator.async_config_entry_first_refresh()
-    except Exception as err:
-        _LOGGER.warning(
-            "Initial connection to device failed (%s); starting offline and retrying in background",
-            err,
-        )
+    except ConfigEntryNotReady:
+        # An entry that has polled successfully before comes up on its last
+        # known entity set and keeps retrying on the normal poll interval,
+        # rather than sitting in setup-retry with a device that reads as
+        # broken and entities that exist only as registry rows (issue #295).
+        #
+        # An entry that has never reached the device has no snapshot, so
+        # there is nothing to show and no device metadata to name it with --
+        # that case still fails, which is also what keeps the door open for
+        # setup flows that need to interact with the device (issue #168).
+        if not await coordinator.async_rehydrate():
+            await coordinator.async_close()
+            raise
+
     hass.data[DOMAIN][entry.entry_id] = coordinator
 
     # Send the DTLS close_notify on Core shutdown, not just on unload (issue
@@ -322,6 +340,16 @@ async def async_remove_config_entry_device(
     return not (device.identifiers & live)
 
 
+async def async_remove_entry(hass: HomeAssistant, entry: ConfigEntry) -> None:
+    """Delete the discovery snapshot this entry accumulated (issue #295).
+
+    Nothing else would: the store is keyed on entry_id, so re-adding the same
+    appliance mints a new one and the old file would linger in .storage
+    forever.
+    """
+    await snapshot_store(hass, entry).async_remove()
+
+
 async def async_unload_entry(hass: HomeAssistant, entry: ConfigEntry) -> bool:
     unloaded = await hass.config_entries.async_unload_platforms(entry, PLATFORMS)
     if unloaded:
diff --git a/custom_components/localthings/coordinator.py b/custom_components/localthings/coordinator.py
index c41785c..e02124a 100644
--- a/custom_components/localthings/coordinator.py
+++ b/custom_components/localthings/coordinator.py
@@ -9,6 +9,7 @@ import logging
 import threading
 import time
 import zlib
+from dataclasses import asdict
 from datetime import timedelta
 from typing import Any, cast
 
@@ -18,6 +19,7 @@ from homeassistant.core import HomeAssistant, callback
 from homeassistant.exceptions import HomeAssistantError, ServiceValidationError
 from homeassistant.helpers import issue_registry as ir
 from homeassistant.helpers.device_registry import DeviceInfo
+from homeassistant.helpers.storage import Store
 from homeassistant.helpers.update_coordinator import DataUpdateCoordinator, UpdateFailed
 from smartthings_local.ocf.state_cache import StateCache
 from smartthings_local.protocol.dtls_session import DtlsCoapSession
@@ -47,7 +49,7 @@ from .const import (
 from .learned import LEARNABLE, LearnedModes, persist
 from .observe import GRACE_PERIOD_S, MODE_OBSERVE, MODE_POLL, ObserveManager
 from .registry import CAPABILITIES
-from .registry.adapter import flatten
+from .registry.adapter import _key, flatten
 from .registry.batch import parse_device0_batch
 from .registry.by_type import resolve as resolve_registry
 from .registry.capabilities.common import (
@@ -83,6 +85,21 @@ _LOGGER = logging.getLogger(__name__)
 
 _SEED_PATH = ["device", "0"]
 
+# Discovery snapshot (issue #295): exactly what the last successful first
+# cycle fed _run_discovery, so a restart can register the same entities
+# while the appliance is unreachable. Kept in .storage rather than on the
+# config entry -- it's device state, not configuration, and runs to tens of
+# kilobytes.
+_SNAPSHOT_VERSION = 1
+_SNAPSHOT_SAVE_DELAY_S = 10.0
+
+
+def snapshot_store(hass: HomeAssistant, entry: ConfigEntry) -> Store[dict[str, Any]]:
+    """This entry's discovery-snapshot store. A free function so
+    `async_remove_entry` can delete the file without standing up a whole
+    coordinator to reach it."""
+    return Store(hass, _SNAPSHOT_VERSION, f"{DOMAIN}.{entry.entry_id}.discovery")
+
 
 class _NoOpDescriptor:
     """No-op: StateCache requires an on_observation hook; this integration
@@ -248,6 +265,11 @@ class LocalThingsCoordinator(DataUpdateCoordinator[dict[str, Any]]):
         self._identity: DeviceIdentity | None = None
         self._discovered = False
         self.bound = []
+        self._snapshot_store = snapshot_store(hass, entry)
+        # Both set only when this entry loaded from a snapshot instead of a
+        # live poll -- see async_rehydrate.
+        self._rehydrate_resources: dict[str, dict] | None = None
+        self._rehydrated_keys: frozenset[tuple[str, str]] | None = None
         # Sibling indoor subdevices on this connection (issue #177); set
         # once at first discovery, narrowed to the ones with live state (see
         # subdevices.discover_partitioned). Never includes MAIN itself.
@@ -405,6 +427,35 @@ class LocalThingsCoordinator(DataUpdateCoordinator[dict[str, Any]]):
         self._canonical_cache[view_key] = view
         return view
 
+    @property
+    def rehydrated(self) -> bool:
+        """True while this entry's entities came from a snapshot rather than
+        a live poll (issue #295)."""
+        return self._rehydrate_resources is not None
+
+    @property
+    def discovery_resources(self) -> dict[str, dict]:
+        """What entity._is_included should judge an entity's existence
+        against: the rehydration snapshot on an offline load, the live cache
+        otherwise.
+
+        Deliberately separate from `last_resources`, which stays empty until
+        the device answers -- that emptiness is what keeps a rehydrated
+        entity `unavailable` instead of rendering a snapshot's stale value.
+        Only read while platforms are being forwarded; nothing consults it
+        once the entities exist.
+        """
+        if self._rehydrate_resources is None:
+            return self.last_resources
+        return self._rehydrate_resources
+
+    def discovery_canonical(self, subdevice: Subdevice) -> dict[str, dict]:
+        """`discovery_resources` in `subdevice`'s canonical view -- the
+        exists_fn counterpart to canonical_resources."""
+        if self._rehydrate_resources is None:
+            return self.canonical_resources(subdevice)
+        return canonical_view(subdevice, self._rehydrate_resources, self.subdevices)
+
     # ------------------------------------------------------------------
     # Learned modes (issue #327)
     # ------------------------------------------------------------------
@@ -999,6 +1050,128 @@ class LocalThingsCoordinator(DataUpdateCoordinator[dict[str, Any]]):
             self._entry, data={**self._entry.data, **identity}
         )
 
+    # ------------------------------------------------------------------
+    # Discovery snapshot (issue #295)
+    # ------------------------------------------------------------------
+
+    def _bound_keys(self) -> frozenset[tuple[str, str]]:
+        """This entity set's identity: one (subdevice, state key) pair per
+        bound entity. `_key` is the unique_id suffix, so two discoveries that
+        agree here would register byte-identical entities."""
+        return frozenset((b.subdevice.key, _key(b)) for b in self.bound)
+
+    @callback
+    def _save_snapshot(self, resources: dict[str, dict], candidates: list[Subdevice]) -> None:
+        """Record what this first cycle handed `_run_discovery`, so the next
+        restart can replay it while the appliance is unreachable.
+
+        `candidates` is the pre-narrowing subdevice list (issue #177):
+        `discover_partitioned` takes candidates and returns the live ones, so
+        replaying against the narrowed list would rediscover nothing for a
+        composite appliance's siblings.
+        """
+        ident = self._identity
+        payload = {
+            "resources": dict(resources),
+            "subdevice_candidates": [asdict(su) for su in candidates],
+            "identity": asdict(ident) if ident is not None else None,
+        }
+        self._snapshot_store.async_delay_save(lambda: payload, _SNAPSHOT_SAVE_DELAY_S)
+
+    async def async_rehydrate(self) -> bool:
+        """Register the last known entity set without reaching the device.
+
+        Replays the stored snapshot through `_run_discovery`, which is what
+        makes this faithful: same code path, same registry resolution, so an
+        offline load produces the entity set the device last actually
+        reported rather than a guess reconstructed from a parallel format.
+
+        Returns False when there's nothing stored (an entry that has never
+        polled successfully) or the replay produced nothing usable -- the
+        caller raises ConfigEntryNotReady in that case, exactly as before.
+        """
+        try:
+            stored = await self._snapshot_store.async_load()
+        except Exception as e:  # corrupt or unreadable store
+            self._log.warning("could not read discovery snapshot: %s", e)
+            return False
+        if not stored or not stored.get("resources"):
+            return False
+
+        ident = stored.get("identity")
+        if ident is not None:
+            self._identity = DeviceIdentity(
+                manufacturer=ident.get("manufacturer") or "",
+                model=ident.get("model") or "",
+                name=ident.get("name") or "",
+                serial=ident.get("serial"),
+                device_types=tuple(ident.get("device_types") or ()),
+                raw=ident.get("raw") or {},
+            )
+        # JSON gives lists back where Subdevice declares tuples, and it's a
+        # frozen (hashable) dataclass used as a dict key in flatten().
+        self.subdevices = [
+            Subdevice(
+                kind=su["kind"],
+                key=su["key"],
+                seed_path=tuple(su.get("seed_path") or ()),
+                flat_hrefs=tuple(su.get("flat_hrefs") or ()),
+            )
+            for su in stored.get("subdevice_candidates") or ()
+        ]
+
+        resources = stored["resources"]
+        try:
+            self._run_discovery(resources)
+        except Exception as e:
+            # A snapshot written by an older release can outlive the registry
+            # shape it was discovered against. Fall back to the pre-#295
+            # behavior rather than failing the entry outright.
+            self._log.warning("discovery snapshot could not be replayed: %s", e, exc_info=True)
+            self.bound = []
+            return False
+
+        # _run_discovery sets _discovered; put it back. The snapshot only
+        # supplied an entity set to register -- the first live poll must
+        # still enumerate subdevices and rediscover for real.
+        self._discovered = False
+        self._rehydrate_resources = resources
+        self._rehydrated_keys = self._bound_keys()
+        if not self.bound:
+            return False
+        self._log.info(
+            "device unreachable; restored %d entities from the last discovery "
+            "snapshot and will keep retrying every %ds",
+            len(self.bound),
+            SUMMARY_INTERVAL_S,
+        )
+        return True
+
+    @callback
+    def _reconcile_rehydrated(self) -> None:
+        """Reload the entry when a live discovery disagrees with the snapshot
+        this load registered from.
+
+        Platforms enumerate `bound` exactly once, at forward time, so a
+        firmware update, a newly-answering sibling subdevice or a different
+        appliance at the same IP can't be picked up in place -- the entry has
+        to come back up against the live set.
+        """
+        if self._rehydrated_keys is None:
+            return
+        stale = self._rehydrated_keys
+        self._rehydrated_keys = None
+        live = self._bound_keys()
+        if live == stale:
+            return
+        self._log.info(
+            "live discovery differs from the snapshot this entry loaded from "
+            "(%d entities gone, %d new); reloading",
+            len(stale - live),
+            len(live - stale),
+        )
+        self.hass.config_entries.async_schedule_reload(self._entry.entry_id)
+
     def _run_discovery(self, resources: dict[str, dict]) -> None:
         # Diagnostics only -- names the firmware generation (e.g. '7.0 Air
         # conditioner' is Tizen Lite); doesn't route, since every device
@@ -1422,7 +1595,10 @@ class LocalThingsCoordinator(DataUpdateCoordinator[dict[str, Any]]):
             # StateCache has no eviction, so the only way to keep them out
             # is to not put them in. Safe to reorder: _run_discovery reads
             # the passed dict, never the cache.
+            candidates = list(self.subdevices)
             self._run_discovery(resources)
+            self._save_snapshot(resources, candidates)
+            self._reconcile_rehydrated()
             resources = self._live_subdevice_resources(resources)
         sweep_mismatch = False
         if self._observe.mode == MODE_OBSERVE:
diff --git a/custom_components/localthings/entity.py b/custom_components/localthings/entity.py
index 770390c..a6fc65c 100644
--- a/custom_components/localthings/entity.py
+++ b/custom_components/localthings/entity.py
@@ -35,12 +35,16 @@ def _is_included(bound: BoundEntity, coordinator: LocalThingsCoordinator) -> boo
     `bound`'s own subdevice's canonical view instead of the raw snapshot,
     same rule as everywhere else a whole-resources-dict scan happens --
     this is a free function, so it can't use self._resources.
+
+    Reads `discovery_resources`, not `last_resources`: on an offline load
+    (issue #295) the live cache is still empty, and judging existence
+    against it would filter every rehydrated entity away.
     """
-    rep = coordinator.last_resources.get(bound.href)
+    rep = coordinator.discovery_resources.get(bound.href)
     if rep is None:
         return False
     if bound.desc.exists_fn is not None:
-        return bound.desc.exists_fn(rep, coordinator.canonical_resources(bound.subdevice))
+        return bound.desc.exists_fn(rep, coordinator.discovery_canonical(bound.subdevice))
     if bound.desc.field:
         if not rep or is_stub_rep(rep):
             return True
diff --git a/docs/investigations/offline-setup.md b/docs/investigations/offline-setup.md
index 2b9c64c..70c334c 100644
--- a/docs/investigations/offline-setup.md
+++ b/docs/investigations/offline-setup.md
@@ -1,14 +1,19 @@
 # Loading a config entry while the appliance is offline
 
-Issue #295 asks for sub-30s recovery when a powered-off appliance comes back,
-instead of HA's `ConfigEntryNotReady` backoff (30s → … → 15 min). PR #303
-tries to get there by catching the first-refresh failure in
-`async_setup_entry` and loading the entry anyway.
+Issue #295 asks for faster recovery when a powered-off appliance comes back,
+instead of waiting out HA's `ConfigEntryNotReady` backoff. PR #303 tried to
+get there by catching the first-refresh failure in `async_setup_entry` and
+loading the entry anyway.
 
 That doesn't work here, and the reason is worth writing down: this
 integration has no static entity list. Every entity comes from discovery,
 and discovery only happens inside a successful poll.
 
+(The issue's "up to 15 minutes" is out of date, incidentally. Current HA
+retries on `2 ** min(tries, 4) * 5` seconds — capped at 80s, not 900. The
+backoff was never the worst part; a device card reading "Retrying setup" with
+no entities behind it is.)
+
 ## What PR #303 produces today
 
 Measured on the PR's branch — set up with `_poll_once` raising, then advance
@@ -55,51 +60,63 @@ device description* to build entities from — ESPHome keeps its entity list in
 `.storage`, Shelly caches device info. The pattern is portable; the mechanism
 underneath it is the part PR #303 is missing.
 
-## What a working version needs
+## How the implemented version works
 
-Three separable pieces, plus a gating rule.
+Three pieces, plus a gating rule.
 
 ### 1. A persisted discovery snapshot
 
-Write the discovery result back onto the config entry after each successful
-`_run_discovery`. There is already precedent for exactly this shape —
-`_persist_identity` (`coordinator.py:973`), `CONF_LEARNED_MODES`,
-`CONF_CLOUD_COURSES`.
+After each successful first cycle, `_save_snapshot` banks exactly the
+`resources` dict that cycle handed `_run_discovery`, along with the
+pre-narrowing subdevice candidate list and the `DeviceIdentity` read from
+`/oic/*`.
 
-`BoundEntity` holds live `Capability`/`SamsungEntityDescription` objects, so
-it isn't directly serializable, but it is re-derivable: persist
-`device_type_name`, the materialized `Subdevice` list (plain str/tuple
-fields), and per entity the `(subdevice key, capability, desc.key, instance,
-key_override, instance_name)` tuple. Rehydrate by re-resolving through
-`resolve_registry` + `CAPABILITIES`.
+Storing the poll input rather than a rendered entity list is the decision
+that keeps this honest. `BoundEntity` holds live
+`Capability`/`SamsungEntityDescription` objects and isn't serializable, so
+the alternative was a parallel format plus a re-resolution path — a second
+implementation of discovery that could drift from the real one. Replaying the
+input through `_run_discovery` means the same code, the same registry
+resolution, and no second source of truth.
 
-Persist the *post-`_is_included`* set, so rehydration doesn't need reps, and
-flag the coordinator as rehydrated so gate 3 above is skipped for that pass.
+Three things ride along because `_run_discovery` reads them off `self`
+rather than out of `resources`, and getting them wrong would silently resolve
+a *different* registry offline than online — which reconciliation below would
+then see as a real change and reload on every restart:
+
+- `_identity.device_types` routes `resolve_registry`.
+- `self.subdevices` is the candidate list `discover_partitioned` narrows;
+  replaying against the already-narrowed list finds no siblings at all.
+- `_identity.manufacturer`/`model` feed `device_info`.
+
+It lives in `.storage` (`Store`, keyed on entry_id) rather than on the config
+entry: it's device state, not configuration, and runs to tens of kilobytes.
+`async_remove_entry` deletes it with the entry.
 
 ### 2. Reconcile on reconnect
 
-The snapshot is a guess about a device we haven't talked to yet. When the
-first successful poll lands, `_run_discovery` computes the real set; if it
-differs from what was rehydrated, the entry has to reload, because gate 2
-means platforms can't add the difference in place. `async_schedule_reload`
-covers it (available well below the 2025.1.0 floor in `hacs.json`).
+The snapshot is a claim about a device we haven't talked to yet. When the
+first live poll lands, `_reconcile_rehydrated` compares the live entity set
+against the rehydrated one — as `(subdevice key, _key(bound))` pairs, which
+is the unique_id identity — and calls `async_schedule_reload` if they differ.
 
-This is the piece that makes the whole thing safe against a firmware update,
-a newly-appearing subdevice, or a different appliance at the same IP. PR #303
-has no equivalent.
+Gate 2 above is why this has to be a reload rather than an in-place fixup.
+It's what makes the feature safe against a firmware update, a sibling
+subdevice that starts answering, or a different appliance at the same IP.
 
 ### 3. Keep polling with no listeners
 
-Independent of the above, and worth doing on its own merits: hold one
-refresh alive for the entry's lifetime so scheduling never depends on entity
-count.
+`async_setup_entry` holds one listener for the entry's lifetime:
 
 ```python
 entry.async_on_unload(coordinator.async_add_listener(lambda: None))
 ```
 
-That alone fixes the measured "never polls again" bug, and covers the case
-where every rehydrated entity happens to be registry-disabled.
+Registered *before* the first refresh, so scheduling survives a refresh that
+fails. This alone fixes the measured "never polls again" bug, and covers a
+rehydrated set whose entities are all registry-disabled. Removing the last
+listener unschedules the timer, and HA runs `async_on_unload` callbacks when
+setup raises, so the setup-retry path doesn't leak a polling coordinator.
 
 ### Gating rule: only load offline when there's a snapshot
 
@@ -110,6 +127,10 @@ without one HA's backoff is still the right behavior. It also leaves room for
 the #168-style flows that need to interact with the device during setup: a
 device that never completed setup still blocks.
 
+It also means `async_remove_config_entry_device` is no longer reachable with
+an empty `coordinator.subdevices`, so an offline load can't offer to delete a
+real-but-unreachable subdevice.
+
 ## What this still won't do
 
 Entities will be present and `unavailable` — not showing their last values.
@@ -124,26 +145,22 @@ Worth being explicit about, because it is the gap between what PR #303
 promises in the thread ("load their previously recorded states") and what any
 correct version can deliver.
 
-## The cheaper alternative
-
-If the goal is only issue #295's — fast recovery — pieces 1 and 2 are
-unnecessary. Keep `ConfigEntryNotReady`, and add a retry that probes on a
-fixed short interval and calls `async_schedule_reload` on the first success.
-No persistence, no reconcile, no new failure modes. The device tile still
-reads "Retrying setup" while the appliance is off, which is true, and
-recovery drops from up-to-15-min to one probe interval.
+## Rejected: zeroconf
 
 The issue's other suggestion — wire zeroconf so the device's own boot
-announcement triggers the retry — is a non-starter as things stand: there is
-no `zeroconf` or `dhcp` key in `manifest.json` and the config flow is
-user-driven only, so HA has no discovery signal for this integration to hang
-a retry on. It would first need a confirmed mDNS service on the appliance.
+announcement triggers a retry, which is the genuinely idiomatic HA answer —
+is a non-starter as things stand: there is no `zeroconf` or `dhcp` key in
+`manifest.json` and the config flow is user-driven only, so HA has no
+discovery signal for this integration to hang a retry on. It would first need
+a confirmed mDNS service on the appliance. Worth revisiting if one turns up;
+it would make recovery near-instant instead of within one poll interval.
 
-## Recommendation
+## Rejected: the cheap version
 
-The cheap alternative solves the filed issue. The three-piece version solves
-the goal stated later in the PR thread, at the cost of a new persisted
-schema, a reconcile path, and a reload-on-mismatch — and it still leaves
-entities unavailable, which is the part that was actually being asked for.
-Do the cheap one first; treat offline entity materialization as a separate
-change with its own issue.
+Keeping `ConfigEntryNotReady` and adding a probe that calls
+`async_schedule_reload` on first success would have fixed the recovery *time*
+in about twenty lines, with no persistence and no reconcile. It was rejected
+because it leaves the device reading as broken for as long as the appliance
+is off, which is the half of issue #295 that actually bites — an appliance
+switched off at the wall is offline for days, not seconds, and a whole
+integration that looks failed for that entire window is the complaint.
diff --git a/tests/localthings/test_coordinator.py b/tests/localthings/test_coordinator.py
index 82cd159..2429e3c 100644
--- a/tests/localthings/test_coordinator.py
+++ b/tests/localthings/test_coordinator.py
@@ -70,8 +70,13 @@ async def test_summary_interval(hass: HomeAssistant, mock_entry, mock_coordinato
     assert coordinator.update_interval == timedelta(seconds=SUMMARY_INTERVAL_S)
 
 
-async def test_offline_initial_setup_succeeds(hass: HomeAssistant, mock_entry) -> None:
-    """ConfigEntry setup succeeds in offline mode when initial poll fails."""
+async def test_update_failed_on_persistent_poll_error(hass: HomeAssistant, mock_entry) -> None:
+    """ConfigEntryNotReady raised when poll fails even after reconnect.
+
+    An entry with no stored discovery snapshot has never reached this device,
+    so there is nothing to load offline from (issue #295) -- it stays on HA's
+    backoff rather than loading empty.
+    """
 
     with (
         patch("custom_components.localthings.coordinator.LocalThingsCoordinator._connect_session"),
@@ -84,10 +89,8 @@ async def test_offline_initial_setup_succeeds(hass: HomeAssistant, mock_entry) -
         result = await hass.config_entries.async_setup(mock_entry.entry_id)
         await hass.async_block_till_done()
 
-    # Setup succeeds — entry should be in hass.data and state is LOADED
-    assert result
-    assert mock_entry.entry_id in hass.data.get(DOMAIN, {})
-    assert mock_entry.state == ConfigEntryState.LOADED
+    # Setup fails — entry should not be in hass.data
+    assert not result or mock_entry.entry_id not in hass.data.get(DOMAIN, {})
 
 
 # ---------------------------------------------------------------------------
@@ -1823,11 +1826,14 @@ async def test_first_refresh_timeout_recovers_via_reconnect(
     assert coordinator._consecutive_poll_timeouts == 0
 
 
-async def test_first_refresh_persistent_timeout_succeeds_offline(
+async def test_first_refresh_persistent_timeout_fails_setup(
     hass: HomeAssistant, mock_entry
 ) -> None:
-    """When initial refresh times out, setup still succeeds in offline mode
-    and retries in background."""
+    """With no snapshot to load from, a reconnect that times out too must
+    fail the first refresh so HA retries on its backoff -- not load an
+    entity-less entry. The session it left open is closed on the way out
+    (`_poll_once` keeps it up on a `TimeoutError`, and the source port is
+    fixed per device)."""
     with (
         patch("custom_components.localthings.coordinator.LocalThingsCoordinator._connect_session"),
         patch(
@@ -1835,11 +1841,21 @@ async def test_first_refresh_persistent_timeout_succeeds_offline(
             side_effect=_HANDSHAKE_TIMEOUT,
         ),
         patch("custom_components.localthings.coordinator.LocalThingsCoordinator._close_session"),
+        # Asserted on `async_close` rather than `_close_session`: the
+        # reconnect path calls the latter on its own, so it is already true
+        # whether or not setup cleans up after itself.
+        patch.object(
+            LocalThingsCoordinator,
+            "async_close",
+            autospec=True,
+        ) as mock_async_close,
     ):
         await hass.config_entries.async_setup(mock_entry.entry_id)
         await hass.async_block_till_done()
 
-    assert mock_entry.state is ConfigEntryState.LOADED
+    assert mock_entry.state is ConfigEntryState.SETUP_RETRY
+    assert not hass.states.async_all()
+    mock_async_close.assert_awaited_once()
 
 
 async def test_post_discovery_timeout_is_still_deferred(
diff --git a/tests/localthings/test_offline_setup.py b/tests/localthings/test_offline_setup.py
new file mode 100644
index 0000000..37a659e
--- /dev/null
+++ b/tests/localthings/test_offline_setup.py
@@ -0,0 +1,366 @@
+"""Loading a config entry while the appliance is unreachable (issue #295).
+
+The device's entity set only exists as the output of a live poll, so coming
+up offline means replaying the last successful discovery from a stored
+snapshot. These tests pin the four things that makes load-bearing: the
+snapshot gets written, it produces the same entity set offline, the
+coordinator keeps polling until the device answers, and a live discovery that
+disagrees with the snapshot reloads the entry rather than silently keeping a
+stale set.
+"""
+
+from __future__ import annotations
+
+from contextlib import contextmanager
+from datetime import timedelta
+from unittest.mock import patch
+
+import pytest
+from homeassistant.config_entries import ConfigEntryState
+from homeassistant.core import HomeAssistant
+from homeassistant.util import dt as dt_util
+from pytest_homeassistant_custom_component.common import async_fire_time_changed
+
+from custom_components.localthings.const import DOMAIN, SUMMARY_INTERVAL_S
+from custom_components.localthings.coordinator import (
+    _SNAPSHOT_SAVE_DELAY_S,
+    LocalThingsCoordinator,
+)
+from custom_components.localthings.registry.identity import DeviceIdentity
+
+from .conftest import _load_fridge_resources as _load_fridge
+
+_COORD = "custom_components.localthings.coordinator.LocalThingsCoordinator"
+
+
+@contextmanager
+def _reachable(resources: dict, identity: DeviceIdentity | None = None):
+    """A device that answers, optionally with an /oic/* identity -- which
+    `_connect_session` is what normally reads, so a test that patches it out
+    otherwise leaves `_identity` None."""
+
+    def _connect(self) -> None:
+        self._identity = identity
+
+    with (
+        patch(f"{_COORD}._connect_session", _connect),
+        patch(f"{_COORD}._poll_once", return_value=resources),
+        patch(f"{_COORD}._close_session"),
+    ):
+        yield
+
+
+@contextmanager
+def _unreachable():
+    with (
+        patch(f"{_COORD}._connect_session"),
+        patch(f"{_COORD}._poll_once", side_effect=OSError("device offline")),
+        patch(f"{_COORD}._close_session"),
+    ):
+        yield
+
+
+def _store_key(entry) -> str:
+    return f"{DOMAIN}.{entry.entry_id}.discovery"
+
+
+async def _flush_snapshot(hass: HomeAssistant) -> None:
+    """Run out `async_delay_save`'s timer without reaching the next poll."""
+    async_fire_time_changed(hass, dt_util.utcnow() + timedelta(seconds=_SNAPSHOT_SAVE_DELAY_S + 1))
+    await hass.async_block_till_done()
+
+
+async def _tick(hass: HomeAssistant) -> None:
+    """Advance past one summary interval so the coordinator polls again.
+
+    `wait_background_tasks` is load-bearing: DataUpdateCoordinator runs its
+    interval refresh as a background task, which a plain block_till_done
+    doesn't await -- the poll would still be in flight at the assertion.
+    """
+    async_fire_time_changed(hass, dt_util.utcnow() + timedelta(seconds=SUMMARY_INTERVAL_S + 1))
+    await hass.async_block_till_done(wait_background_tasks=True)
+
+
+async def _setup_online_then_unload(hass: HomeAssistant, entry, resources: dict) -> set[str]:
+    """Bring the entry up against a live device, bank the snapshot, and take
+    it back down. Returns the entity_ids that run produced."""
+    with _reachable(resources):
+        await hass.config_entries.async_setup(entry.entry_id)
+        await hass.async_block_till_done()
+        entity_ids = {s.entity_id for s in hass.states.async_all()}
+        await _flush_snapshot(hass)
+        await hass.config_entries.async_unload(entry.entry_id)
+        await hass.async_block_till_done()
+    return entity_ids
+
+
+# ---------------------------------------------------------------------------
+# Writing the snapshot
+# ---------------------------------------------------------------------------
+
+
+async def test_snapshot_written_after_first_discovery(
+    hass: HomeAssistant, mock_entry, mock_coordinator_session, hass_storage
+) -> None:
+    """A successful first cycle banks what it handed _run_discovery."""
+    await hass.config_entries.async_setup(mock_entry.entry_id)
+    await hass.async_block_till_done()
+    await _flush_snapshot(hass)
+
+    stored = hass_storage[_store_key(mock_entry)]["data"]
+    assert stored["resources"]
+    assert "/information/vs/0" in stored["resources"]
+    assert "subdevice_candidates" in stored
+
+
+async def test_snapshot_not_written_when_device_never_answers(
+    hass: HomeAssistant, mock_entry, hass_storage
+) -> None:
+    """Nothing to bank, so nothing is -- this is what keeps the no-snapshot
+    gate meaningful on a brand-new entry."""
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+        # Asserted before the flush: running the save timer out also runs out
+        # HA's own setup-retry timer, which puts the entry back into
+        # SETUP_IN_PROGRESS.
+        assert mock_entry.state is ConfigEntryState.SETUP_RETRY
+        await _flush_snapshot(hass)
+
+    assert _store_key(mock_entry) not in hass_storage
+
+
+async def test_snapshot_removed_when_entry_removed(
+    hass: HomeAssistant, mock_entry, mock_coordinator_session, hass_storage
+) -> None:
+    """The store is keyed on entry_id, so re-adding the appliance mints a new
+    one -- the old file has to go with the entry that wrote it."""
+    await hass.config_entries.async_setup(mock_entry.entry_id)
+    await hass.async_block_till_done()
+    await _flush_snapshot(hass)
+    assert _store_key(mock_entry) in hass_storage
+
+    await hass.config_entries.async_remove(mock_entry.entry_id)
+    await hass.async_block_till_done()
+
+    assert hass_storage.get(_store_key(mock_entry), {}).get("data") is None
+
+
+# ---------------------------------------------------------------------------
+# Loading from it
+# ---------------------------------------------------------------------------
+
+
+async def test_offline_load_restores_the_same_entity_set(
+    hass: HomeAssistant, mock_entry, hass_storage
+) -> None:
+    """The whole point: a restart with the appliance powered off comes up on
+    the entity set the device last actually reported."""
+    resources = _load_fridge()
+    online_ids = await _setup_online_then_unload(hass, mock_entry, resources)
+    assert online_ids  # guard: the online run must actually produce entities
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    assert mock_entry.state is ConfigEntryState.LOADED
+    assert {s.entity_id for s in hass.states.async_all()} == online_ids
+
+
+async def test_offline_entities_are_unavailable_not_stale(hass: HomeAssistant, mock_entry) -> None:
+    """Restored entities must not render the snapshot's values -- the
+    appliance is unreachable, so `unavailable` is the honest state and the
+    live cache stays empty to enforce it."""
+    resources = _load_fridge()
+    await _setup_online_then_unload(hass, mock_entry, resources)
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    states = hass.states.async_all()
+    assert states
+    assert all(s.state == "unavailable" for s in states)
+
+    coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+    assert coordinator.rehydrated
+    assert not coordinator.last_resources
+
+
+async def test_offline_load_without_snapshot_still_fails(hass: HomeAssistant, mock_entry) -> None:
+    """No snapshot means no device metadata to build anything from, so the
+    entry stays on HA's backoff rather than loading empty."""
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    assert mock_entry.state is ConfigEntryState.SETUP_RETRY
+    assert not hass.states.async_all()
+
+
+async def test_corrupt_snapshot_falls_back_to_setup_retry(
+    hass: HomeAssistant, mock_entry, hass_storage
+) -> None:
+    """A snapshot whose resources no longer replay cleanly must not take the
+    entry down with it."""
+    key = _store_key(mock_entry)
+    hass_storage[key] = {
+        "version": 1,
+        "minor_version": 1,
+        "key": key,
+        "data": {"resources": {"/information/vs/0": "not-a-rep"}, "subdevice_candidates": []},
+    }
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    assert mock_entry.state is ConfigEntryState.SETUP_RETRY
+
+
+async def test_snapshot_restores_identity(hass: HomeAssistant, mock_entry) -> None:
+    """`/oic/d`'s device types route the registry, so an offline load that
+    lost them could resolve a different one than the live poll did -- which
+    would show up as a spurious reconcile reload every restart."""
+    resources = _load_fridge()
+    identity = DeviceIdentity(
+        manufacturer="Samsung",
+        model="TEST-MODEL",
+        name="Fridge",
+        serial=None,
+        device_types=("oic.d.refrigerator",),
+        raw={"/oic/p": {}, "/oic/d": {}, "/oic/res": []},
+    )
+    with _reachable(resources, identity):
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+        await _flush_snapshot(hass)
+        await hass.config_entries.async_unload(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+    assert coordinator._identity is not None
+    assert coordinator._identity.device_types == ("oic.d.refrigerator",)
+
+
+# ---------------------------------------------------------------------------
+# Recovery and reconciliation
+# ---------------------------------------------------------------------------
+
+
+async def test_offline_load_keeps_polling_and_recovers(hass: HomeAssistant, mock_entry) -> None:
+    """The failure the PR this replaces actually shipped: with zero live
+    listeners the base coordinator stops rescheduling, and the entry never
+    polls again. Entities must go available on the next interval once the
+    appliance answers."""
+    resources = _load_fridge()
+    await _setup_online_then_unload(hass, mock_entry, resources)
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+    assert coordinator._unsub_refresh is not None  # a poll is actually queued
+    assert not coordinator.last_update_success
+
+    with _reachable(resources):
+        await _tick(hass)
+
+    assert coordinator.last_update_success
+    assert any(s.state != "unavailable" for s in hass.states.async_all())
+
+
+async def test_reconcile_reloads_when_live_discovery_differs(
+    hass: HomeAssistant, mock_entry
+) -> None:
+    """Platforms enumerate `bound` once, so a live set that disagrees with the
+    snapshot can only be adopted by bringing the entry back up."""
+    resources = _load_fridge()
+
+    with _reachable(resources):
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+        coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+        victim = coordinator.bound[0].href
+        await _flush_snapshot(hass)
+        await hass.config_entries.async_unload(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    reduced = {href: rep for href, rep in resources.items() if href != victim}
+    with (
+        _reachable(reduced),
+        patch.object(hass.config_entries, "async_schedule_reload") as reload,
+    ):
+        await _tick(hass)
+
+    reload.assert_called_once_with(mock_entry.entry_id)
+
+
+async def test_reconcile_is_quiet_when_live_discovery_agrees(
+    hass: HomeAssistant, mock_entry
+) -> None:
+    """The common case -- same appliance, same firmware -- must not reload,
+    or every offline restart would cost a second setup cycle."""
+    resources = _load_fridge()
+    await _setup_online_then_unload(hass, mock_entry, resources)
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    with (
+        _reachable(resources),
+        patch.object(hass.config_entries, "async_schedule_reload") as reload,
+    ):
+        await _tick(hass)
+
+    reload.assert_not_called()
+
+
+async def test_live_load_never_reconciles(
+    hass: HomeAssistant, mock_entry, mock_coordinator_session
+) -> None:
+    """An entry that came up against a live device has nothing to reconcile
+    against; the reload path must stay out of the normal startup entirely."""
+    with patch.object(hass.config_entries, "async_schedule_reload") as reload:
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+
+    coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+    assert not coordinator.rehydrated
+    reload.assert_not_called()
+
+
+@pytest.mark.parametrize("failures", [1, 3])
+async def test_offline_load_survives_repeated_poll_failures(
+    hass: HomeAssistant, mock_entry, failures: int
+) -> None:
+    """Recovery isn't one-shot: the entry keeps its entities and keeps
+    retrying across however many intervals the appliance stays dark."""
+    resources = _load_fridge()
+    online_ids = await _setup_online_then_unload(hass, mock_entry, resources)
+
+    with _unreachable():
+        await hass.config_entries.async_setup(mock_entry.entry_id)
+        await hass.async_block_till_done()
+        for _ in range(failures):
+            await _tick(hass)
+
+    assert mock_entry.state is ConfigEntryState.LOADED
+    assert {s.entity_id for s in hass.states.async_all()} == online_ids
+
+    with _reachable(resources):
+        await _tick(hass)
+
+    coordinator: LocalThingsCoordinator = hass.data[DOMAIN][mock_entry.entry_id]
+    assert coordinator.last_update_success
diff --git a/tests/test_airconditioner_ailp_fac_capabilities.py b/tests/test_airconditioner_ailp_fac_capabilities.py
index 9316c8f..4cff87a 100644
--- a/tests/test_airconditioner_ailp_fac_capabilities.py
+++ b/tests/test_airconditioner_ailp_fac_capabilities.py
@@ -36,6 +36,15 @@ class _FakeCoordinator:
     def canonical_resources(self, subdevice):
         return self.last_resources
 
+    # _is_included judges existence against the discovery view, which is the
+    # live cache for everything but an offline load (issue #295).
+    @property
+    def discovery_resources(self):
+        return self.last_resources
+
+    def discovery_canonical(self, subdevice):
+        return self.canonical_resources(subdevice)
+
 
 def _resources():
     return _load_device("airconditioner_ailp_fac")
diff --git a/tests/test_entity.py b/tests/test_entity.py
index f0c0b9b..2eb234e 100644
--- a/tests/test_entity.py
+++ b/tests/test_entity.py
@@ -29,6 +29,15 @@ class _FakeCoordinator:
         # subdevices (issue #177).
         return self.last_resources
 
+    # _is_included judges existence against the discovery view, which is the
+    # live cache for everything but an offline load (issue #295).
+    @property
+    def discovery_resources(self):
+        return self.last_resources
+
+    def discovery_canonical(self, subdevice):
+        return self.canonical_resources(subdevice)
+
 
 def _coord(last_resources) -> LocalThingsCoordinator:
     return cast(LocalThingsCoordinator, _FakeCoordinator(last_resources))
diff --git a/tests/test_unique_ids.py b/tests/test_unique_ids.py
index f679508..643a406 100644
--- a/tests/test_unique_ids.py
+++ b/tests/test_unique_ids.py
@@ -34,6 +34,15 @@ class _FakeCoordinator:
 
         return canonical_view(subdevice, self.last_resources, self._subdevices)
 
+    # _is_included judges existence against the discovery view, which is the
+    # live cache for everything but an offline load (issue #295).
+    @property
+    def discovery_resources(self):
+        return self.last_resources
+
+    def discovery_canonical(self, subdevice):
+        return self.canonical_resources(subdevice)
+
 
 @pytest.mark.parametrize("name", _FIXTURE_NAMES)
 def test_key_is_unique_across_all_bound_entities(name):