diff --git a/crud.py b/crud.py
index 6eda100..a622d96 100644
--- a/crud.py
+++ b/crud.py
@@ -5,7 +5,7 @@
 # through dca_machines.operator_user_id. See plan section "Identity & multi-
 # machine model".
 
-from datetime import datetime
+from datetime import datetime, timezone
 
 from lnbits.db import Database
 from lnbits.helpers import urlsafe_short_hash
@@ -1442,7 +1442,7 @@ async def upsert_fleet_snapshot(
 # Cassette configs — operator-driven ATM cassette inventory (#29 v1.1).
 # =============================================================================
 # Row lifecycle per #29:
-#   - First population for a (machine_id, position) pair → apply_bootstrap_state
+#   - First population for a (machine_id, position) pair → apply_reported_state
 #     (consumer reading the ATM's one-shot bitspire-cassettes-state event)
 #   - Operator edit of denomination or count → update_cassette_config
 #     (refuses to create new rows; the slot count is hardware-determined)
@@ -1450,25 +1450,56 @@ async def upsert_fleet_snapshot(
 #     re-provisioning + new bootstrap event (not exposed in v1 here)
 
 
-def _should_apply_bootstrap_state(
-    existing_state_event_id: str | None, incoming_event_id: str
-) -> bool:
-    """Pure-function dedup gate for apply_bootstrap_state.
+def _as_unix(value) -> float | None:
+    """Normalise whatever the driver hands back for state_at to a unix float.
 
-    Returns False if any existing row for this machine already references
-    the incoming event_id (relay re-delivery after restart). True otherwise.
-
-    Extracted as a pure function so the dedup decision is unit-testable
-    without a database round-trip. The actual idempotency check in
-    apply_bootstrap_state fetches one existing row and passes its
-    state_event_id here.
+    SQLite stores these as integers and Postgres as timestamps, and an event's
+    created_at arrives tz-aware, so comparing the raw values risks either a
+    TypeError (aware vs naive) or a silently wrong answer. Everything is
+    compared as seconds since the epoch instead.
     """
-    return existing_state_event_id != incoming_event_id
+    if value is None:
+        return None
+    if isinstance(value, datetime):
+        if value.tzinfo is None:
+            value = value.replace(tzinfo=timezone.utc)
+        return value.timestamp()
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return None
 
 
-async def get_cassette_config(
-    machine_id: str, position: int
-) -> CassetteConfig | None:
+def _should_apply_state_event(oldest_state_at, incoming_created_at) -> bool:
+    """Ordering gate for apply_reported_state.
+
+    Applies only when the incoming event is strictly newer than the OLDEST
+    state stamp on file for the machine.
+
+    Two deliberate choices:
+
+    - Compare created_at, not event ids. The old gate asked only whether the
+      incoming id differed from one stored row, which is a one-event memory:
+      a re-delivered A, B, A applied three times, and an event that arrived
+      late overwrote newer state because nothing ever looked at the clock.
+      Strict `>` also subsumes exact-replay dedup, since a replay carries the
+      same stamp.
+    - Oldest, not newest. Every execute in this data layer commits on its own,
+      so a multi-row apply cannot be made atomic here; a crash mid-loop leaves
+      some rows advanced and some not. Gating on the oldest means a partial
+      apply is re-applied on the next event rather than being mistaken for a
+      complete one, and the ATM republishes on a heartbeat, so it converges.
+    """
+    oldest = _as_unix(oldest_state_at)
+    if oldest is None:
+        return True
+    incoming = _as_unix(incoming_created_at)
+    if incoming is None:
+        return False
+    return incoming > oldest
+
+
+async def get_cassette_config(machine_id: str, position: int) -> CassetteConfig | None:
     return await db.fetchone(
         "SELECT * FROM spirekeeper.cassette_configs "
         "WHERE machine_id = :mid AND position = :pos",
@@ -1497,7 +1528,7 @@ async def update_cassette_config(
 ) -> CassetteConfig | None:
     """Operator-driven row update: change denomination and/or count for a
     single cassette slot. Refuses to create new rows — those only land via
-    apply_bootstrap_state() consuming an ATM bootstrap event (per #29 row
+    apply_reported_state() consuming an ATM bootstrap event (per #29 row
     lifecycle: hardware-determined slot count, not operator-creatable).
     Returns None if the (machine_id, position) row doesn't exist.
     """
@@ -1520,41 +1551,60 @@ async def update_cassette_config(
     return await get_cassette_config(machine_id, position)
 
 
-async def apply_bootstrap_state(
+async def apply_reported_state(
     machine_id: str,
     event_id: str,
     event_created_at: datetime,
     payload: PublishCassettesPayload,
 ) -> bool:
     """Consume an ATM-published kind-30078 bitspire-cassettes-state:<m> event
-    and upsert one cassette_configs row per position in the payload.
+    and reconcile cassette_configs for the machine against it.
 
-    Returns True if the upsert ran; False if any existing row for this
-    machine already references this event_id (idempotent on relay
-    re-delivery / restart).
+    Returns True if the state was applied, False if the event was not newer
+    than what is already on file (see _should_apply_state_event).
+
+    The payload is the machine's full bay set, and the machine owns that
+    layout — the bay count is hardware-determined. So positions absent from
+    the payload are DELETED here. Leaving them behind was half the problem a
+    shrinking bay count caused: the stale row stayed, the dashboard showed a
+    mix of old and new, and every later publish was rejected for a position
+    mismatch with no way out but hand-written SQL.
 
     Populates both the operator-believed columns (denomination, count,
-    updated_at, updated_by='atm-bootstrap') AND the v2 reverse-channel
-    columns (state_denomination, state_count, state_at, state_event_id)
-    so the operator's initial view matches the ATM's reported state. v2
-    reconciliation UI will diverge them when continuous reverse-channel
-    events land + the operator subsequently edits.
+    updated_at, updated_by) and the reported columns (state_denomination,
+    state_count, state_at, state_event_id), so the UI can show reported
+    against believed.
     """
-    existing_first: dict | None = await db.fetchone(
-        "SELECT state_event_id FROM spirekeeper.cassette_configs "
-        "WHERE machine_id = :mid LIMIT 1",
+    oldest: dict | None = await db.fetchone(
+        "SELECT state_at FROM spirekeeper.cassette_configs "
+        "WHERE machine_id = :mid AND state_at IS NOT NULL "
+        "ORDER BY state_at ASC LIMIT 1",
         {"mid": machine_id},
     )
-    existing_event_id: str | None = None
-    if existing_first is not None:
-        existing_event_id = (
-            existing_first.get("state_event_id")
-            if isinstance(existing_first, dict)
-            else getattr(existing_first, "state_event_id", None)
+    oldest_state_at = None
+    if oldest is not None:
+        oldest_state_at = (
+            oldest.get("state_at")
+            if isinstance(oldest, dict)
+            else getattr(oldest, "state_at", None)
         )
-    if not _should_apply_bootstrap_state(existing_event_id, event_id):
+    if not _should_apply_state_event(oldest_state_at, event_created_at):
         return False
 
+    # Drop bays the machine no longer reports, before writing the rest. A crash
+    # between the two leaves rows missing rather than stale, and the next
+    # heartbeat re-inserts them — the safe direction.
+    keep = sorted(payload.positions.keys())
+    placeholders = ", ".join(f":p{i}" for i in range(len(keep)))
+    delete_values: dict = {"mid": machine_id}
+    for i, pos in enumerate(keep):
+        delete_values[f"p{i}"] = pos
+    await db.execute(
+        "DELETE FROM spirekeeper.cassette_configs "
+        f"WHERE machine_id = :mid AND position NOT IN ({placeholders})",
+        delete_values,
+    )
+
     now = datetime.now()
     for pos, row in payload.positions.items():
         await db.execute(
@@ -1581,7 +1631,7 @@ async def apply_bootstrap_state(
                 "denom": row.denomination,
                 "count": row.count,
                 "now": now,
-                "by": "atm-bootstrap",
+                "by": "atm-report",
                 "state_denom": row.denomination,
                 "state_count": row.count,
                 "state_at": event_created_at,
diff --git a/tasks.py b/tasks.py
index e3865d2..61695e1 100644
--- a/tasks.py
+++ b/tasks.py
@@ -244,20 +244,20 @@ async def _record_rejected(payment: Payment, machine: Machine, exc: Exception) -
 
 
 # =============================================================================
-# Cassette bootstrap consumer (#29 v1)
+# Cassette state consumer (#29)
 # =============================================================================
 # Subscribes to kind-30078 bitspire-cassettes-state:<atm_pubkey_hex> events
-# published by each active machine's ATM on first boot (lamassu-next#56's
-# bootstrap publish path). Decrypts the NIP-44 v2 content with the operator's
-# privkey + ATM sender pubkey, validates as PublishCassettesPayload, and
-# upserts cassette_configs via apply_bootstrap_state.
+# published by each active machine's ATM. Decrypts the NIP-44 v2 content with
+# the operator's privkey + ATM sender pubkey, validates as
+# PublishCassettesPayload, and reconciles cassette_configs via
+# apply_reported_state.
 #
-# v1 = one-shot per machine (ATM's meta.bootstrapPublishedAt makes the
-# publish idempotent on ATM-side restart; spirekeeper's apply_bootstrap_
-# state dedups on state_event_id for relay re-delivery).
-#
-# v2 (separate issue) = continuous reverse-channel consumer with a
-# last_state_created_at watermark for reconciliation UI.
+# This is continuous, not one-shot: the ATM publishes on startup, after every
+# change to its bays, and on a heartbeat, and each event replaces the last.
+# The comments here used to call it a "v1 bootstrap" consumer, which was
+# misleading — there has never been a once-per-machine guard, so every event
+# an ATM published was already being applied. Ordering is enforced in
+# apply_reported_state by created_at; it is not inferred from arrival order.
 #
 # Implementation: polls nostrclient.router.NostrRouter.received_subscription_
 # events keyed by our subscription_id. nostrclient's NostrRouter design is
@@ -285,7 +285,7 @@ async def wait_for_cassette_state_events() -> None:
         between retries (operator may install it later)
       - inbound event fails sig-verify / decrypt / parse → log + skip
         the event, continue the loop
-      - apply_bootstrap_state errors → log + skip
+      - apply_reported_state errors → log + skip
     """
     logger.info(
         "spirekeeper v2: cassette bootstrap consumer starting "
@@ -335,7 +335,7 @@ async def _cassette_consumer_tick(current_filter_key: str | None) -> str:
 
     from .cassette_transport import build_state_d_tags_for_machines
     from .crud import (
-        apply_bootstrap_state,
+        apply_reported_state,
         get_machine_by_atm_pubkey_hex,
         list_all_active_machines,
     )
@@ -372,7 +372,7 @@ async def _cassette_consumer_tick(current_filter_key: str | None) -> str:
                 await _handle_cassette_state_event(
                     event_message,
                     get_machine_by_atm_pubkey_hex,
-                    apply_bootstrap_state,
+                    apply_reported_state,
                 )
             except Exception as exc:
                 logger.warning(
@@ -386,7 +386,7 @@ async def _cassette_consumer_tick(current_filter_key: str | None) -> str:
 async def _handle_cassette_state_event(
     event_message,
     get_machine_by_atm_pubkey_hex,
-    apply_bootstrap_state,
+    apply_reported_state,
 ) -> None:
     """Verify signature, resolve the operator's signer, decrypt via the
     signer abstraction (bunker round-trip for RemoteBunkerSigner; direct
@@ -482,7 +482,7 @@ async def _handle_cassette_state_event(
     created_at_unix = event_obj.get("created_at", 0)
     event_created_at = _datetime.fromtimestamp(int(created_at_unix), tz=_timezone.utc)
 
-    applied = await apply_bootstrap_state(
+    applied = await apply_reported_state(
         machine.id, event_id, event_created_at, payload
     )
     if applied:
diff --git a/tests/test_cassette_configs.py b/tests/test_cassette_configs.py
index 703c3bb..242aa13 100644
--- a/tests/test_cassette_configs.py
+++ b/tests/test_cassette_configs.py
@@ -5,11 +5,11 @@ Covers the pure pieces that don't need a live DB:
   - Pydantic validator behaviour on PublishCassettesPayload + the row /
     upsert models (position key coercion, integer ranges, multiple-same-
     denomination payloads, wire-format round-trip)
-  - _should_apply_bootstrap_state dedup helper (extracted from
-    apply_bootstrap_state so the relay-re-delivery decision is testable
-    without a database round-trip)
+  - _should_apply_state_event ordering gate (extracted from
+    apply_reported_state so the decision is testable without a database
+    round-trip)
 
-DB-touching tests (apply_bootstrap_state actually upserting, list-by-
+DB-touching tests (apply_reported_state actually reconciling, list-by-
 machine ordering, etc.) follow the project convention from
 test_deposit_currency.py: "Layer 2 is an endpoint-level behaviour better
 covered by an integration test against a running LNbits; tracked in #26
@@ -21,9 +21,11 @@ denomination + count are operator-editable per row, multiple same-denom
 cassettes are valid.
 """
 
+from datetime import datetime, timedelta, timezone
+
 import pytest
 
-from ..crud import _should_apply_bootstrap_state
+from ..crud import _as_unix, _should_apply_state_event
 from ..models import (
     CassettePayloadRow,
     PublishCassettesPayload,
@@ -186,35 +188,62 @@ class TestUpsertCassetteConfigData:
 
 
 # =============================================================================
-# _should_apply_bootstrap_state — relay re-delivery dedup
+# _should_apply_state_event — ordering gate
 # =============================================================================
 
+NOW = datetime(2026, 9, 22, 12, 0, 0, tzinfo=timezone.utc)
 
-class TestShouldApplyBootstrapState:
-    """Pure-function dedup gate extracted from apply_bootstrap_state so the
-    decision is testable without a DB. Logic: apply if-and-only-if the
-    existing row's state_event_id differs from the incoming event_id.
 
-    In v1.1 the ATM publishes the bootstrap event exactly once per machine,
-    so this is sufficient for replay protection. v2 will need a
-    `last_state_created_at` watermark in addition (per bitspire's
-    `meta.lastKnownConfigCreatedAt` on the ATM side) — flagged in #29's
-    v2 forward-look section.
+class TestShouldApplyStateEvent:
+    """Ordering gate extracted from apply_reported_state so the decision is
+    testable without a DB.
+
+    It compares created_at against the OLDEST state stamp on file. The gate it
+    replaced compared event ids against a single arbitrary row, which is a
+    one-event memory: a re-delivered A, B, A applied three times, and a late
+    event overwrote newer state because nothing looked at the clock.
     """
 
-    def test_applies_when_no_existing_row(self):
-        assert _should_apply_bootstrap_state(None, "new-event-id") is True
+    def test_applies_when_machine_has_no_state_yet(self):
+        assert _should_apply_state_event(None, NOW) is True
 
-    def test_applies_when_existing_event_id_differs(self):
-        assert _should_apply_bootstrap_state("old-event-id", "new-event-id") is True
+    def test_applies_when_strictly_newer(self):
+        assert _should_apply_state_event(NOW, NOW + timedelta(seconds=1)) is True
 
-    def test_skips_when_existing_event_id_matches(self):
-        """The same bootstrap event re-delivered after a relay reconnect
-        or spirekeeper restart should no-op, not re-upsert the same
-        rows (which would clobber any operator edits since)."""
-        assert _should_apply_bootstrap_state("same-event", "same-event") is False
+    def test_skips_an_exact_replay(self):
+        """A re-delivered event carries the same stamp, so strict `>` covers
+        replay without needing to remember ids."""
+        assert _should_apply_state_event(NOW, NOW) is False
 
-    def test_applies_when_existing_is_empty_string_and_incoming_is_id(self):
-        """Defensive — a sentinel empty-string existing_state_event_id
-        shouldn't block a real incoming event from applying."""
-        assert _should_apply_bootstrap_state("", "real-event-id") is True
+    def test_skips_an_event_that_arrives_late(self):
+        """The failure the id-only gate allowed: an older event landing after
+        a newer one used to overwrite it."""
+        assert _should_apply_state_event(NOW, NOW - timedelta(seconds=30)) is False
+
+    def test_reapplies_over_a_partial_write(self):
+        """Every execute in this data layer commits on its own, so a crash
+        mid-apply can leave some rows advanced and some not. Gating on the
+        oldest stamp means the next event re-applies rather than treating the
+        partial write as complete."""
+        partially_applied_oldest = NOW - timedelta(seconds=300)
+        assert _should_apply_state_event(partially_applied_oldest, NOW) is True
+
+    def test_handles_naive_and_epoch_stamps(self):
+        """SQLite hands these back as integers and Postgres as timestamps, and
+        the incoming stamp is tz-aware. Comparing raw would raise or silently
+        mislead."""
+        assert _as_unix(NOW) == NOW.timestamp()
+        assert _as_unix(NOW.replace(tzinfo=None)) == NOW.timestamp()
+        assert _as_unix(int(NOW.timestamp())) == float(int(NOW.timestamp()))
+        assert _as_unix(None) is None
+        assert _as_unix("not-a-time") is None
+        assert _should_apply_state_event(int(NOW.timestamp()), NOW) is False
+        assert (
+            _should_apply_state_event(int(NOW.timestamp()), NOW + timedelta(seconds=5))
+            is True
+        )
+
+    def test_skips_when_incoming_stamp_is_unusable(self):
+        """Fail closed: an unparseable incoming stamp must not overwrite
+        state that is known-good."""
+        assert _should_apply_state_event(NOW, "nonsense") is False
diff --git a/tests/test_cassette_state_consumer.py b/tests/test_cassette_state_consumer.py
index a0840bc..d7cdd8e 100644
--- a/tests/test_cassette_state_consumer.py
+++ b/tests/test_cassette_state_consumer.py
@@ -22,7 +22,7 @@ demonstrates the project lacks an asyncio plugin in CI; using asyncio.run
 inside the test body sidesteps that without changing project config).
 
 Full handler tests (the dispatch through verify_event →
-get_machine_by_atm_pubkey_hex → apply_bootstrap_state) need a live LNbits
+get_machine_by_atm_pubkey_hex → apply_reported_state) need a live LNbits
 DB; smoke-tested manually via the dev container per the project
 convention (see test_deposit_currency.py rationale).
 """