Make the missed-sweep watermark claim safe across processes

The sweep guarded its announcement with a read-then-set watermark
inside a per-process asyncio lock, so two processes sharing one SQLite
file (the HA integration's rollover sweep and the API scheduler) could
both read the old watermark and both announce the same missed batch.

Add MetaStateDao.claim: a single conditional upsert
(INSERT ... ON CONFLICT DO UPDATE ... WHERE value IS NOT excluded.value)
whose rowcount says whether THIS caller changed the row — the
read-compare-write is one SQLite statement, atomic at the database
level, not just within one process. run_missed_sweep now builds the
events optimistically and gates the announcement on that claim: the
caller that advances the watermark returns the events, a loser returns
[]. The claim happens only after the events are built, so a failure
before it leaves the watermark unadvanced and a later run retries;
get/set are intact and the in-process sweep lock still serialises
same-process callers.

Tests prove two independent NestQuestDatabase wrappers (the
two-process shape) racing one seeded database yield exactly one
announcer, plus the unchanged same-night no-op and post-downtime
window-once semantics.

.dockerignore

@@ -1,27 +0,0 @@
-# Keep the image build context minimal: only api/ and the bundled
-# custom_components/nestquest/core/ are copied (see the Dockerfile);
-# everything else — VCS state, virtualenvs, tests, the frontend, and
-# any local secrets or database files — never enters the build.
-.git
-.gitignore
-.venv
-**/__pycache__
-**/*.py[cod]
-.pytest_cache
-**/*.db
-**/*.db-wal
-**/*.db-shm
-# Local configuration — never baked into an image.
-.env
-**/.env
-deploy
-design
-docs
-frontend
-orchestration_tests
-scripts
-templates
-tests
-uv.lock
-hacs.json
-*.md

.gitignore

@@ -5,12 +5,3 @@ __pycache__/
 .pytest_cache/
 *.metadata.json
 frontend/node_modules/
-
-# Feature 17 (API deployment): local secret config and runtime data
-# are never committed — compose reads deploy/.env; SQLite/WAL sidecar
-# files are container/volume artifacts.
-.env
-**/.env
-*.db
-*.db-wal
-*.db-shm

Dockerfile

@@ -1,47 +0,0 @@
-# Container image for the NestQuest API service (Feature 17).
-#
-# Build context is the REPOSITORY ROOT (compose sets it), because the
-# API loads the bundled domain layer from
-# custom_components/nestquest/core/ by file location
-# (api/nestquest_core.py): the image must carry both api/ and the
-# bundled core at their relative positions.  Everything else in the
-# repository is excluded via .dockerignore.
-#
-# Multi-stage so no pip/build toolchain survives into the runtime:
-# the builder installs the deployable's own manifest
-# (api/requirements.txt) into a staging prefix and the runtime copies
-# only the installed packages.  The service runs as a non-root user;
-# the SQLite database lives on the mounted volume at
-# /var/lib/nestquest (compose supplies NESTQUEST_DB_PATH pointing
-# there — configuration is environment-supplied, never baked in).
-FROM python:3.12-slim AS builder
-
-COPY api/requirements.txt /tmp/requirements.txt
-RUN pip install --no-cache-dir --prefix=/install -r /tmp/requirements.txt
-
-FROM python:3.12-slim
-
-# The staged dependency install from the builder (no pip, no wheels).
-COPY --from=builder /install /usr/local
-
-# /app is the repository-root equivalent: api/ resolves imports as
-# ``api.*`` and nestquest_core.py finds the bundled core one level up.
-WORKDIR /app
-COPY api/ ./api/
-COPY custom_components/nestquest/core/ ./custom_components/nestquest/core/
-
-# Non-root runtime user; owns the database volume mount point.
-RUN useradd --system --create-home --home-dir /home/nestquest nestquest \
-    && mkdir -p /var/lib/nestquest \
-    && chown nestquest:nestquest /var/lib/nestquest
-USER nestquest
-
-ENV PYTHONUNBUFFERED=1 PYTHONDONTWRITEBYTECODE=1
-
-# The database path and the service token / OIDC settings are supplied
-# by the environment (compose env_file), never committed into the image.
-ENV NESTQUEST_DB_PATH=/var/lib/nestquest/nestquest.db
-VOLUME /var/lib/nestquest
-
-EXPOSE 8080
-CMD ["uvicorn", "--factory", "api.app:create_app", "--host", "0.0.0.0", "--port", "8080"]

api/README.md

@@ -1,32 +0,0 @@
-# NestQuest API
-
-Standalone FastAPI service owning the NestQuest database (Feature 16).
-Served with ``uvicorn --factory api.app:create_app``; configuration is
-environment-supplied (``api/config.py``) and every variable is required.
-
-## Runtime dependencies
-
-``api/requirements.txt`` is the API deployable's OWN dependency
-manifest (fastapi, uvicorn, pyjwt[crypto], httpx) — deliberately kept
-out of the integration's ``pyproject.toml`` dev group, which exists only
-so the monorepo test venv can run the API tests.
-
-## Container deployment (scaffold)
-
-The image is built from the repository ROOT context (the API loads the
-bundled domain core from ``custom_components/nestquest/core/`` by file
-location, so both trees must be in the build context):
-
-    docker build -t nestquest-api:latest .
-
-Run it with compose — the SQLite file lives on the named volume
-``nestquest-db`` so recreating the container keeps household data:
-
-    cp deploy/.env.example deploy/.env   # then fill in real values
-    docker compose -f deploy/compose.yaml up -d --build
-
-``deploy/.env`` is git-ignored and supplies the secrets (panel service
-token, Authentik OIDC settings); ``NESTQUEST_DB_PATH`` is fixed to
-``/var/lib/nestquest/nestquest.db`` by compose. The port is bound to
-localhost only — TLS/reverse proxy, Authentik, network separation,
-backup and restore are later Feature 17 tasks.

api/requirements.txt

@@ -1,15 +0,0 @@
-# Runtime dependency manifest for the NestQuest API deployable.
-#
-# The monorepo's pyproject.toml carries fastapi/uvicorn/pyjwt/httpx in
-# the INTEGRATION's dev group only (so the test venv can exercise the
-# API); they are deliberately NOT integration runtime dependencies.
-# Per that manifest's own comment, the API deployable declares its own
-# runtime dependencies HERE — this file is what the container image
-# installs (see the Dockerfile at the repository root).
-#
-# Compatible ranges mirroring the dev group's floors; the API is a
-# server, so these are the only packages its runtime needs.
-fastapi>=0.115
-uvicorn>=0.30
-pyjwt[crypto]>=2.9
-httpx>=0.28

custom_components/nestquest/core/dao_meta.py

@@ -41,4 +41,38 @@ class MetaStateDao:
                     f"INSERT INTO {_TABLE} (key, value) VALUES (?, ?) "
                     "ON CONFLICT (key) DO UPDATE SET value = excluded.value",
                     (key, value),
-                )
+                )
+
+    async def claim(self, key: str, value: str) -> bool:
+        """Atomically set ``key`` to ``value`` unless it already is.
+
+        The single upsert statement carries a ``WHERE`` guard on the
+        conflict path (``t.value IS NOT excluded.value``), so the
+        read-compare-write is ONE SQLite statement — atomic at the
+        DATABASE level, not just within one process: two connections
+        (two processes sharing one SQLite file) cannot both observe
+        "not yet set" and both write, because the second writer's
+        statement re-evaluates the guard against the first writer's
+        committed value and matches zero rows.
+
+        Returns True only when THIS call changed the row: the first
+        claimant wins and every later claimant of the same value
+        returns False.  A DIFFERENT value always wins (an advance, e.g.
+        the sweep watermark moving to a new date) and re-arms the
+        same-value loser check for the new value.
+
+        Runs under the shared connection lock and inside one
+        transaction, like :meth:`set`; ``cursor.rowcount`` (1 for a
+        matched write, 0 for a guarded no-op) is captured by the
+        wrapper on the executor thread, so no sqlite3 object crosses
+        back to the event loop.
+        """
+        async with _connection_lock(self._database):
+            async with self._database.transaction():
+                result = await self._database.execute(
+                    f"INSERT INTO {_TABLE} (key, value) VALUES (?, ?) "
+                    "ON CONFLICT (key) DO UPDATE SET value = excluded.value "
+                    f"WHERE {_TABLE}.value IS NOT excluded.value",
+                    (key, value),
+                )
+        return result.rowcount == 1

custom_components/nestquest/core/sweep.py

@@ -26,12 +26,17 @@ later completed simply gains its completion event; history keeps both
 rows (append-only), and the sweep's later runs skip it via the
 no-completion-event rule.
 
-The whole watermark read → query → build → write span runs inside the
+The whole watermark read → query → build span runs inside the
 per-database sweep lock (one lock per connection wrapper and running
 loop, the settings store's pattern): an overlapping startup sweep and
-a rollover sweep queue instead of interleaving, and the loser re-reads
-the watermark INSIDE the lock — sees the winner's date — and announces
-nothing twice.
+a rollover sweep of the SAME process queue instead of interleaving.
+The announce decision itself is the DATABASE-level claim
+(:meth:`MetaStateDao.claim`), so two PROCESSES sharing one SQLite file
+(the integration's rollover sweep and the API service's scheduler)
+cannot both announce the same missed batch either: both may read the
+old watermark and build the events, but the single conditional upsert
+lets exactly one of them win the claim — the loser returns [] and
+announces nothing.
 
 The caller supplies ``today`` as a plain :class:`datetime.date`: the
 household-local calendar date, read from ``hass.config.time_zone`` by
@@ -60,11 +65,11 @@ LOGGER = logging.getLogger(__name__)
 SWEEP_WATERMARK_KEY = "missed_sweep_last_run_date"
 
 #: One asyncio.Lock per (connection wrapper, running loop), mirroring
-#: the settings store's pattern: the whole watermark read → query →
-#: build → write span is one critical section, so an overlapping
-#: startup sweep and rollover sweep queue instead of double-announcing.
-#: The loser re-reads the watermark INSIDE the lock, sees the winner's
-#: date, and fires nothing.
+#: the settings store's pattern: the watermark read → query → build
+#: span is one critical section, so an overlapping startup sweep and
+#: rollover sweep of the SAME process queue instead of interleaving.
+#: Same-process callers are fully serialised by it; the cross-process
+#: case is closed by the database-level claim, not by this lock.
 _SWEEP_LOCKS: dict[tuple[int, int], asyncio.Lock] = {}
 
 
@@ -99,15 +104,24 @@ async def run_missed_sweep(
     the caller decides how to publish them and announces nothing on
     its own.
 
-    The whole watermark read → query → build → write span runs inside
-    the per-database sweep lock: an overlapping startup sweep and a
-    rollover sweep queue instead of interleaving, and the loser
-    re-reads the watermark inside the lock — sees the winner's date —
-    and announces nothing twice.  Every event in one run shares one
+    The watermark read → query → build span runs inside the
+    per-database sweep lock, so overlapping callers of ONE process
+    queue instead of interleaving.  The announce decision is the
+    DATABASE-level claim (:meth:`MetaStateDao.claim` — a single
+    conditional upsert on the watermark, atomic across processes):
+    every caller builds the events optimistically, then only the
+    caller whose claim actually advanced the watermark returns them;
+    a loser (another process already announced for ``today``) returns
+    [] and announces nothing twice.  Every event in one run shares one
     ``occurred_at`` stamp (one run, one announcement moment), and the
     payload comes from the shared
     :func:`~.events._instance_payload` builder so the shape can never
     drift from the documented §3 contract.
+
+    Failure ordering: the watermark is advanced ONLY by the claim,
+    AFTER the events are fully built — a failure before the claim
+    leaves the watermark unadvanced, so a later run re-sweeps the same
+    window and nothing is lost.
     """
     today_iso = today.isoformat()
     dao = MetaStateDao(database)
@@ -134,10 +148,23 @@ async def run_missed_sweep(
                     ),
                 )
             )
-        await dao.set(SWEEP_WATERMARK_KEY, today_iso)
+        # The database-level claim gates the announcement: only the
+        # caller that actually advanced the watermark may publish the
+        # built events, so two processes racing one SQLite file cannot
+        # both announce the same missed batch.  A same-value loser's
+        # claim matches zero rows and this run announces nothing.
+        claimed = await dao.claim(SWEEP_WATERMARK_KEY, today_iso)
+        if not claimed:
+            LOGGER.info(
+                "NestQuest missed sweep for %s lost the watermark claim; "
+                "announcing nothing",
+                today_iso,
+            )
+            return []
     LOGGER.info(
         "NestQuest missed sweep for %s built %d event(s)",
         today_iso,
         len(events),
     )
     return events
+

deploy/.env.example

@@ -1,17 +0,0 @@
-# Example environment for the NestQuest API container deployment.
-#
-# Copy to ``deploy/.env`` (git-ignored) and fill in REAL values there;
-# this example carries placeholder names only and is never a source of
-# secrets.  Compose loads ``deploy/.env`` into the api service; the
-# API fails fast on startup when any of these is unset or empty
-# (api/config.py — there are no silent defaults).
-
-# The static panel-plane service token the HA integration presents as
-# a Bearer credential on /api/v1/panel/* routes. Generate one, e.g.:
-#   python -c "import secrets; print(secrets.token_urlsafe(32))"
-NESTQUEST_PANEL_TOKEN=
-
-# The Authentik OIDC settings the admin plane verifies JWTs against.
-NESTQUEST_OIDC_ISSUER=
-NESTQUEST_OIDC_AUDIENCE=
-NESTQUEST_OIDC_JWKS_URL=

deploy/compose.yaml

@@ -1,41 +0,0 @@
-# Compose deployment scaffold for the NestQuest API service (Feature
-# 17).  The SQLite database lives on the NAMED VOLUME ``nestquest-db``
-# mounted at /var/lib/nestquest — recreating the container keeps the
-# household data (``docker compose down`` leaves named volumes alone).
-#
-# Configuration is supplied by the environment: the fixed database
-# path here, and the secret values (panel service token, Authentik
-# OIDC settings) from ``deploy/.env`` — a git-ignored file you create
-# from ``deploy/.env.example`` (placeholder names only; never commit
-# real values).  TLS/reverse proxy, Authentik wiring and network
-# separation are later Feature 17 tasks; the port is bound to
-# localhost only until those land.
-services:
-  api:
-    build:
-      context: ..
-      dockerfile: Dockerfile
-    image: nestquest-api:latest
-    init: true
-    restart: unless-stopped
-    ports:
-      - "127.0.0.1:8080:8080"
-    environment:
-      # The SQLite file inside the named volume; fixed, not a secret.
-      NESTQUEST_DB_PATH: /var/lib/nestquest/nestquest.db
-    env_file:
-      # Secret values (NESTQUEST_PANEL_TOKEN, NESTQUEST_OIDC_*); absent
-      # until you create deploy/.env from the example.
-      - path: .env
-        required: false
-    volumes:
-      - nestquest-db:/var/lib/nestquest
-    healthcheck:
-      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8080/health')"]
-      interval: 10s
-      timeout: 5s
-      retries: 5
-      start_period: 15s
-
-volumes:
-  nestquest-db:

tests/test_core_sweep.py

@@ -0,0 +1,295 @@
+"""Cross-process safety of the missed-sweep watermark claim.
+
+Feature 11's sweep guards its announcement with the watermark; the
+original guard was a read-then-set inside a per-PROCESS asyncio lock,
+so two PROCESSES sharing one SQLite file (the HA integration's
+rollover sweep and the API service's scheduler) could both read the
+old watermark and both announce the same missed batch.  These tests
+pin the fix:
+
+- :meth:`MetaStateDao.claim` is the atomic, database-level primitive:
+  one conditional upsert whose ``rowcount`` says whether THIS caller
+  changed the row.  The first claimant of a value wins; a second
+  claimant of the same value — even on an INDEPENDENT connection to
+  the same file, the shape two processes produce — loses.
+- :func:`run_missed_sweep` gates the announcement on that claim: two
+  independent ``NestQuestDatabase`` wrappers (two processes' shape)
+  racing one seeded database yields exactly one announcer.
+- The documented sweep semantics are unchanged: a same-night rerun
+  announces nothing, and a run after days of downtime announces the
+  accumulated ``[watermark, today)`` window exactly once.
+
+Everything here runs against the bundled HA-free core copy
+(``custom_components.nestquest.core``) with a plain
+``asyncio.to_thread`` executor — no Home Assistant, no API server.
+"""
+from __future__ import annotations
+
+import asyncio
+import datetime
+from types import SimpleNamespace
+
+import pytest
+
+from custom_components.nestquest.core.dao_instances import (
+    QuestInstancesDao,
+)
+from custom_components.nestquest.core.dao_meta import MetaStateDao
+from custom_components.nestquest.core.db import NestQuestDatabase
+from custom_components.nestquest.core.migrations import apply_migrations
+from custom_components.nestquest.core.children import create_child
+from custom_components.nestquest.core.quest_definitions import (
+    create_quest_definition,
+)
+from custom_components.nestquest.core.materialize import materialize
+from custom_components.nestquest.core.recurrence import ScheduleRule
+from custom_components.nestquest.core.sweep import (
+    SWEEP_WATERMARK_KEY,
+    run_missed_sweep,
+)
+
+#: One missed event per unswept past-due instance — the sweep's only
+#: event type.
+EVENT_QUEST_MISSED = "nestquest_quest_missed"
+
+
+async def _executor(fn, *args, **kwargs):
+    """Run a synchronous callable on a worker thread and await it.
+
+    The :class:`NestQuestDatabase` executor contract (the shape of
+    ``hass.async_add_executor_job``), so sqlite3 calls never block the
+    event loop.
+    """
+    return await asyncio.to_thread(fn, *args, **kwargs)
+
+
+async def _open_database(db_path: str) -> NestQuestDatabase:
+    """Open (without migrating) one independent connection wrapper."""
+    database = NestQuestDatabase(_executor)
+    await database.open(db_path)
+    return database
+
+
+async def _seed_missed_household(
+    db_path: str, today: datetime.date
+) -> SimpleNamespace:
+    """Seed a household with two past-due open instances on ``db_path``.
+
+    One child, two quests whose ONLY instances are due YESTERDAY
+    (materialized with ``today`` pinned to yesterday, ``end_date``
+    closing the rules so the later today-walk cannot duplicate them).
+    Seeding uses its OWN connection, opened and closed here, so the
+    connections the tests race with start fresh.
+    """
+    database = await _open_database(db_path)
+    try:
+        await apply_migrations(database)
+        today_iso = today.isoformat()
+        yesterday = today - datetime.timedelta(days=1)
+        yesterday_iso = yesterday.isoformat()
+        rule = ScheduleRule.from_dict(
+            {
+                "rule_type": "daily",
+                "start_date": yesterday_iso,
+                "end_date": yesterday_iso,
+            }
+        )
+        ada = await create_child(database, "Ada", sort_order=0)
+        await create_quest_definition(
+            database, "Stale chore", rule, [ada.id], [("morning", "09:00")]
+        )
+        await create_quest_definition(
+            database, "Old chore", rule, [ada.id], [("evening", "18:30")]
+        )
+        await materialize(database, yesterday_iso, yesterday_iso, today=yesterday)
+        # A quest due TODAY (created after the yesterday walk so it has
+        # no past-due instance): the post-downtime catch-up run's
+        # target — inside [watermark=today, later_today) but not
+        # inside the day-one sweep's window.
+        await create_quest_definition(
+            database,
+            "Recoverable chore",
+            ScheduleRule.from_dict(
+                {"rule_type": "daily", "start_date": today_iso}
+            ),
+            [ada.id],
+            [("afternoon", "17:00")],
+        )
+        await materialize(database, today_iso, today_iso, today=today)
+        instances = QuestInstancesDao(database)
+        rows = await instances.list_by_child_and_date(ada.id, yesterday_iso)
+        assert len(rows) == 2
+        today_rows = await instances.list_by_child_and_date(
+            ada.id, today_iso
+        )
+        assert len(today_rows) == 1
+        return SimpleNamespace(
+            child=ada,
+            instances=rows,
+            recoverable_instance=today_rows[0],
+            yesterday_iso=yesterday_iso,
+            today_iso=today_iso,
+        )
+    finally:
+        await database.close()
+
+
+# --- the claim primitive: atomic at the database level ------------------
+
+
+async def test_claim_first_caller_wins_and_second_same_value_loses(
+    temp_db_path: str,
+) -> None:
+    """claim() is a conditional upsert: first True, same-value second False.
+
+    A different value always wins (an advance re-arms the check), and
+    the stored value follows the winner.
+    """
+    database = await _open_database(temp_db_path)
+    try:
+        await apply_migrations(database)
+        dao = MetaStateDao(database)
+        assert await dao.get(SWEEP_WATERMARK_KEY) is None
+
+        assert await dao.claim(SWEEP_WATERMARK_KEY, "2026-09-01") is True
+        # Same value again: the row is already there — nothing changed.
+        assert await dao.claim(SWEEP_WATERMARK_KEY, "2026-09-01") is False
+        assert await dao.get(SWEEP_WATERMARK_KEY) == "2026-09-01"
+        # A DIFFERENT value is an advance, not a re-claim: it wins.
+        assert await dao.claim(SWEEP_WATERMARK_KEY, "2026-09-02") is True
+        assert await dao.get(SWEEP_WATERMARK_KEY) == "2026-09-02"
+        assert await dao.claim(SWEEP_WATERMARK_KEY, "2026-09-02") is False
+    finally:
+        await database.close()
+
+
+async def test_claim_is_atomic_across_independent_connections(
+    temp_db_path: str,
+) -> None:
+    """Two connections to one file cannot both claim the same value.
+
+    This is the two-process shape: independent sqlite3 connections (one
+    per ``NestQuestDatabase`` wrapper), racing the SAME key on the SAME
+    event loop with no shared lock.  The single conditional upsert is
+    serialized by SQLite itself, so exactly one caller's rowcount is 1.
+    """
+    first = await _open_database(temp_db_path)
+    second = await _open_database(temp_db_path)
+    try:
+        await apply_migrations(first)
+        won = await asyncio.gather(
+            MetaStateDao(first).claim(SWEEP_WATERMARK_KEY, "2026-09-01"),
+            MetaStateDao(second).claim(SWEEP_WATERMARK_KEY, "2026-09-01"),
+        )
+        assert sorted(won) == [False, True]
+        # One stored row, holding the winner's value.
+        assert await MetaStateDao(first).get(SWEEP_WATERMARK_KEY) == (
+            "2026-09-01"
+        )
+        assert await MetaStateDao(second).get(SWEEP_WATERMARK_KEY) == (
+            "2026-09-01"
+        )
+    finally:
+        await first.close()
+        await second.close()
+
+
+# --- the sweep: the announcement is gated by the atomic claim -----------
+
+
+async def test_two_process_shape_race_exactly_one_announces(
+    temp_db_path: str,
+) -> None:
+    """Two independent connections racing the sweep: ONE announcer.
+
+    Each wrapper carries its own connection AND its own per-process
+    sweep lock, so nothing in-process serialises the two runs — the
+    only guard left is the database-level claim.  Both read the unset
+    watermark and both build the events, but exactly one claim wins:
+    one run returns the missed events, the loser returns [].
+    """
+    today = datetime.datetime.now().astimezone().date()
+    seed = await _seed_missed_household(temp_db_path, today)
+    first = await _open_database(temp_db_path)
+    second = await _open_database(temp_db_path)
+    try:
+        results = await asyncio.gather(
+            run_missed_sweep(first, today=today),
+            run_missed_sweep(second, today=today),
+        )
+        announcers = [r for r in results if r]
+        assert len(announcers) == 1
+        event_type, payload = announcers[0][0]
+        assert event_type == EVENT_QUEST_MISSED
+        assert payload["due_date"] == seed.yesterday_iso
+        # Both instances of the seeded window, one event each.
+        assert len(announcers[0]) == 2
+        losers = [r for r in results if not r]
+        assert losers == [[]]
+        # The watermark holds the run's date for whoever looks next.
+        assert (
+            await MetaStateDao(first).get(SWEEP_WATERMARK_KEY)
+            == seed.today_iso
+        )
+    finally:
+        await first.close()
+        await second.close()
+
+
+async def test_same_night_rerun_announces_nothing(temp_db_path: str) -> None:
+    """A second sweep for the same date is an empty no-op (the watermark)."""
+    today = datetime.datetime.now().astimezone().date()
+    await _seed_missed_household(temp_db_path, today)
+    database = await _open_database(temp_db_path)
+    try:
+        first = await run_missed_sweep(database, today=today)
+        assert len(first) == 2
+        second = await run_missed_sweep(database, today=today)
+        assert second == []
+    finally:
+        await database.close()
+
+
+async def test_post_downtime_run_announces_accumulated_window_once(
+    temp_db_path: str,
+) -> None:
+    """After days of downtime the catch-up run sweeps the window once.
+
+    The first run sweeps yesterday's instances (watermark := today);
+    ``today`` then jumps two days and the catch-up run — on a
+    DIFFERENT connection, the two-process shape — announces the
+    accumulated ``[watermark, later_today)`` window exactly once, and
+    a rerun on the original connection is silent again.
+    """
+    today = datetime.datetime.now().astimezone().date()
+    later_today = today + datetime.timedelta(days=2)
+    seed = await _seed_missed_household(temp_db_path, today)
+    first = await _open_database(temp_db_path)
+    second = await _open_database(temp_db_path)
+    try:
+        # Day one: yesterday's two instances fire once.
+        day_one = await run_missed_sweep(first, today=today)
+        assert len(day_one) == 2
+        assert all(
+            payload["due_date"] == seed.yesterday_iso
+            for _, payload in day_one
+        )
+
+        # The downtime catch-up: the window is [watermark, later_today).
+        catchup = await run_missed_sweep(second, today=later_today)
+        assert len(catchup) == 1
+        event_type, payload = catchup[0]
+        assert event_type == EVENT_QUEST_MISSED
+        assert payload["due_date"] == seed.today_iso
+        assert payload["instance_id"] == seed.recoverable_instance.id
+        assert payload["child_id"] == seed.child.id
+
+        # A further rerun announces nothing.
+        assert await run_missed_sweep(first, today=later_today) == []
+        assert (
+            await MetaStateDao(first).get(SWEEP_WATERMARK_KEY)
+            == later_today.isoformat()
+        )
+    finally:
+        await first.close()
+        await second.close()