diff --git a/tests/test_soak.py b/tests/test_soak.py
new file mode 100644
index 0000000..8603348
--- /dev/null
+++ b/tests/test_soak.py
@@ -0,0 +1,111 @@
+"""Soak test — long-run daemon health (issue #101).
+
+A compressed daemon lifetime: 3,000 scheduler ticks (~10 simulated days)
+against real SQLite, one watch with the full stage-2 kit, flips and new
+commits along the way. Guards the slow failure modes: steady-state memory
+growth, per-tick latency drift, retention efficacy, DB growth. The
+budgets are generous tripwires, not benchmarks; widening one is a
+decision that requires a finding, not a fix.
+"""
+
+from __future__ import annotations
+
+import time
+import tracemalloc
+from pathlib import Path
+
+from fakes import COMPLETED, FAILURE, FakeClock, FakeGitHub, FakeSink
+from fakes import SUCCESS
+from nanodot.core.activity import ActivityLog
+from nanodot.core.runner import OBSERVATION_RETENTION
+from nanodot.core.runner import DIGEST, FLAKY, STALE, TaskLoop
+from nanodot.core.tasks import PRTarget, Task, TaskStore
+from nanodot.native.daemon import RunnerDaemon
+
+TARGET = PRTarget.parse("thinkflowlab/nanodot#9")
+TICKS = 3_000
+CADENCE = 300
+DAY = 86_400
+FLIP_EVERY = 500
+COMMIT_EVERY = 1_500
+
+
+def test_soak_long_run_health(home: Path) -> None:
+    clock = FakeClock()
+    store = TaskStore(path=home / "nanodot.db", clock=clock)
+    activity = ActivityLog(path=home / "nanodot.db")
+    github = FakeGitHub(TARGET, required=())  # never completes: no required
+    github.set_pr("open", head_sha="s0")
+    github.checks["s0"] = []
+    github.add_check("ci", FAILURE, sha="s0")
+    sink = FakeSink()
+    loop = TaskLoop(store, github, sink, activity)
+    task = store.create(
+        Task(
+            target=TARGET, purpose="soak", cadence_seconds=CADENCE,
+            digest_interval_seconds=DAY, stale_after_seconds=2 * DAY,
+            flaky_alerts=True, next_check_at=0.0,
+        )
+    )
+    daemon = RunnerDaemon(loop, store, clock=clock)
+
+    # Warmup (schema, first tick) before tracing so setup noise stays out.
+    daemon.tick()
+    clock.advance(CADENCE)
+    tracemalloc.start()
+    setup_current, _ = tracemalloc.get_traced_memory()
+    tracemalloc.reset_peak()
+
+    timings: list[float] = []
+    try:
+        for tick in range(1, TICKS + 1):
+            if tick % FLIP_EVERY == 0:
+                runs = github.checks[github.head_sha]
+                flipped = SUCCESS if runs[-1].conclusion == FAILURE else FAILURE
+                github.checks[github.head_sha] = [
+                    type(runs[-1])(
+                        name=runs[-1].name, status=COMPLETED,
+                        conclusion=flipped, sha=github.head_sha,
+                    )
+                ]
+            if tick % COMMIT_EVERY == 0:
+                new_sha = f"s{tick // COMMIT_EVERY}"
+                github.set_pr("open", head_sha=new_sha)
+                github.checks[new_sha] = []
+                github.add_check("ci", FAILURE, sha=new_sha)
+            start = time.monotonic()
+            daemon.tick()
+            timings.append(time.monotonic() - start)
+            clock.advance(CADENCE)
+        end_current, peak = tracemalloc.get_traced_memory()
+    finally:
+        tracemalloc.stop()
+
+    # Memory: no steady-state growth, bounded peak.
+    assert end_current - setup_current < 1_000_000, (
+        f"steady-state growth {(end_current - setup_current) / 1e6:.2f} MB"
+    )
+    assert peak < 10_000_000, f"traced peak {peak / 1e6:.2f} MB"
+
+    # Latency: the last stretch must not drift from the first.
+    head = sum(timings[:200]) / 200
+    tail = sum(timings[-200:]) / 200
+    assert tail < 3 * head + 0.005, f"drift: first {head*1e3:.2f}ms, last {tail*1e3:.2f}ms"
+
+    # Retention: observations pruned to the cap across the whole run.
+    observed = activity.query(task_id=task.id, kinds=("check-observed",), limit=1000)
+    assert len(observed) <= OBSERVATION_RETENTION, len(observed)
+
+    # DB growth bounded for a simulated fortnight of full-kit watching.
+    size = (home / "nanodot.db").stat().st_size
+    assert size < 10_000_000, f"db {size / 1e6:.2f} MB"
+
+    # Functional sanity: all three kinds fired; the watch stayed healthy.
+    kinds = [e.kind for e in sink.events]
+    simulated_days = TICKS * CADENCE / DAY
+    assert kinds.count(DIGEST) >= simulated_days - 2, kinds.count(DIGEST)
+    assert FLAKY in kinds and STALE in kinds
+    final = store.get(task.id)
+    assert final.state.value == "active" and final.next_check_at is not None
+    store.close()
+    activity.close()