test(decay): deflake reinforced-memory guard against coarse clocks (#190)

The test simulated reinforcement by setting last_access = now_ts(), then
asserted the next decay pass leaves stability bit-identical (1e-9). On
hosts with coarse clock resolution (observed on Windows: consecutive
time.time() calls can return the same value), last_access can equal the
just-written last_decay anchor, so the strict 'last_access > anchor'
guard misses and the pass legitimately decays one tick's interval
(~2.1e-9 stability drift) — a race, not the compounding regression this
test pins.

Set last_access one second past the pass anchor so the guard condition
holds deterministically on every host, and extend coverage to 50 rapid
passes to pin the guard, not just the first one after the update.

Reproduced: full-suite runs failed this test intermittently (stability
drift 2.13e-9 vs 1e-9 threshold) while it passed in isolation; 20/20
same-tick stress trials now stable across 8 consecutive test runs.

Author: Coding-Dev-Tools

tests/test_decay_idempotent.py

@@ -67,10 +67,18 @@ def test_reinforced_memory_is_not_decayed_that_interval(monkeypatch, tmp_path):
     decayed = _stability("ns", "hot")
     assert decayed < 5.0
     # Simulate reinforcement: the memory is accessed now (last_access moves past the anchor).
+    # Bump one second past the pass's anchor: time.time() can return the SAME
+    # value as the anchor on coarse-clock hosts (observed on Windows), and a
+    # last_access equal to the anchor makes the subsequent pass legitimately
+    # decay one tick's interval — a race, not the regression this pins.
     conn = get_conn()
     conn.execute("UPDATE memories SET last_access=? WHERE namespace=? AND document_id=?",
-                 (now_ts(), "ns", "hot"))
+                 (now + 1.0, "ns", "hot"))
     conn.commit()
     # A subsequent pass must NOT decay it further — it was just accessed.
     mem_store.apply_decay_to_all("ns", 7.0)
     assert abs(_stability("ns", "hot") - decayed) < 1e-9
+    # And the guard must hold across many rapid passes, not just one.
+    for _ in range(50):
+        mem_store.apply_decay_to_all("ns", 7.0)
+    assert abs(_stability("ns", "hot") - decayed) < 1e-9