diff --git a/hypothesis-python/RELEASE-sample.rst b/hypothesis-python/RELEASE-sample.rst
index 373c6b801f..8e2d6d1c47 100644
--- a/hypothesis-python/RELEASE-sample.rst
+++ b/hypothesis-python/RELEASE-sample.rst
@@ -25,7 +25,10 @@ which should:
   - :func:`package.function` for link to functions, where the link text will
     be ``package.function``, or :func:`~package.function` to show ``function``.
   - :class:`package.class` for link to classes (abbreviated as above).
-  - :issue:`issue-number` for referencing issues or :pull:`pr-number` for PRs.
+  - :issue:`issue-number` for referencing issues.
+  - Similarly, :pull:`pr-number` can be used for PRs, but it's usually
+    preferred to refer to version numbers such as :ref:`version 6.98.9 <v6.98.9>,
+    as they are meaningful to end users.
   - :doc:`link text <chapter#anchor>` for documentation references.
   - `link text <https://hypothesis.readthedocs.io/en/latest/chapter.html#anchor>`__
     is the same link, for general web addresses.
diff --git a/hypothesis-python/RELEASE.rst b/hypothesis-python/RELEASE.rst
index ee6c398c14..b21de76dda 100644
--- a/hypothesis-python/RELEASE.rst
+++ b/hypothesis-python/RELEASE.rst
@@ -1,5 +1,8 @@
 RELEASE_TYPE: patch
 
 Account for time spent in garbage collection during tests, to avoid
-flaky ``DeadlineExceeded`` errors as seen in :issue:`3975`. Also fixes
-overcounting of stateful run times resulting from :issue:`3890`.
+flaky ``DeadlineExceeded`` errors as seen in :issue:`3975`.
+
+Also fixes overcounting of stateful run times,
+a minor observability bug dating to :ref:`version 6.98.9 <v6.98.9>`
+(:issue:`3890`).
diff --git a/hypothesis-python/tests/nocover/test_recursive.py b/hypothesis-python/tests/nocover/test_recursive.py
index 077da8fbb3..ab01f07d24 100644
--- a/hypothesis-python/tests/nocover/test_recursive.py
+++ b/hypothesis-python/tests/nocover/test_recursive.py
@@ -195,10 +195,11 @@ def test_gc_hooks_do_not_cause_unraisable_recursionerror():
 
     # This test is potentially flaky, because the stack usage of a function is not
     # constant. Regardless, if the test passes just once that's sufficient proof that
-    # it's not the GC (or accounting of it) that is at fault.
+    # it's not the GC (or accounting of it) that is at fault. Note, I haven't actually
+    # seen it fail/flake, but I believe it could happen in principle.
 
-    # The number of cycles sufficient to reliably trigger a GC cycle, experimentally
-    # found to be a few hundred on CPython. Multiply by 10 for safety margin.
+    # The number of cycles sufficient to reliably trigger GC, experimentally found
+    # to be a few hundred on CPython. Multiply by 10 for safety margin.
     NUM_CYCLES = 5_000
 
     def probe_depth():
@@ -251,12 +252,16 @@ def inner_test(_):
                 max_depth -= 1
             else:
                 break
-        # Note that PyPy is a bit weird, in that it raises RecursionError at
-        # (maxdepth - n) for small positive n, but not at exactly (maxdepth).
-        # In general, it is really finicky to get the details right in this
-        # test, so be careful.
-
-        # Now check that the limit is unchanged with gc enabled
+            # Note that PyPy is a bit weird, in that it raises RecursionError at
+            # (maxdepth - n) for small positive n, but not at exactly (maxdepth).
+            # In general, it is really finicky to get the details right in this
+            # test, so be careful.
+
+        # Now check that the limit is unchanged with gc enabled, and also that
+        # leaving a few frames for the callbacks does not fail.
+        if hasattr(gc, "callbacks"):  # see comment above
+            for n in range(1, 4):
+                gen_cycles_at_depth(max_depth - n, gc_disable=False)
         gen_cycles_at_depth(max_depth, gc_disable=False)
         with pytest.raises(RecursionError):
             gen_cycles_at_depth(max_depth + 1, gc_disable=False)