diff --git a/.github/workflows/pytest.yml b/.github/workflows/pytest.yml
index a587bf8..d06c02d 100644
--- a/.github/workflows/pytest.yml
+++ b/.github/workflows/pytest.yml
@@ -96,7 +96,7 @@ jobs:
       # carries, or a change to them is not exercised until after it merges.
       # Consumers use the action; this repository owns the rules.
       - name: Refuse a skip, and a suite that shrank
-        run: tools/check-test-outcome.py "$RUNNER_TEMP/pytest.log" --min-tests 315
+        run: tools/check-test-outcome.py "$RUNNER_TEMP/pytest.log" --min-tests 317
 
       # The rules earn their place by refusing a log that carries what they
       # name. Both fixtures are written here rather than tracked, and the
diff --git a/parser/shapeinfer.py b/parser/shapeinfer.py
index 652de4b..d017a01 100644
--- a/parser/shapeinfer.py
+++ b/parser/shapeinfer.py
@@ -110,40 +110,57 @@ def _input_arrays(func: dict) -> list:
     length from.
 
     An input array is a parameter that is an array of pointers (``TYPE **``) or
-    of by-value scalars (``uint8_t *``, ``int64_t *``), immediately followed by
-    a by-value integer.  That the length is by VALUE is what tells an argument
-    apart from a written-back out-array, whose length is by POINTER — the same
-    distinction this module already reads in the other direction.
+    of by-value scalars (``uint8_t *``, ``int64_t *``), followed by a by-value
+    integer.  That the length is by VALUE is what tells an argument apart from
+    a written-back out-array, whose length is by POINTER — the same distinction
+    this module already reads in the other direction.
 
     Without it a binding matches the LENGTH PARAMETER'S NAME, and the names
     disagree: ``count``, ``size``, ``ngeoms``, ``keys_len``, ``path_len``,
     ``pixels_size``, ``wkb_size``, ``count1``.  Every one of them is a length,
     and a binding that knows only some of them silently drops the rest.
+
+    A RUN of arrays shares the one length that follows it.  Arrays read in
+    parallel are declared together and counted once — ``jsonb_make_two_arg(text
+    **keys, text **values, int count)`` pairs the two element by element, and
+    ``tpointseq_make_coords`` reads four — so the length belongs to every array
+    of the run, not only to the one the count happens to sit beside.  Where a
+    family counts each array separately the run is one long and this says what
+    it always said: ``edwithin_tgeoarr_tgeoarr(arr1, count1, arr2, count2, …)``
+    keeps ``arr1`` on ``count1``.
     """
     params = func.get("params", [])
-    out = []
-    for i, prm in enumerate(params[:-1]):
+
+    def is_array(prm) -> bool:
         ctype = _bare(prm.get("cType"))
-        if _bare(params[i + 1].get("cType")) not in _LENGTH_TYPES:
-            continue
         if ctype.endswith("**"):
-            if ctype in ("char **", "void **"):
-                continue
-        elif not (ctype.endswith("*")
-                  and ctype[:-1].strip() in _ELEMENT_SCALARS):
+            return ctype not in ("char **", "void **")
+        return ctype.endswith("*") and ctype[:-1].strip() in _ELEMENT_SCALARS
+
+    out = []
+    start = 0
+    while start < len(params):
+        if not is_array(params[start]):
+            start += 1
             continue
-        out.append({
-            "param": prm["name"],
-            "lengthFrom": {"kind": "param", "name": params[i + 1]["name"]},
-            # The element reads as the return's does — the type with one
-            # pointer level off and no `const`, which belongs to the argument
-            # rather than to the element type a binding marshals.
-            "element": {
-                "c": _strip_one_ptr(_bare(prm.get("cType"))),
-                "canonical": _strip_one_ptr(
-                    _bare(prm.get("canonical") or prm.get("cType"))),
-            },
-        })
+        end = start
+        while end < len(params) and is_array(params[end]):
+            end += 1
+        if end < len(params) and _bare(params[end].get("cType")) in _LENGTH_TYPES:
+            for prm in params[start:end]:
+                out.append({
+                    "param": prm["name"],
+                    "lengthFrom": {"kind": "param", "name": params[end]["name"]},
+                    # The element reads as the return's does — the type with one
+                    # pointer level off and no `const`, which belongs to the
+                    # argument rather than to the element type a binding marshals.
+                    "element": {
+                        "c": _strip_one_ptr(_bare(prm.get("cType"))),
+                        "canonical": _strip_one_ptr(
+                            _bare(prm.get("canonical") or prm.get("cType"))),
+                    },
+                })
+        start = end
     return out
 
 
diff --git a/tests/test_shapeinfer.py b/tests/test_shapeinfer.py
index 9dcb837..d50d91b 100644
--- a/tests/test_shapeinfer.py
+++ b/tests/test_shapeinfer.py
@@ -162,6 +162,44 @@ def test_a_value_beside_a_number_is_not_an_array(self):
         for f in idl["functions"]:
             self.assertNotIn("inputArrays", f.get("shape", {}))
 
+    def test_a_run_of_arrays_shares_the_length_that_follows_it(self):
+        # `jsonb_make_two_arg(text **keys, text **values, int count)` pairs the
+        # two arrays element by element and counts them once, so `count` is the
+        # length of BOTH; reading only the array it sits beside drops the other.
+        idl = {"functions": [_fn(
+            "jsonb_make_two_arg", "Jsonb *",
+            [("keys", "text **"), ("values", "text **"), ("count", "int")]),
+            _fn("tpointseq_make_coords", "TSequence *",
+                [("xcoords", "const double *"), ("ycoords", "const double *"),
+                 ("zcoords", "const double *"), ("times", "const TimestampTz *"),
+                 ("count", "int"), ("srid", "int32_t")])]}
+        idl, stats = infer_shapes(idl)
+        self.assertEqual(
+            [(a["param"], a["lengthFrom"]["name"])
+             for a in idl["functions"][0]["shape"]["inputArrays"]],
+            [("keys", "count"), ("values", "count")])
+        self.assertEqual(
+            [(a["param"], a["lengthFrom"]["name"])
+             for a in idl["functions"][1]["shape"]["inputArrays"]],
+            [("xcoords", "count"), ("ycoords", "count"),
+             ("zcoords", "count"), ("times", "count")])
+        self.assertEqual(stats["inputArrays"], 6)
+
+    def test_an_array_counted_on_its_own_keeps_its_own_length(self):
+        # The counter-case the run rule must not swallow: a family that counts
+        # each array separately declares each one beside ITS length, so every
+        # run is one long and each array keeps the count it is declared with.
+        idl = {"functions": [_fn(
+            "edwithin_tgeoarr_tgeoarr", "int *",
+            [("arr1", "const Temporal **"), ("count1", "int"),
+             ("arr2", "const Temporal **"), ("count2", "int"),
+             ("dist", "double"), ("count", "int *")])]}
+        idl, _ = infer_shapes(idl)
+        self.assertEqual(
+            [(a["param"], a["lengthFrom"]["name"])
+             for a in idl["functions"][0]["shape"]["inputArrays"]],
+            [("arr1", "count1"), ("arr2", "count2")])
+
     def test_an_out_array_is_not_read_as_an_input_one(self):
         # `temporal_time_split(temp, ..., TimestampTz **bins, int *count)` writes
         # `bins` back, and its length being BY POINTER is what says so.