diff --git a/docs/modal-compatibility-matrix.md b/docs/modal-compatibility-matrix.md
index 49cf337..94165bd 100644
--- a/docs/modal-compatibility-matrix.md
+++ b/docs/modal-compatibility-matrix.md
@@ -74,9 +74,9 @@ real gap — should not stay this way) · ⬜ not present at all.
 | `@modal.enter(snap=False)` | Runs once per container at startup, before any input. `snap=True` marks pre-snapshot code (see §I memory snapshots). | 🔥 (wherever `@cls` is used) | ✅ body carried as `ir.Class.EnterBody`, actually run once by `warmd`. `snap=` kwarg itself unrecognized (falls through to generic "unmodeled arg"). | Memory-snapshot semantics (`snap=True` vs default) aren't distinguished — low-risk since calque doesn't do container snapshotting at all. | — |
 | `@modal.exit()` | Runs on container shutdown; gets a grace period on preemption specifically for cleanup. | 🟡 (paired with `@enter` wherever teardown matters) | ✅ (fixed 2026-08-07, calque#86) recognized in `visit_ClassDef`, excluded from `cls.Methods` (confirmed via a live repro: before the fix, an exit-only class was picked as the warm unit's sole method by `pickWarmUnit`'s fallback; after the fix, `run()` correctly refuses with "no mapped @cls+@enter warm unit found" instead). Teardown itself is leaked as unreproduced (`ir.Class.HasExit`). | The warm supervisor has no shutdown-hook concept — teardown logic itself still doesn't run, just no longer silently misclassified. | closed |
 | `@modal.method(is_generator=None)` | Converts an instance method into an invokable Modal Function scoped to the class. | 🔥 (wherever `@cls` is used) | ✅ | `is_generator=` kwarg not specifically recognized. | — |
-| `@modal.batched(max_batch_size=, wait_ms=)` | Dynamic input batching; all inputs/outputs must be equal-length lists; at most one batched method per class. | 🧊 (2/212 files in modal-examples; not found in independent-repo sample) | ⬜ | — | [#91](https://github.com/spore-host/calque/issues/91) |
+| `@modal.batched(max_batch_size=, wait_ms=)` | Dynamic input batching; all inputs/outputs must be equal-length lists; at most one batched method per class. | 🧊 (2/212 files in modal-examples; not found in independent-repo sample) | 🟨 recognized, not modeled — detected via trailing decorator name (same pattern as `_SERVE_DECOS`), tagged with a distinct `modal.batched` leak in `_describe_fn`; the function/method still runs, just without Modal's request-coalescing behavior. | Real batching execution (coalescing N concurrent calls into one list-valued call) is not reproduced — out of scope per this construct's rare (🧊) frequency. | [#91](https://github.com/spore-host/calque/issues/91) |
 | `@modal.concurrent(max_inputs=, target_inputs=None)` | **Replaces the deprecated `allow_concurrent_inputs=N` kwarg** (v0.73.148) — now a separate decorator, not a function kwarg. Sync functions get separate OS threads (must be thread-safe); async get coroutines on one thread. | 🟡 (a common tuning pattern in production) | ✅ (fixed 2026-08-07, calque#82) `_describe_fn` already captured every decorator's kwargs on a plain function, so `max_inputs`/`target_inputs` just needed adding to `autoscalingKwargs`. The class-level case needed a real fix: `visit_ClassDef` only read `@app.cls`'s OWN kwargs — a separate `@modal.concurrent(...)` stacked on the same class was invisible; now merged into `cls_kwargs`. | — | closed |
-| `@app.batched`/`max_batch_size=` | See `@modal.batched` above (same construct, different framing in the original audit). | 🧊 | ⬜ | — | dup of above |
+| `@app.batched`/`max_batch_size=` | See `@modal.batched` above (same construct, different framing in the original audit). | 🧊 | 🟨 see `@modal.batched` above — same trailing-name match covers this spelling too. | — | dup of above |
 | Web decorators: `@modal.fastapi_endpoint` (renamed from `@modal.web_endpoint`, v0.73.89), `@modal.asgi_app()`, `@modal.wsgi_app()`, `@modal.web_server(port)` | Long-lived, request-driven, no fixed N, autoscaling-driven termination — fundamentally different execution model from batch `.map()`. | 🟡 (a first-class use case — own top-level directory in modal-examples, ~19% of files) | 🟨 detected via `_SERVE_DECOS` (matches both old and new decorator names by trailing attribute), sets `entry_kind: "serve"`; `run.go` refuses gracefully with a leak, the long-lived server is never built (by design — see `docs/serve-architecture.md`). | Working as intended per the project's own documented scope decision. | — |
 
 ---
diff --git a/internal/parse/parse_test.go b/internal/parse/parse_test.go
index f094951..ca7055b 100644
--- a/internal/parse/parse_test.go
+++ b/internal/parse/parse_test.go
@@ -1158,6 +1158,38 @@ func TestParseRareConstructsAreTaggedNotSilent(t *testing.T) {
 	}
 }
 
+// TestParseModalBatchedDecoratorLeaks (calque#91): @modal.batched(...) had
+// ZERO recognition at all before this fix — unlike its four from_name/
+// CloudBucketMount siblings tested above (TestParseRareConstructsAreTaggedNotSilent),
+// it fell through completely unnoticed, no leak, no tag. Proves it now gets
+// the SAME distinguishable leak treatment, AND that the leak is purely
+// additive: the decorated function still resolves normally (still findable,
+// still runnable) — batching itself is out of scope, only the leak is new.
+func TestParseModalBatchedDecoratorLeaks(t *testing.T) {
+	r, args := runner(t)
+	rep := &leak.Report{}
+	script, _ := filepath.Abs("../../testdata/scripts/batched_function.py")
+
+	app, err := Parse(context.Background(), script, rep, r, args...)
+	if err != nil {
+		t.Fatalf("Parse: %v", err)
+	}
+
+	found := false
+	for _, l := range rep.Leaks {
+		if strings.Contains(l.Detail, "modal.batched") && (strings.Contains(l.Detail, "batching") || strings.Contains(l.Detail, "coalescing")) {
+			found = true
+		}
+	}
+	if !found {
+		t.Errorf("expected a leak mentioning modal.batched + batching/coalescing; leaks=%+v", rep.Leaks)
+	}
+
+	if _, ok := app.FindFunction("process"); !ok {
+		t.Error(`function "process" not found in parsed app — the @modal.batched leak must be additive, not a refusal to parse the function`)
+	}
+}
+
 // TestParseMapIterables (calque#136): a real .map()/.starmap() iterable that
 // pyast could statically resolve (a literal list, a literal list of tuples,
 // or a range()) must land in ir.Function.Items; an unresolvable one (a
diff --git a/testdata/scripts/batched_function.py b/testdata/scripts/batched_function.py
new file mode 100644
index 0000000..296cce5
--- /dev/null
+++ b/testdata/scripts/batched_function.py
@@ -0,0 +1,20 @@
+"""batched_function.py — calque#91 fixture: @modal.batched(...) stacked on a
+plain @app.function, Modal's automatic request-coalescing decorator.
+
+Before this fixture existed, @modal.batched had ZERO recognition at all —
+unlike its four from_name/CloudBucketMount siblings in rare_constructs.py,
+it fell through completely unnoticed (no leak, no tag). This fixture only
+proves the decorator is now a distinct, greppable leak (`where` ==
+"modal.batched" in helper_leaks) — batching itself is NOT modeled; `process`
+still runs, just without Modal's list-coalescing behavior.
+"""
+
+import modal
+
+app = modal.App("batched-function")
+
+
+@app.function()
+@modal.batched(max_batch_size=4, wait_ms=100)
+def process(items: list[int]) -> list[int]:
+    return [i * 2 for i in items]
diff --git a/tools/pyast/pyast.py b/tools/pyast/pyast.py
index b6bfded..8af71a5 100755
--- a/tools/pyast/pyast.py
+++ b/tools/pyast/pyast.py
@@ -651,6 +651,28 @@ def _describe_fn(
                 "lineno": getattr(d, "lineno", node.lineno),
             }
         )
+    # calque#91: @modal.batched(...) requests automatic request-coalescing —
+    # it's a MODIFIER stacked on top of whatever shape the function already
+    # has (a plain @app.function, an @app.cls method, etc.), not a new
+    # entry_kind value (unlike _SERVE_DECOS/_entry_kind above). Before this,
+    # it fell through completely unnoticed — zero leak at all, unlike its
+    # four from_name/CloudBucketMount siblings above, which already get a
+    # distinguishable "where" tag. Matched by trailing decorator name, same
+    # style as the @app.function check in _try_expand_decorated_loop.
+    if leaks is not None:
+        for d in node.decorator_list:
+            if _decorator_name(d).rsplit(".", 1)[-1] == "batched":
+                leaks.append(
+                    {
+                        "where": "modal.batched",
+                        "detail": (
+                            "@modal.batched(...) requests automatic request-coalescing/batching of concurrent "
+                            "calls into one call with list-valued args; calque does not reproduce this -- the "
+                            "function still runs, just without Modal's batching behavior"
+                        ),
+                        "lineno": getattr(d, "lineno", node.lineno),
+                    }
+                )
     return {
         "name": node.name,
         "lineno": node.lineno,