diff --git a/packages/pickled-core/src/pickled_core/mine/features_stage.py b/packages/pickled-core/src/pickled_core/mine/features_stage.py
index 49f1ab1..1354622 100644
--- a/packages/pickled-core/src/pickled_core/mine/features_stage.py
+++ b/packages/pickled-core/src/pickled_core/mine/features_stage.py
@@ -47,7 +47,7 @@ def _filter_story_paths(
     return filtered
 
 
-async def _draft_one(
+def _draft_one_sync(
     story_path: Path,
     feature_path: Path,
     llm: LLMClient,
@@ -55,6 +55,15 @@ async def _draft_one(
     overwrite: bool,
     interactive: bool,
 ) -> FeatureResult:
+    """Draft one feature synchronously (used by both serial and parallel paths).
+
+    This function performs blocking I/O (LLM HTTP call + file writes). The
+    parallel path schedules it on a worker thread via :func:`asyncio.to_thread`
+    so that ``--max-parallel`` actually overlaps LLM calls. Calling the LLM
+    directly inside an ``async def`` would block the event loop and serialise
+    every draft, defeating the flag (and producing N×slower mining runs for
+    workspaces with many surfaces).
+    """
     surface_id = _story_surface_id(story_path)
     if feature_path.is_file() and not overwrite:
         return FeatureResult(
@@ -122,7 +131,8 @@ async def _run_parallel(
     async def _one(story_path: Path, feature_path: Path) -> None:
         async with sem:
             results.append(
-                await _draft_one(
+                await asyncio.to_thread(
+                    _draft_one_sync,
                     story_path,
                     feature_path,
                     llm,
@@ -178,19 +188,16 @@ def run_features(
             _run_parallel(jobs, llm, overwrite=overwrite, max_parallel=max_parallel)
         )
     else:
-        results = []
-        for story_path, feature_path in jobs:
-            results.append(
-                asyncio.run(
-                    _draft_one(
-                        story_path,
-                        feature_path,
-                        llm,
-                        overwrite=overwrite,
-                        interactive=True,
-                    )
-                )
+        results = [
+            _draft_one_sync(
+                story_path,
+                feature_path,
+                llm,
+                overwrite=overwrite,
+                interactive=True,
             )
+            for story_path, feature_path in jobs
+        ]
 
     return FeaturesStageResult(output_dir=paths.root, results=results, skipped_entire_stage=False)
 
diff --git a/packages/pickled-core/tests/test_mine_features_stage.py b/packages/pickled-core/tests/test_mine_features_stage.py
index f2a2f7a..6ad0f12 100644
--- a/packages/pickled-core/tests/test_mine_features_stage.py
+++ b/packages/pickled-core/tests/test_mine_features_stage.py
@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import threading
+import time
 from pathlib import Path
 from typing import Any
 
@@ -92,3 +94,70 @@ class _Client:
     assert not result.skipped_entire_stage
     assert len(result.results) == 1
     assert result.results[0].surface_id == "tiny_target_greet"
+
+
+def test_max_parallel_actually_overlaps_llm_calls(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """`--max-parallel N` must let up to N LLM drafts run concurrently.
+
+    Regression for a bug where ``_draft_one`` was an ``async def`` that
+    invoked the synchronous ``FeatureDrafter.draft_from_story`` directly:
+    every coroutine blocked the event loop for the entire LLM call, so
+    the asyncio.Semaphore never released to a waiting task and the
+    advertised ``--max-parallel`` flag silently degraded to 1-way
+    serial execution. With 50 surfaces × 5s LLM calls, that turns a
+    ~63 s mining run with ``--max-parallel 8`` into ~250 s.
+    """
+    stories = tmp_path / "stories"
+    stories.mkdir()
+    for index in range(8):
+        (stories / f"surf_{index}.story.md").write_text("# story\n", encoding="utf-8")
+
+    state: dict[str, int] = {"current": 0, "peak": 0}
+    lock = threading.Lock()
+
+    class _SlowDrafter:
+        def __init__(self, llm: Any) -> None:
+            _ = llm
+
+        def draft_from_story(self, story: str) -> DraftResult:
+            _ = story
+            with lock:
+                state["current"] += 1
+                if state["current"] > state["peak"]:
+                    state["peak"] = state["current"]
+            try:
+                time.sleep(0.2)
+            finally:
+                with lock:
+                    state["current"] -= 1
+            return DraftResult(
+                text="Feature: x\n\n  Scenario: y\n    Given z\n",
+                rationale="",
+                warnings=(),
+            )
+
+    monkeypatch.setattr("pickled_core.mine.features_stage.FeatureDrafter", _SlowDrafter)
+
+    class _Client:
+        pass
+
+    started = time.time()
+    result = run_features(
+        tmp_path,
+        llm=_Client(),  # type: ignore[arg-type]
+        quick=True,
+        overwrite=True,
+        max_parallel=4,
+    )
+    elapsed = time.time() - started
+
+    assert len(result.results) == 8
+    # 4-way parallelism over 8 jobs of 0.2s each: ideal ~0.4s, accept up to 1.5s
+    # for slow CI jitter; serial would take ~1.6s and we want a clear gap.
+    assert state["peak"] >= 2, (
+        f"max-parallel=4 should overlap LLM calls; observed peak={state['peak']}"
+    )
+    assert elapsed < 1.5, f"max-parallel=4 should run in well under serial time; got {elapsed:.2f}s"