fix: serialize decision ledger writes

harness/tests/test_decision_join_set_cli.py

@@ -1,4 +1,7 @@
+import fcntl
 import json
+import multiprocessing
+import time
 
 import pytest
 from typer.testing import CliRunner
@@ -8,6 +11,16 @@ from tht.decisions import append_decision, append_decisions, list_decisions
 from tht.phase import current_phase
 
 
+def _concurrent_append_worker(session, subject, start, ready, done):
+    ready.put(subject)
+    start.wait()
+    try:
+        append_decision(session, type="concept_clarified", subject=subject)
+        done.put((subject, None))
+    except Exception as error:  # pragma: no cover - surfaced through the parent assertion
+        done.put((subject, repr(error)))
+
+
 def _walk_to_phase(session, target):
     while current_phase(session) < target:
         append_decision(session, type="phase_approved", subject=f"phase:{current_phase(session)}")
@@ -91,3 +104,41 @@ def test_atomic_replace_failure_keeps_the_original_ledger(tmp_path, monkeypatch)
         )
 
     assert (session / "review_decisions.jsonl").read_text() == original
+
+
+def test_concurrent_appends_wait_for_the_session_lock_and_keep_both_records(tmp_path):
+    session = tmp_path / "s1"
+    session.mkdir()
+    context = multiprocessing.get_context("fork")
+    start = context.Event()
+    ready = context.Queue()
+    done = context.Queue()
+    processes = [
+        context.Process(
+            target=_concurrent_append_worker,
+            args=(session, subject, start, ready, done),
+        )
+        for subject in ("first", "second")
+    ]
+    for process in processes:
+        process.start()
+    assert {ready.get(timeout=5), ready.get(timeout=5)} == {"first", "second"}
+
+    lock_path = session / ".review_decisions.lock"
+    with lock_path.open("a+") as lock:
+        fcntl.flock(lock, fcntl.LOCK_EX)
+        start.set()
+        time.sleep(0.2)
+        writers_waited = not (session / "review_decisions.jsonl").exists()
+        fcntl.flock(lock, fcntl.LOCK_UN)
+
+    results = [done.get(timeout=5), done.get(timeout=5)]
+    for process in processes:
+        process.join(timeout=5)
+        assert process.exitcode == 0
+
+    assert writers_waited
+    assert all(error is None for _subject, error in results)
+    records = list_decisions(session)
+    assert {record.subject for record in records} == {"first", "second"}
+    assert [record.seq for record in records] == [1, 2]