DocsGPT/tests/test_message_journal.py

"""Unit tests for ``application/streaming/message_journal.py``.

The journal hook is best-effort by contract — its failure modes are
the most important thing to lock down so a streaming hiccup never
crashes ``complete_stream``.
"""

from __future__ import annotations

import json
from unittest.mock import MagicMock, patch

import pytest

from application.streaming.message_journal import (
    BatchedJournalWriter,
    record_event,
)


@pytest.mark.unit
class TestRecordEvent:
    def test_invalid_args_return_false(self):
        assert record_event("", 0, "answer") is False
        assert record_event("msg-1", 0, "") is False

    def test_happy_path_writes_and_publishes(self):
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            result = record_event(
                "msg-1", 5, "answer", {"type": "answer", "answer": "ok"}
            )

            assert result is True
            mock_repo.record.assert_called_once_with(
                "msg-1", 5, "answer", {"type": "answer", "answer": "ok"}
            )
            mock_topic_cls.assert_called_once_with("channel:msg-1")
            mock_topic.publish.assert_called_once()
            wire = mock_topic.publish.call_args[0][0]
            envelope = json.loads(wire)
            assert envelope["sequence_no"] == 5
            assert envelope["event_type"] == "answer"
            assert envelope["payload"] == {"type": "answer", "answer": "ok"}

    def test_publish_attempted_even_when_journal_fails(self):
        """A DB hiccup must not stop the live tail — currently-attached
        subscribers should still receive the live event so their UI is
        live even if a future reconnect's snapshot is missing this row.
        """
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ), patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.side_effect = RuntimeError("pg down")
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            result = record_event(
                "msg-1", 1, "answer", {"type": "answer", "answer": "x"}
            )

            assert result is False
            mock_topic.publish.assert_called_once()

    def test_publish_failure_does_not_raise(self):
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo_cls.return_value.record = MagicMock()
            mock_topic = MagicMock()
            mock_topic.publish.side_effect = RuntimeError("redis down")
            mock_topic_cls.return_value = mock_topic

            # Must not raise.
            result = record_event("msg-1", 0, "answer", {"answer": "y"})
            assert result is True  # Journal still committed.

    def test_payload_none_treated_as_empty_dict(self):
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            record_event("msg-1", 0, "end", None)
            mock_repo.record.assert_called_once_with("msg-1", 0, "end", {})

    def test_payload_non_dict_rejected_at_gate(self):
        """Contract: payload must be a dict (or None). Lists, strings,
        ints, and other shapes are rejected without writing or
        publishing.

        Background: the live path (``base.py::_emit``) and the replay
        path (``event_replay``) previously reconstructed non-dicts
        differently — ``{"value": payload}`` live vs.
        ``{"type": event_type}`` on replay — so a reconnecting client
        would receive a different envelope than the one originally
        streamed. Rejecting at this gate keeps the two paths
        byte-identical.
        """
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            result = record_event("msg-1", 0, "end", ["unexpected", "list"])

            assert result is False
            mock_repo.record.assert_not_called()
            mock_topic.publish.assert_not_called()

    def test_integrity_error_retries_with_seq_plus_one(self):
        """Composite-PK collision on (message_id, sequence_no) is
        recovered by one retry against ``latest_sequence_no + 1`` —
        the most likely cause is a stale seq seed on a continuation
        retry, where the route read MAX(seq) from a separate
        connection before another writer committed past it.

        On success the live pubsub publish uses the retried seq so
        the journal row and the live frame agree, even though the
        caller's original POST stream still carries the pre-retry id.
        """
        from sqlalchemy.exc import IntegrityError

        # Two repo instances: one per ``with db_session()`` block.
        # repo_first.record raises IntegrityError; the readonly
        # session returns latest=7; repo_retry.record succeeds at 8.
        repo_first = MagicMock(name="repo_first")
        repo_first.record.side_effect = IntegrityError("stmt", {}, Exception())
        repo_readonly = MagicMock(name="repo_readonly")
        repo_readonly.latest_sequence_no.return_value = 7
        repo_retry = MagicMock(name="repo_retry")

        repo_instances = iter([repo_first, repo_readonly, repo_retry])

        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.db_readonly"
        ) as mock_readonly, patch(
            "application.streaming.message_journal.MessageEventsRepository",
            side_effect=lambda conn: next(repo_instances),
        ), patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_readonly.return_value.__enter__.return_value = MagicMock()
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            result = record_event("msg-1", 3, "answer", {"text": "hi"})

            assert result is True
            # First INSERT attempted at the caller's seq=3.
            repo_first.record.assert_called_once_with(
                "msg-1", 3, "answer", {"text": "hi"}
            )
            # Latest probed for the retry seed.
            repo_readonly.latest_sequence_no.assert_called_once_with("msg-1")
            # Retry INSERT at latest+1 = 8.
            repo_retry.record.assert_called_once_with(
                "msg-1", 8, "answer", {"text": "hi"}
            )
            # Live publish uses the materialised seq so the wire and
            # the journal row stay in lockstep on the retry path.
            mock_topic.publish.assert_called_once()
            wire = json.loads(mock_topic.publish.call_args[0][0])
            assert wire["sequence_no"] == 8

    def test_integrity_error_retry_failure_drops_silently(self):
        """If the retry collides again (truly concurrent writers in
        lockstep) the journal write is dropped but the function still
        returns ``False`` without raising — the streaming loop must
        not be killed by a journal hiccup.
        """
        from sqlalchemy.exc import IntegrityError

        repo_first = MagicMock(name="repo_first")
        repo_first.record.side_effect = IntegrityError("stmt", {}, Exception())
        repo_readonly = MagicMock(name="repo_readonly")
        repo_readonly.latest_sequence_no.return_value = 3
        repo_retry = MagicMock(name="repo_retry")
        repo_retry.record.side_effect = IntegrityError("stmt", {}, Exception())

        repo_instances = iter([repo_first, repo_readonly, repo_retry])

        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.db_readonly"
        ) as mock_readonly, patch(
            "application.streaming.message_journal.MessageEventsRepository",
            side_effect=lambda conn: next(repo_instances),
        ), patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_readonly.return_value.__enter__.return_value = MagicMock()
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            result = record_event("msg-1", 0, "answer", {"text": "hi"})

            assert result is False
            # Both INSERT attempts fired; the second raised too.
            assert repo_first.record.call_count == 1
            assert repo_retry.record.call_count == 1
            # Live publish still fires on the materialised seq path —
            # subscribers downgrade to keepalives if they were waiting
            # on this event, which is correct since the journal can't
            # serve it on a future reconnect anyway.
            mock_topic.publish.assert_called_once()

    def test_payload_with_null_byte_stripped_before_insert(self):
        """``\\x00`` in a string value would crash JSONB INSERT — strip
        at the journal boundary so the row lands and reconnecting
        clients see the chunk.
        """
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            record_event(
                "msg-1", 0, "answer", {"answer": "hel\x00lo", "type": "answer"}
            )

            mock_repo.record.assert_called_once_with(
                "msg-1", 0, "answer", {"answer": "hello", "type": "answer"}
            )

    def test_payload_with_nested_null_bytes_stripped_recursively(self):
        """Strip walks nested dicts and lists — and string keys, to be
        safe against an LLM emitting structured data with NULs.
        """
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            record_event(
                "msg-1",
                0,
                "answer",
                {
                    "outer": {"inner\x00key": "a\x00b"},
                    "items": ["x\x00y", {"k": "v\x00"}],
                    "clean": "ok",
                },
            )

            mock_repo.record.assert_called_once_with(
                "msg-1",
                0,
                "answer",
                {
                    "outer": {"innerkey": "ab"},
                    "items": ["xy", {"k": "v"}],
                    "clean": "ok",
                },
            )

    def test_payload_without_null_bytes_unchanged(self):
        """Common-case: no NUL, no allocation churn — payload passed
        through identically.
        """
        with patch(
            "application.streaming.message_journal.db_session"
        ) as mock_session, patch(
            "application.streaming.message_journal.MessageEventsRepository"
        ) as mock_repo_cls, patch(
            "application.streaming.message_journal.Topic"
        ) as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            payload = {"a": "1", "nested": {"b": [1, 2, "three"]}}
            record_event("msg-1", 0, "answer", payload)

            mock_repo.record.assert_called_once_with(
                "msg-1", 0, "answer", payload
            )


@pytest.mark.unit
class TestBatchedJournalWriter:
    """``BatchedJournalWriter`` amortizes per-emit PG commits.

    Tests cover the four flush triggers (size, time, lifecycle,
    explicit), the bulk→per-row IntegrityError fallback, and the
    live-publish-stays-synchronous invariant.
    """

    def _patch_io(self):
        """Standard patch set: stub PG sessions and Redis pubsub.

        Returns a tuple ``(session_cm, readonly_cm, repo_factory, topic)``.
        The repo_factory is a list — each ``MessageEventsRepository(...)``
        call inside the writer pops one element off the front. Tests
        push fakes onto it before triggering the path that needs them.
        """
        from unittest.mock import patch as _patch

        session_cm = _patch("application.streaming.message_journal.db_session")
        readonly_cm = _patch(
            "application.streaming.message_journal.db_readonly"
        )
        repo_cls = _patch(
            "application.streaming.message_journal.MessageEventsRepository"
        )
        topic_cls = _patch("application.streaming.message_journal.Topic")
        return session_cm, readonly_cm, repo_cls, topic_cls

    def test_size_trigger_flushes_at_batch_size(self):
        """Buffer reaches ``batch_size`` → one bulk_record call covers
        all rows; publishes fire after the bulk commit, one per row.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter(
                "msg-1", batch_size=3, batch_interval_ms=10_000
            )
            # First two records: buffered, no bulk_record yet, no publish.
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "answer", {"text": "b"})
            mock_repo.bulk_record.assert_not_called()
            assert mock_topic.publish.call_count == 0
            # Third record: triggers a size-based flush.
            writer.record(2, "answer", {"text": "c"})
            mock_repo.bulk_record.assert_called_once()
            args = mock_repo.bulk_record.call_args
            assert args.args[0] == "msg-1"
            buffered = args.args[1]
            assert len(buffered) == 3
            assert [seq for seq, _, _ in buffered] == [0, 1, 2]
            # Publishes fire after the bulk commit, one per row in order.
            assert mock_topic.publish.call_count == 3
            published_seqs = [
                json.loads(call.args[0])["sequence_no"]
                for call in mock_topic.publish.call_args_list
            ]
            assert published_seqs == [0, 1, 2]

    def test_time_trigger_flushes_after_interval(self):
        """When the elapsed time since the last flush exceeds
        ``batch_interval_ms``, the next ``record()`` flushes — even if
        the buffer is well below ``batch_size``. Drives reconnect
        visibility for slow producers.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls, patch(
            "application.streaming.message_journal.time.monotonic"
        ) as mock_mono:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            # First call from __init__ at t=0; subsequent calls drive
            # the time-trigger check inside ``_should_flush``.
            #   t=0.000  → __init__ snapshots last_flush
            #   t=0.001  → record(0) — 1ms elapsed → no flush
            #   t=0.001  → after record, last_flush snapshot when flush called (n/a)
            #   t=0.200  → record(1) — 200ms elapsed → flush
            #   t=0.200  → flush updates last_flush
            mock_mono.side_effect = iter([0.000, 0.001, 0.200, 0.200])

            writer = BatchedJournalWriter(
                "msg-1", batch_size=100, batch_interval_ms=100
            )
            writer.record(0, "answer", {"text": "a"})
            mock_repo.bulk_record.assert_not_called()
            writer.record(1, "answer", {"text": "b"})
            mock_repo.bulk_record.assert_called_once()

    def test_close_drains_remaining_buffer(self):
        """``close()`` is the lifecycle flush — at end of stream every
        buffered event must commit before the writer goes silent.
        Idempotent so it's safe in multiple finally clauses.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            writer = BatchedJournalWriter(
                "msg-1", batch_size=100, batch_interval_ms=100_000
            )
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "end", {"type": "end"})
            mock_repo.bulk_record.assert_not_called()

            writer.close()
            mock_repo.bulk_record.assert_called_once()
            assert len(mock_repo.bulk_record.call_args.args[1]) == 2

            # Second close is a no-op.
            writer.close()
            assert mock_repo.bulk_record.call_count == 1

    def test_record_after_close_returns_false(self):
        """Writing past ``close()`` must not silently land in a stale
        buffer — return False so the caller's flow surfaces the bug.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo_cls.return_value = MagicMock()
            mock_topic_cls.return_value = MagicMock()

            writer = BatchedJournalWriter("msg-1", batch_size=100)
            writer.close()
            assert writer.record(0, "answer", {"text": "a"}) is False

    def test_record_rejects_non_dict_payload(self):
        """Same contract as ``record_event``: non-dict payloads are
        rejected to keep live and replay paths byte-identical.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo_cls.return_value = MagicMock()
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter("msg-1")
            assert writer.record(0, "answer", ["bad", "list"]) is False
            # Nothing publishes either — the gate fires before the wire.
            mock_topic.publish.assert_not_called()

    def test_bulk_collision_falls_back_to_per_row(self):
        """Bulk INSERT fails with IntegrityError → writer retries each
        row individually via the legacy ``record_event`` retry path,
        so a single colliding seq doesn't drop the whole batch.
        """
        from sqlalchemy.exc import IntegrityError

        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm as mock_readonly, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_readonly.return_value.__enter__.return_value = MagicMock()

            # First call: bulk_record raises. Subsequent calls: per-row
            # record succeeds. Use a single fake whose first
            # ``bulk_record`` call raises and whose ``record`` succeeds.
            bulk_repo = MagicMock(name="bulk_repo")
            bulk_repo.bulk_record.side_effect = IntegrityError(
                "stmt", {}, Exception()
            )
            per_row_repo = MagicMock(name="per_row_repo")
            # MessageEventsRepository(conn) is called once per session
            # opened. Bulk path = 1; per-row fallback = 2 rows × 1 each.
            mock_repo_cls.side_effect = [bulk_repo, per_row_repo, per_row_repo]

            mock_topic_cls.return_value = MagicMock()

            writer = BatchedJournalWriter("msg-1", batch_size=2)
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "answer", {"text": "b"})

            bulk_repo.bulk_record.assert_called_once()
            # Per-row fallback wrote each row in its own session.
            assert per_row_repo.record.call_count == 2
            assert per_row_repo.record.call_args_list[0].args[1] == 0
            assert per_row_repo.record.call_args_list[1].args[1] == 1

    def test_flush_clears_buffer_even_on_total_failure(self):
        """A flush that fails with a non-IntegrityError exception must
        still clear the buffer — leaving rows would grow memory
        unbounded across the remainder of the stream. Degraded UX
        (missing snapshot rows) beats runaway memory.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo.bulk_record.side_effect = RuntimeError("PG gone")
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            writer = BatchedJournalWriter("msg-1", batch_size=2)
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "answer", {"text": "b"})
            # Bulk failed, but the buffer is cleared so a subsequent
            # close() is a no-op and memory stays bounded.
            assert writer._buffer == []
            writer.close()
            # Only the one flush we forced — no double-attempt on
            # close after the buffer was already drained.
            assert mock_repo.bulk_record.call_count == 1

    def test_record_does_not_publish_synchronously(self):
        """Regression guard: ``record()`` must not publish before the
        journal INSERT commits. Buffering an event with no flush yet
        means zero pubsub frames on the wire — otherwise a reconnect
        snapshot could miss a row that subscribers already received.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo_cls.return_value = MagicMock()
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter("msg-1", batch_size=100)
            for i in range(5):
                writer.record(i, "answer", {"text": str(i)})
            mock_topic.publish.assert_not_called()

    def test_flush_publishes_each_buffered_event_after_bulk_insert_in_order(
        self,
    ):
        """Happy path: flush issues one bulk INSERT, then publishes each
        buffered frame in the order it was recorded.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter(
                "msg-1", batch_size=100, batch_interval_ms=100_000
            )
            for i in range(4):
                writer.record(i, "answer", {"text": str(i)})
            mock_topic.publish.assert_not_called()

            writer.flush()
            mock_repo.bulk_record.assert_called_once()
            assert mock_topic.publish.call_count == 4
            published_seqs = [
                json.loads(call.args[0])["sequence_no"]
                for call in mock_topic.publish.call_args_list
            ]
            assert published_seqs == [0, 1, 2, 3]

    def test_flush_bulk_exception_drops_both_row_and_publish(self):
        """Non-IntegrityError from bulk_record → both the journal row
        and the live publish are dropped together. Matches the
        publish_user_event "drop live publish when durable write fails"
        contract.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo.bulk_record.side_effect = RuntimeError("PG gone")
            mock_repo_cls.return_value = mock_repo
            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter("msg-1", batch_size=2)
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "answer", {"text": "b"})

            mock_repo.bulk_record.assert_called_once()
            mock_topic.publish.assert_not_called()

    def test_flush_per_row_publishes_on_commit_skips_on_persistent_drop(self):
        """IntegrityError-after-retry path: per-row fallback publishes
        each row that lands in PG (including those rewritten at
        latest+1), and skips publish for any row dropped after the
        retry collision.
        """
        from sqlalchemy.exc import IntegrityError

        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm as mock_readonly, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_readonly.return_value.__enter__.return_value = MagicMock()

            # Bulk attempt collides. Per-row: row 0 commits at its seq;
            # row 1 collides, readonly probe returns latest=9, retry at
            # 10 collides too → dropped without publish.
            bulk_repo = MagicMock(name="bulk_repo")
            bulk_repo.bulk_record.side_effect = IntegrityError(
                "stmt", {}, Exception()
            )
            row0_repo = MagicMock(name="row0_repo")
            row1_first_repo = MagicMock(name="row1_first_repo")
            row1_first_repo.record.side_effect = IntegrityError(
                "stmt", {}, Exception()
            )
            row1_readonly_repo = MagicMock(name="row1_readonly_repo")
            row1_readonly_repo.latest_sequence_no.return_value = 9
            row1_retry_repo = MagicMock(name="row1_retry_repo")
            row1_retry_repo.record.side_effect = IntegrityError(
                "stmt", {}, Exception()
            )
            mock_repo_cls.side_effect = [
                bulk_repo,
                row0_repo,
                row1_first_repo,
                row1_readonly_repo,
                row1_retry_repo,
            ]

            mock_topic = MagicMock()
            mock_topic_cls.return_value = mock_topic

            writer = BatchedJournalWriter("msg-1", batch_size=2)
            writer.record(0, "answer", {"text": "a"})
            writer.record(1, "answer", {"text": "b"})

            # Only row 0 publishes — row 1 was dropped after the retry
            # collision, so no live frame reaches the wire for it.
            assert mock_topic.publish.call_count == 1
            wire = json.loads(mock_topic.publish.call_args[0][0])
            assert wire["sequence_no"] == 0
            assert wire["payload"] == {"text": "a"}

    def test_record_strips_null_bytes_before_buffer(self):
        """Null-byte sanitization applies at the writer boundary too —
        otherwise a NUL-bearing chunk would land in the buffer, the
        bulk INSERT would fail with ``DataError``, and the per-row
        fallback would re-attempt the same broken row. Strip at the
        gate so the buffered tuple is JSONB-safe.
        """
        session_cm, readonly_cm, repo_cls_p, topic_cls_p = self._patch_io()
        with session_cm as mock_session, readonly_cm, repo_cls_p as mock_repo_cls, topic_cls_p as mock_topic_cls:
            mock_session.return_value.__enter__.return_value = MagicMock()
            mock_repo = MagicMock()
            mock_repo_cls.return_value = mock_repo
            mock_topic_cls.return_value = MagicMock()

            writer = BatchedJournalWriter("msg-1", batch_size=2)
            writer.record(0, "answer", {"answer": "hel\x00lo"})
            writer.record(
                1, "answer", {"nested": {"k\x00": ["a\x00b", "ok"]}}
            )

            mock_repo.bulk_record.assert_called_once()
            buffered = mock_repo.bulk_record.call_args.args[1]
            assert buffered[0][2] == {"answer": "hello"}
            assert buffered[1][2] == {"nested": {"k": ["ab", "ok"]}}