test(renderer): pin master-parity for malformed trailing-text table rows

Round-3 gate finding: the downstream table regex lacks an end-of-line anchor, so
it accepts a prefix that the physical-line guard rejects (e.g.
'| h<br>x | n |\n|---|---| trailing').

Measured against MASTER rather than against the intermediate commit: master does
not render that input as a table either, so the stage is at parity, not
regressed. Across the edge-case matrix (indented, CRLF, trailing-space, padded
separator, blank-line delimited, mid-document) the stage is equal to master on
the two malformed inputs and strictly better on six genuine tables.

Pinned as a regression so any future change to the table guard has to make a
deliberate decision about this input instead of changing it silently.

tests/test_renderer_js_behaviour.py

@@ -1014,3 +1014,19 @@ class TestRendererGateRegressions7618:
         assert out.count("<tr>") == 3
         assert "OAuth<br>API keys<br>tokens" in out
         assert "LRU<br>60s TTL" in out
+
+    def test_row_with_trailing_text_after_closing_pipe_matches_master(self, driver_path):
+        """Pins parity with master for a malformed row that has text past the last pipe.
+
+        The downstream table regex lacks an end-of-line anchor, so it accepts a
+        prefix the physical-line guard rejects. Master does NOT render this as a
+        table either, so preserving prose here is parity, not a regression. Pinned
+        so a future table-guard change has to make a deliberate decision about it.
+        """
+        out = _render(driver_path, "| h<br>x | n |\n|---|---| trailing")
+        assert "<table>" not in out
+
+    def test_indented_and_padded_table_still_preserves_br(self, driver_path):
+        out = _render(driver_path, "  | a | b |\n  |---|---|\n  | x<br>y | z |")
+        assert "<table>" in out
+        assert "x<br>y" in out