Let a jsonl export keep a record a value

`--format jsonl` existed, but `_cell` ran json.dumps at row-build time,
before a format was chosen — so a nested value was a string by then and
jsonl only re-escaped it, leaving a consumer against csv's 128KB field
limit either way. Stringifying moved to the csv writer, so csv is
byte-identical and jsonl carries json. The runs TUI followed, or a record
would draw as a Python repr.

backend/tests/api/routes/test_runs.py

@@ -4,6 +4,7 @@ The pipeline half — what a hit restores and what a key is made of — is in
 `tests/flow/test_runs.py`, which runs without one.
 """
 
+import csv
 import json
 from datetime import UTC, datetime, timedelta
 
@@ -954,6 +955,15 @@ def test_an_exported_run_row_carries_every_recorded_input(
         "metric.test_metrics.known.perfect"
     )
 
+    # `final_metrics` alone, short of `.train_loss`, is a record rather than a
+    # leaf — the shape a large nested blob is. jsonl keeps it a json value;
+    # csv still has nowhere to put it but a quoted string, unchanged.
+    nested = _lines(export(format="jsonl", metrics="final_metrics"))
+    assert nested[0]["metric.final_metrics"] == {"train_loss": 0.5}
+
+    rows = list(csv.reader(export(metrics="final_metrics").text.splitlines()))
+    assert rows[1][-1] == '{"train_loss": 0.5}'
+
 
 @pytest.fixture
 def paged():