`ValidationColumnConfig` misaligns rows and discards the whole batch after an upstream row is dropped

まず data_designer/engine/column_generators/generators/validation.py、特に ValidationColumnGenerator.generate() を確認し、その結果インデックスを data_designer/engine/dataset_builders/async_scheduler.py:2294-2296 から渡される残存行ラベルと比較してください。添付の repro_validation_index.py を使い、最後の行ではない --fail-row を指定して再現してください。preview() が残存行を answer_check の結果とともに返し、バッチサイズエラーが発生しなければ完了です。

Priority Level

Medium (Annoying but has workaround)

Describe the bug

If a column that runs before a ValidationColumnConfig fails non-retryably for one row (e.g. the LLM
endpoint returns a 400), the validation column returns more rows than the batch contains. The
scheduler then rejects the validation result and the entire batch is lost, not just the failed row.

[WARNING] Non-retryable failure on answer[rg=0, row=1]
[WARNING] Non-retryable failure on answer_check[rg=0, row=None]: Batch generator for column
'answer_check' returned 4 rows but 3 were expected (rg=0).

With 4 records and one failed LLM call, preview() produces 0 records and raises
DataDesignerGenerationError instead of returning the 3 surviving rows.

It only happens when the dropped row is not the last one in the row group. Dropping the last row
leaves the index contiguous (0, 1, 2), so the failure looks intermittent when real API errors hit
random rows.

Root cause. The async scheduler builds the batch input with the surviving rows' original labels:

data_designer/engine/dataset_builders/async_scheduler.py:2294-2296

active_row_indices = [ri for ri in range(rg_size) if ri not in pre_dropped]
...
input_df = lazy.pd.DataFrame(records, index=active_row_indices)

ValidationColumnGenerator.generate() builds its results with a fresh RangeIndex and joins
column-wise on labels:

data_designer/engine/column_generators/generators/validation.py:102-103

validation_results = lazy.pd.DataFrame({self.config.name: outputs_as_dicts})
return lazy.pd.concat([data, validation_results], axis=1)

With data.index == [0, 2, 3] and validation_results.index == [0, 1, 2], the outer join yields
4 rows (labels 0–3), and the misaligned rows hold NaN in the data columns or in the validation column.

Steps/Code to reproduce bug

repro_validation_index.py
The attached script uses a local mock OpenAI-compatible endpoint (stdlib http.server) that returns
HTTP 400 for one chosen row. Real Data Designer handles everything else: the seed dataset, the LLM
column, the async scheduler and the validation column. No credentials or network access are needed.

uv run --script repro_validation_index.py                # row 1 fails  -> BUG REPRODUCED (0 rows)
uv run --script repro_validation_index.py --fail-row 0   # row 0 fails  -> BUG REPRODUCED (0 rows)
uv run --script repro_validation_index.py --fail-row 3   # last row     -> PASS (3 rows)
uv run --script repro_validation_index.py --fail-row -1  # no failure   -> PASS (4 rows)

Pipeline: DataFrameSeedSource(item_id 0..3) → LLMTextColumnConfig → ValidationColumnConfig
(LOCAL_CALLABLE).

Results (same for every version tested):

data-designer	fail row 0	fail row 1	fail row 3 (last)	no failure
0.9.5	bug	bug	pass	pass
0.9.6 (latest)	bug	bug	pass	pass
1.0.0rc1	bug	bug	pass	pass

Tested with Python 3.12, pandas 3.0.6, Linux. Pin a version with
uv run --with data-designer==<version> --script repro_validation_index.py.

Expected behavior

The dropped row is excluded and the validation column is generated for the surviving rows. In the
repro, preview() returns 3 rows, each with an answer_check result.

Agent Diagnostic / Prior Investigation

This bug was found while building a multi-turn text-to-SQL trajectory pipeline on Data Designer, when
rate-limit errors on single rows wiped out whole batches. An AI coding agent helped write the issue and
the repro; I reviewed both and ran the repro myself.

• Searched existing issues and PRs (ValidationColumnConfig, "rows but … were expected", validation
  index/reset_index). Found nothing related.
• Diffed the installed 0.9.5 engine against 0.9.6 and 1.0.0rc1. validation.py and
  async_scheduler.py are unchanged.
• Confirmed the fix direction by monkeypatching ValidationColumnGenerator.generate to call the
  original with data.reset_index(drop=True) and then restore data.index on the result. With that
  patch, the repro passes on 0.9.6 and 1.0.0rc1 (3 rows, all validated).

Suggested fix (in ValidationColumnGenerator.generate):

validation_results = lazy.pd.DataFrame({self.config.name: outputs_as_dicts}, index=data.index)
return lazy.pd.concat([data, validation_results], axis=1)

Other full-column generators that rebuild a frame and concat(axis=1) may need the same check.

Additional context

Workaround: do the check in a cell-by-cell CustomColumnConfig (@dd.custom_column_generator)
instead of ValidationColumnConfig. It is row-safe.

Checklist

• I reproduced this issue or provided a minimal example
• I searched the docs/issues myself, or had my agent do so
• If I used an agent, I included its diagnostics above