perf(table): make localized Polars exports native-speed

[Light2Dark]

This pull request was authored by a coding agent.

📝 Summary

This is a stacked follow-up to #10661 and targets its feature branch. The parent PR makes CSV and TSV exports locale-aware, but non-default dialects route dataframe rows through Python's csv.writer. That cost is especially visible for Polars, whose native writer is otherwise implemented in Rust.

This PR keeps pandas and Polars on their native CSV writers for ordinary numeric and text data with the common . and , decimal separators. pandas prepares only values that its native decimal option does not format. Polars falls back to the shared compatibility writer when native output would change established export text.

Complexity and compatibility tradeoff

The adapter-specific fallbacks are intentional:

• Unicode decimal separators such as ٫ use the shared compatibility writer.
• Polars versions without native decimal-comma support use the compatibility writer.
• Polars temporal and binary columns use it because native Polars emits different text representations.
• Polars float columns containing NaN use it because native Polars emits an empty field instead of nan.
• pandas normalizes object-dtype numbers, NaN-bearing float columns, and exported numeric indexes before calling its native writer.

The ordinary path still performs one native dataframe serialization. pandas does not materialize or inspect its index unless the index is included in the export. The redundant download_as(locale=...) path and the private Narwhals separator/dialect dual interface were removed to offset the compatibility branches and leave one locale and dialect flow.

Benchmark

Local microbenchmark on Apple Silicon with 200,000 rows and five columns. Each case had one warm-up serialization followed by one timed serialization, using the same environment for the parent PR and this branch.

Export	Parent PR	This PR	Improvement
pandas pt-BR CSV	0.724s	0.518s	1.4×
Polars pt-BR CSV	0.641s	0.005s	128×
Polars TSV	0.285s	0.004s	71×

The repository does not have a Python benchmark harness, so unit tests avoid flaky wall-clock assertions. Deterministic routing tests verify the native writer paths, while exact-output tests cover NaN-bearing floats, numeric indexes, temporal values, binary values, nested data, Unicode decimal separators, and older-Polars capability detection.

📋 Pre-Review Checklist

• This work is a refinement of the approved locale-export work in feat(table): locale-aware CSV and TSV exports #10661.
• Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
• No visual evidence is needed; this changes backend serialization only.

✅ Merge Checklist

• I have read the contributor guidelines.
• No public documentation changes are required; internal comments describe the compatibility boundary.
• Tests cover compatibility and native-writer routing.

Written by GPT-5 on Codex

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
marimo-docs	Ready	Preview	Aug 27, 2026 4:57pm

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

All reported issues were addressed across 10 files

_{You’re at about 95% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>