Finish longform `filter`: section → translate → stitch (close TODO in filter.py)

[telivity-otaip]

Why

There is an explicit TODO in the filter path:

    # TODO(longform): align article inference with translator train sections —
    # section → translate each → stitch via split_translator_source_sections +
    # stitch_filter_chunks. Avoid a half-built whole-article echo path.

Today long drafts trip should_chunk_filter but the section→translate→stitch path is incomplete. Closing this makes filter useful for article-length rewrite (LinkedIn articles, newsletters) instead of echoing whole pieces.

Goal

Implement the longform filter path that:

1. Splits source into translator-compatible sections
2. Filters/translates each section with the local adapter/base path
3. Stitches with stitch_filter_chunks (or equivalent)
4. Preserves voice guards / force mode semantics

Acceptance criteria

• Article-length Contoso fixtures filter section-wise, not as one giant echo
• Stitched output has no duplicated section headers / separator chrome
• Short drafts still use the existing single-shot path
• Tests cover split + stitch + chunk threshold boundary
• TODO comment removed once done
• README / Advanced notes mention longform filter behavior

Pointers

• filter.py — should_chunk_filter, chunk helpers
• Translator split/stitch helpers (see imports/usages around SFT translator sections)
• tests/test_filter_voice.py

Difficulty

Medium — existing helpers are partially there; needs careful tests.