fix: consolidated rows carry per-cell provenance, not the first row's (26.6.9)

[ancongui]

Summary

Fixes a provenance-loss bug in the post-extraction consolidation transform: a cross-document consolidated group (e.g. a "current cap table" merged across several deeds) gave every output row the first input row's bounding box, page, confidence and judge verdict.

Root cause

_rebuild_rows (llm_transformer.py) already computed each output row's row-level pages/confidence from its cited contributors (26.6.8), but every cell still borrowed its bbox/pages/confidence from rows[0] through a single template_by_name map, and dropped each cell's judge/notes to the default uncertain/0.

Observable symptoms in GET /workflows/{id} → context.consolidated_field_groups:

• All rows of cap_table_vigente shared one identical bbox/pages/confidence fingerprint — byte-for-byte equal to the first source document's first row.
• Row-level pages (from real contributors) no longer matched the cell-level pages (from rows[0]).
• Every cell came back judge.status = uncertain, confidence = 0.

UI overlays drawn from these bboxes pointed every member at the same rectangle on the same (wrong) page.

Fix

Resolve each output row's provenance source in priority order:

1. input rows cited via _source_rows;
2. for a pure 1:1 rewrite (no row in the batch cites anything and the counts match) — the input row at the same index;
3. input rows matched by a distinctive identity token (an uncited consolidation row);
4. rows[0] as a last resort.

Within a row, each field takes its best-grounded contributor cell (one with a bbox wins over one without, then higher confidence), and judge/notes now survive the transform. Value-rewriting transforms are unchanged when no contributor is matched.

Test plan

• 5 new tests in test_llm_transformer_provenance.py (best-per-cell across contributors, single cited contributor, judge propagation, positional rewrite, identity fallback) — written failing first, then green.
• Full unit suite: 292 passed.
• ruff check / ruff format --check: clean. pyright: 0 errors.

🤖 Generated with Claude Code