renderer(tier-1): row anchors + number-format expansion + merged-cell propagation

[arnav2]

Summary

Three surgical changes to text_renderer.render_block() — each addresses a distinct way the parser/chunker was losing signal between the workbook and the chunk's render_text. Stacked on #11. Together they move the parser-quality metric by +0.6 pp on the full 912 and lift text@5 by +1.0 pp / text@5 in-scope by +0.9 pp — the first measurable parser+chunker improvement of this session.

What changed

1. Row-number anchors (r<N> prefix per data row)

The renderer already emitted a block header with the A1 range, but data rows had no row number. Downstream consumers — especially the agent on ks-backend — couldn't compute cell coordinates from chunk text. Now:

[Sheet1!A1:D10] (table)
     | A    | B   | C   | D    |
     |------|-----|-----|------|
r1   | Name | Q1  | Q2  | Q3   |
r2   | Wgt  | 100 | 150 | 200  |

Row prefix width is sized to the largest row number in the block so the grid stays aligned regardless of block depth.

2. Number-format-aware rendering

When a cell's number_format produces a meaningfully different displayed string from the raw value, we now emit BOTH:

raw value	number_format	rendered
0.06	0%	0.06 [6%]
1272	#,##0.00	1272 [1,272.00]
46022	yyyy-mm-dd	already handled (date)

Substring-match retrieval can hit either form — the question may quote either, and answer.xlsx often uses the display form even though input.xlsx keeps the raw.

Trivial diffs (1272 → "1272.00", "1272.0") are NOT expanded — they add no retrieval-relevant tokens, only noise.

3. Merged-cell value propagation

Slave cells in a merged region used to render blank (openpyxl returns None for them). Questions that referenced the cell by a slave coordinate could never match. Now:

r1   | Total | ← Total | ← Total |  (master + 2 slaves)

The merged region's visible value appears at every position it appears in Excel.

Bench on full 912 with text-embedding-3-large

Metric	Before Tier-1	After Tier-1	Δ
Parser-quality (rank IS NOT None, in-scope)	0.843	0.849	+0.006 (+4 instances surfaced)
recall_text@5	0.750	0.760	+0.010
recall_text@5_in_scope	0.838	0.847	+0.009
recall_text@3	0.746	0.755	+0.009
recall_text@1	0.638	0.641	+0.003
recall_geometric@5	0.482	0.484	+0.002 (noise)
mean parse_ms	156	184	+27 ms
Per-instance flips	—	—	6 miss→hit, 0 hit→miss on text@5

The +0.9 pp text@5_in_scope move is modest but clean — zero regressions, no parser internals reshaped, all gains come from making the chunk text more faithful to what's visually in the workbook.

Type of change

• ✨ New feature (row anchors, format expansion, merged-cell propagation)
• 🚀 Performance (slight regression: +27ms parse for the format/merge lookups)
• 🧪 Parser edge case / new regression test (7 new cases)

Test plan

• make test passes — 1079 → 1086 (+7 new)
• Full 912 SpreadsheetBench v0.1 with text-embedding-3-large: +6 instances flip miss→hit on text@5, 0 regressions
• tests/test_renderer_tier1.py covers all 3 changes + boundary cases (trivial diff suppression, no spurious markers on unmerged sheets)

Notes for reviewers

• Stacks on chunker: range-tighten clip + opt-in row-group splitter (no recall move; correctness only) #11 — review that PR first (chunker correctness fix; 0 recall move).
• No new dependencies. All changes are deterministic, no env vars, no opt-in flags. The renderer just emits more faithful chunks.
• render_text gets ~5-10% longer because of the row anchors + format-expansion brackets + merged-cell ← markers. Within the embedder's context window for any reasonable block; doesn't trigger any chunk-size cap.

🤖 Generated with Claude Code