fix(layout): restore column detection on table pages and sparse pages

[abimaelmartell]

What

Two gates in relative-valley column detection were starving real multi-column pages of their gutters, dropping them to single-column Y-sorting where the columns interleave line by line:

1. Table pages: the page_has_table guard predates item claiming — by grouping time the detected table's items have already left the flow, so a table cannot fake a gutter in the histogram the guard protects. The prose gate inside relative-valley acceptance is the real defense, and it stays. Without the fallback, the prose remainder of a table-bearing two-column page wove line by line.
2. Sparse pages: the 100-item floor guarded against shallow histogram dips, but OCR'd multi-column pages arrive as few long line-runs (a two-column French academic page is ~60 items) and never qualified. Lowered to 30; the same prose gate judges the split.

Results

• Multi-column reading order: 460 → 462 / 884; table structure tests unchanged (446/1022); all other categories unchanged.
• Regression corpus: all pass. Five documents change, four are clear recoveries — a French two-column bioethics text that was fully woven now reads perfectly, a refrigerant datasheet's intro column untangles from its sidebar, an IRS publication's three-column region largely un-weaves, an annual report's prose recovers around its tables. One decorative magazine-style page that was already woven is now differently (not worse) woven; it has no semantic baseline to arbitrate, disclosed here for the record.
• Semantic scorer on baselined movers: reading order up (0.10 → 0.18 on the datasheet), no decliners.
• Full test suite passes; local review clean.

🤖 Generated with Claude Code

---

Summary by cubic

Restores relative‑valley column detection on table-bearing and sparse pages to prevent interleaved reading order. Previously it required ≥100 items and no detected tables; now it runs with ≥30 items regardless of tables, and the XY‑cut fallback is re‑guarded on table pages.

• Remove the table guard from the relative‑valley path: table items are out of flow and the prose gate rejects table‑shaped splits.
• Lower the item floor from 100 to 30 to handle OCR’d multi‑column pages with few long line‑runs.
• Re‑guard the XY‑cut fallback on table pages because it has no prose gate, preventing unvalidated splits after a rejected valley.
• Impact: improved multi‑column reading order; table handling unaffected; benchmarks and tests unchanged.

^{Written for commit dfe64b8. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 1 file

<sub>Shadow auto-approve: would not auto-approve because issues were found.

Fix all with cubic | Re-trigger cubic</sub>

[cubic-dev-ai]

0 issues found across 1 file (changes from recent commits).

<sub>Shadow auto-approve: would require human review. Expands relative-valley column detection to table-bearing and sparse pages (removes table guard, lowers item floor 100→30). This broader heuristic rollout and threshold change require human sign-off; the claimed improvements rely on external benchmarks not visible in the diff.

Re-trigger cubic</sub>