Side-by-side ruled tables still share projected text lines

[Chengyunlai]

Summary

After side-by-side ruled grids are identified as independent geometry
components, same-y text from adjacent table rectangles can still be grouped
into shared projected lines. Each ruled candidate then sees text from the
other table as overhang, is rejected, and fallback emits a merged table.

The completed fix is now included directly in #392 at 26b79a6.

Public reproduction

• GitHub preview / download
• Raw PDF
• SHA-256: e9b56e7a3f1d990d9e1f1b63b46e5d89b8ce1afaa85c9e17a09d83d1f5563b8d

The PDF is a self-authored, synthetic one-page Word-to-PDF document. It has no
customer, personal, credential, internal-project, or business-specific data.

Command

lit parse /input.pdf \
  --no-ocr \
  --format markdown \
  --output /results/output.md \
  --quiet

Observed pipeline before the fix

At the previous #392 head (9d740a4), geometry is already separate but text
projection is not:

two independent vector grids
-> two ruled components
-> shared cross-table projected text lines
-> ruled candidates rejected for adjacent-table overhang
-> fallback emits one merged 6-column table

Expected output

## Release channels

| Channel | Status | Owner |
|---|---|---|
| Stable | Ready | Team A |
| Beta | Testing | Team B |
| Nightly | Active | Team C |

## Support windows

| Region | Window | Contact |
|---|---|---|
| East | Morning | Desk 1 |
| West | Afternoon | Desk 2 |
| Central | Evening | Desk 3 |

Fix in #392

The updated projection path receives the detected ruled-table rectangles and
uses them as structural boundaries:

• same-y items owned by different table rectangles cannot become one line;
• nearby labels can belong to exactly one table;
• page-spanning labels remain unowned;
• table-owned lines are ordered left-to-right within a side-by-side band and
  top-to-bottom across bands;
• ordinary non-table line grouping keeps its existing y tolerance.

A/B result

Version	Geometry	Public Markdown
main@2fd644a	one merged component	one merged 7-column table
previous #392 head 9d740a4	two components	one merged 6-column fallback table
current #392 head 26b79a6	two components	two headings and two independent 3-column tables

Regression coverage

#392 now covers the published side-by-side shape plus:

• one wide ruled table;
• vertically stacked ruled tables;
• ordinary two-column prose;
• one spanning heading above two tables;
• multiple table bands and a non-transitive three-table overlap.

This issue should close when #392 is merged. The separate same-y Markdown
heading-continuation boundary found during adjacent testing is tracked in
#418.

[Chengyunlai]

Additional local comparison (fixture not uploaded):

Version	Geometry components	Final Markdown
LiteParse 2.10.1	1 merged component	1 merged 7-column table with an empty gutter column
PR #392	2 independent components	1 merged 6-column table, because projected lines still span both regions
current main	same geometry behavior as baseline for this sample	1 merged 7-column table

With a local follow-up prototype that keeps projected lines from spanning independent table rectangles, the same PDF produces two independent Markdown tables (3 columns each). This is only local evidence at this point; the prototype is not being proposed as final code until regression coverage is complete.

[logan-markewich]

@Chengyunlai without a document attached, its hard to verify what this fixes or how realistic this is

[idiot499]

Superseded by the same update posted from the issue author's account: #414 (comment)

[idiot499]

Superseded by the same update posted from the issue author's account: #414 (comment)

[Chengyunlai]

Thanks — I have now published the actual reproduction document.

• GitHub preview / download
• Raw PDF
• SHA-256: e9b56e7a3f1d990d9e1f1b63b46e5d89b8ce1afaa85c9e17a09d83d1f5563b8d

This is a self-authored, synthetic one-page document. It contains two independent 3-column ruled tables placed side by side, with identical row y-coordinates and a clear horizontal gutter. It contains no customer data or internal material.

Local comparison on this exact PDF:

Version	Ruled geometry	Final Markdown
LiteParse 2.10.1	one merged component	one merged 7-column table, including an empty gutter column
PR #392	two independent components	one merged 6-column table, because projected text lines still span both rectangles
local follow-up prototype	two independent components	two independent 3-column Markdown tables

The remaining failure is downstream of the geometry split: build_projected_lines groups same-y items from the two known table rectangles into one line. Each ruled candidate then sees text overhang from the other table and is rejected; fallback detection merges them.

I have a focused regression test and a small local fix that uses the detected table rectangles as structural line-grouping boundaries. The full liteparse library test suite passes locally (295 tests). I will link the follow-up draft PR here once the code branch is published.

The PDF lives on a dedicated reproduction-only branch so it is downloadable without mixing the binary fixture into the code change.

[Chengyunlai]

The focused follow-up is now available as a stacked draft PR:

• Chengyunlai/liteparse#2 — fix: keep projected lines within ruled tables

It is based directly on the head branch of #392, so its diff contains only the downstream projection fix and regression test. I kept it as a draft in the fork because it depends on #392; it can be retargeted/rebased once the geometry change is resolved.

Validation on the published PDF:

• focused regression: 1 passed
• full liteparse library suite: 295 passed
• full workspace release build: passed
• end-to-end Markdown output: two independent 3-column tables

[Chengyunlai]

Update: the completed projection fix and expanded regressions are now folded
directly into #392 at 26b79a6.

On the published PDF (e9b56e...) and the public lit parse --no-ocr --format markdown path, the updated PR now emits two independent headings and two
independent 3-column tables. The PR body includes the baseline / previous-head
/ current-head A/B, the two-stage root cause, exact output, and validation.

The earlier stacked draft PR in the fork is therefore superseded. I am leaving
this issue open until #392 is merged; #392 now includes Fixes #414.

The adjacent same-y heading-continuation boundary discovered during negative
testing is tracked separately in #418 with a public 3-row x 2-column PDF.