PDF: reuse table grids and lane evidence

[myhloli]

Summary

• Reuse frozen rule-grid components across deferred rule candidates, closed-grid detection, and owned merging.
• Reduce repeated marker-context scanning, lane coverage sorting, single-lane short-tail validation, and per-cell glyph sorting.
• Add an exact-type fast path for owned Bbox normalization with unchanged fallback semantics.
• Preserve protocol 28, public SDK/API, output schema, and explicit compute/render backend selection.

Correctness

• Public replay: 32 PDFs / 299 pages, two candidate runs per document; ModelJson, MiddleJson, assets, and diagnostics matched Stage 21 exactly.
• MinerU dev@f504cff: 31 eligible Flash text outputs and 32 actual medium shared outputs all matched the main baseline.
• Rust/session full tests: 5472 passed, 14 skipped.
• Python/legacy full tests: 4762 passed, 724 skipped.
• Cargo workspace tests, Clippy -D warnings, rustfmt, and Ruff check passed. Changed-file formatting passed; 20 pre-existing unrelated files still fail full-repository ruff format and were not modified.
• ABI3 wheel verified in an independent CPython 3.14.4 environment. Source and wheel extension SHA-256 both bc42f1546006c340cb7175b5bbfa6698641a59057758c046a8f5f4c0bf5a222e; DocVortex core and MinerU Flash smoke outputs matched for auto and rust.

Measured performance

One warmup plus five hot runs per document, alternating baseline/candidate execution order by document, with isolated process-tree RSS sampling. First-pass ratios above 5% were reverse-order retested; none persisted above 5%.

Chain	main	Stage 22	Reduction	Max time	Max RSS
DocVortex public parse, 32 PDFs	16.535675 s	16.049855 s	2.94%	1.19798	1.00948
MinerU Flash text, 31 PDFs	16.379259 s	15.990359 s	2.37%	1.10413	1.05150
MinerU medium shared, 32 PDFs	19.056479 s	18.956444 s	0.52%	1.26994	1.04256

Reverse retests:

• Public: demo4.pdf 0.94217; mixed_text_layout_sample.pdf.xor 1.04183; quarterly_report_financial_tables.pdf 0.98873.
• Flash: demo4.pdf 0.96987; annual_report_research_projects_table.pdf 0.93490; fund_asset_and_transaction_tables.pdf 1.01703.
• Shared: mixed_elements_pages_03_06.pdf 1.00299; small_ocr.pdf 0.98000; pollutant_discharge_tables.pdf 0.95873.

One-pass 32-PDF profile reductions include _detect_table_candidates 6.821288 → 4.258895 s, _materialize_table_blocks 5.383532 → 3.768801 s, _infer_text_lanes 3.251807 → 2.474491 s, and _merge_owned_table_candidates 2.223635 → 1.366970 s.

The original 2× objective remains incomplete; this PR reports only measured gains.

Not done

• No merge, release, version bump, PyPI publication, or downstream dependency synchronization.
• MinerU remains integration-only and its untracked examples/ directory is preserved.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
🔒 Security Review	✅ Completed 2026-09-29T18:22:42.038222Z	f3b53e6	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.