perf: avoid repeated line feature work in Flash PDF

[myhloli]

Motivation

Native inline-script matching recomputes font-size medians for every candidate pair, and ordinary lines compute structural script evidence that is discarded.

Modification

Precompute font scales, upright bounds and reference-marker features within each matching pass. Compute strong script roles only when consumed. Prefilter same-baseline pairs by necessary classification conditions and delay table-obstacle checks until geometry matches, preserving pair order and merge decisions.

This is stage 2/6 of the Flash PDF equivalence refactor, based on codex/flash-pdf-01-baseline. Review and integrate stages in order.

Validation

31 documents / 298 pages retain identical complete model-list and Middle JSON output. 65 focused tests pass, including cache lifetime and linear font-statistic work. Four representative five-run medians improve by 11.8%–34.4%; peak RSS remains within 0.4% of baseline.

The original test_demo_sparse_table_confidence_manifest bbox expectation failure is documented separately and its expected data is not modified.

Compatibility

The existing PdfModel.predict(), PDFDocument methods, ModelJson/MiddleJson output and renderer contracts remain unchanged. Recognition thresholds, candidate priority and fallback behavior are preserved.

Checklist

• Relevant Ruff checks and git diff --check pass.
• Focused regression and complete corpus output comparisons recorded.
• New functions have Chinese docstrings and explicit type annotations.
• Documentation or module responsibility docstrings updated.
• Remote CI and CLA checks completed.