fix(pptx): omit fully occluded searchable text

[developer0hye]

Summary

• omit an earlier wrapping, upright PPTX text box from generated PDF markup only when a later opaque, untransformed raster covers both the page and the text frame
• preserve text for partial-page images, transparent rasters, rotated boxes, and no-wrap boxes whose glyphs can escape their declared frame
• add a PDF-level searchable-text regression and strengthen the visual contract so Text-layer-only: Yes requires real zero decoded-pixel difference between before.jpg and after.jpg
• document and store the text-layer-only evidence workflow for page 1 of the reported deck

Related issue

Fixes #1432

Testing

• Red first: cargo test -p office2pdf render::typst_gen::tests::fixed_page_tests::test_fixed_page_omits_text_fully_occluded_by_later_opaque_jpeg retained Sensitivity: in generated markup before the implementation.
• Red first: python3 -m unittest scripts.tests.test_check_visual_pr.LayoutAuditTests.test_text_layer_only_fix_rejects_different_decoded_before_and_after_images showed that the initial declaration-only exception accepted different rasters.
• cargo test -p office2pdf render::typst_gen::tests::fixed_page_tests — 20 passed on the rebased head, including PDF text extraction and the partial/transparent/no-wrap safety cases.
• python3 -m unittest scripts.tests.test_check_visual_pr — 42 passed.
• CARGO_TARGET_DIR=/Volumes/T7/targets/office2pdf-issue-1432 cargo clippy --workspace --all-targets -- -D warnings — passed on the rebased head.
• OFFICE2PDF_VALIDATE_PDF=1 CARGO_TARGET_DIR=/Volumes/T7/targets/office2pdf-issue-1432 cargo test --workspace — passed before the conflict-free rebase: core 2,753; DOCX 188; PDF validation 7; PPTX 130; XLSX 206; CLI, package roundtrip, model roundtrip, and doctests all passed with zero failures. Focused tests and clippy were then repeated after the rebase.
• Converted the reported PPTX with the rebuilt CLI; pdftotext -layout no longer contains Sensitivity: Internal on page 1 while the visible title remains searchable.
• python3 scripts/compare_layout.py --page 1 --audit --json native-powerpoint.pdf office2pdf.pdf — page count 1/1, missing 0, extra 0, reflow 0, visibility mismatches 0, large shifts 0.
• magick compare -metric AE assets/bugfixes/issue-1432/before.jpg assets/bugfixes/issue-1432/after.jpg null: — 0 (0).
• Re-rendered the current page at 300 DPI, encoded it with the evidence settings, and compared it to committed after.jpg — decoded-pixel AE 0 (0).

Visual impact

• No rendered PDF change
• Rendered PDF change or visual evidence added

Visual audit

• Issue: PPTX fully occluded footer text remains searchable in exported PDF #1432
• Fixture: GENERAL SERVICES.pptx reported through PPTX table cell fills lose transparency over a slide photograph #1371
• Page(s): 1
• Renderer and DPI: pdftoppm, 300 DPI
• Evidence mode: fix
• Text-layer-only: Yes
• Pixel delta: 0
• Layout audit report: assets/bugfixes/issue-1432/layout-audit.json
• Layout audit page count: Pass
• Layout audit text flow: Pass
• Layout audit large shifts: Pass
• New follow-up issues found in this audit: None
• Model vision findings: At 300 DPI Codex inspected the full GT, before, and after pages, the 5% fuzz diff, and matched footer, title, and rule crops. The footer has no painted pixels in either before or after, and those two renders are decoded-pixel identical. Remaining GT/output edge clusters are limited to text and rule antialiasing or subpixel rasterization below the audit thresholds; no untracked visible deviation remains on page 1.
• GT: assets/bugfixes/issue-1432/gt.jpg
• Before: assets/bugfixes/issue-1432/before.jpg
• After: assets/bugfixes/issue-1432/after.jpg

Visual comparison

GT	Before	After

Required inspection

• Rendered all evidence at 150 DPI or higher
• Stored progressive JPEG quality 86 assets with metadata stripped
• Used Codex/Claude vision to inspect the full GT/output pages, diff, and matched crops
• Inspected matched region crops at full resolution
• Ran compare_layout.py --audit and dispositioned every large text-instance shift and painted-visibility mismatch
• Ran the 5% fuzz pixel-difference sweep
• Inventoried hairlines and border dash styles
• Inventoried font weight, italic, and underline emphasis

Deviation audit

Check	Result
Page count/order	Matches GT — one compared page in the page-1 audit
Element presence	Fixed — the fully covered footer is absent from the searchable layer and remains absent from the pixels
Position/size	No deviation observed — before/after pixels are identical; GT text baselines and x/width deltas remain below audit thresholds
Rotation/flip	Matches GT — no rotated or flipped page-1 element differs
Fill	Matches GT — no before/after pixel changed
Stroke/border	No deviation observed — the title rule remains present; only accepted rasterization edge noise appears against GT
Shape outline geometry	Matches GT — no outline geometry changed
Text content	Fixed — Sensitivity: Internal is no longer extra searchable text; the three visible lines still match GT
Font family/weight/style	No deviation observed — full-resolution text crops retain the same visible styling
Text color	Matches GT — no before/after pixel changed
Alignment	No deviation observed — matched text remains within the layout audit thresholds
Line/paragraph spacing	No deviation observed — no wrap or reflow finding remains
Clipping/overflow	Fixed — fully occluded wrapping text is culled conservatively while no-wrap, rotated, partial-cover, and transparent-cover cases stay searchable

Checklist

• Commits include a Signed-off-by line
• PR scope contains one root cause
• Remaining visual deviations each reference an open issue