fix(ocr): keep scanned/image-only marker for mixed scanned+failed PDFs

[dev07060]

Summary

Fixes Finding #1 from the OCR-detection review: a scanned/image-only PDF that also had a page fail to extract took the page-failure error branch, which omits the scanned/image-only marker the Dart layer keys on — so OCR guidance was silently dropped for a recoverable document.

The fix

extract_text_from_pdf now decides the marker via a small predicate, should_include_scanned_marker:

• keep the marker whenever at least one page parsed successfully (a readable-but-text-less page is the scanned signature OCR can recover), and
• withhold it only when every page failed to parse (genuine corruption — OCR cannot help).

The classification is intentionally count-based and recall-favoring: we never silently drop a potentially recoverable scanned PDF just because one page is corrupt. Pure-scanned and degenerate zero-page behavior is unchanged.

The below-threshold message construction is extracted into below_threshold_error_message, a pure function, so the exact wording the Dart classifier depends on is unit-testable without a PDF fixture.

Messages emitted (below threshold)

Case	Message contains marker?
Pure scanned (0 failed)	✅ …; PDF may be scanned/image-only
Mixed (some parsed, some failed)	✅ …; N of M page(s) failed to extract (pages: […]); PDF may be scanned/image-only
All pages failed	❌ …; N of N page(s) failed to extract (pages: […])

Dart isOcrRequiredPdfExtractionError is unchanged — it already keys on the marker substring, which Rust now emits for the mixed case.

Tests

• test_should_include_scanned_marker — predicate boundaries (single-page, multi-page, mixed ratios, zero-page, all-failed).
• test_below_threshold_error_message — pins the exact emitted strings for pure-scanned, mixed (1-of-5, 3-of-5), and all-failed, tying the Rust producer to the Dart matcher contract so the format can't drift undetected across the FFI boundary.
• Dart: rewrote the stale negative test to a genuine all-failed case; added a mixed scanned+corrupt regression test.
• The #[ignore]'d scanned-fixture test now also asserts the scanned/image-only marker.

Verification

• Rust: cargo test --lib -- --test-threads=1 → 147 passed / 0 failed / 10 ignored.
• Dart: flutter test test/unit → 59/59 passed; flutter analyze lib/ test/ → clean.