[snaka]

Summary

• Replace the OCR text detectors (SensitiveDetectors + the broad Apple Intelligence prompt) with privmask's DetectionPipeline, pinned to 0.4.x.
• OCR lines are joined into one text for the pipeline; JoinedLines maps each finding's UTF-16 range back to its line so the existing glyph-rect lookup still applies.
• The on-device model is used for Japanese personal names only, and its findings are reconciled in the same pipeline (a model "name" wrapping a phone number is dropped). All candidates are masked regardless of confidence.
• Debug overlay colours now come from each region's label instead of the detector, since one detector now yields both deterministic and model findings.
• README gains a before/after sample of detection.

Why

NLTagger has no Japanese NER, so Japanese names were effectively never detected; the broad AI prompt produced false positives (port numbers and IPs as addresses). privmask measured 9/9 Japanese names on text input and filters implausible phone numbers (order numbers, timestamps).

Coverage changes

• Added: postal codes, My Number, API keys / secrets.
• Dropped: usernames/handles and credit-card numbers — tracked in Payment card numbers are not detected; verify them by issuer prefix and Luhn, as My Number is by its check digit privmask#25 and Usernames and @handles are not detected, and a plain @-pattern would mask decorators and scoped packages privmask#26.
• English-only screenshots no longer invoke the model (English names still come from NLTagger).

Test plan

• swift test --package-path Packages/BokashiCore (39 tests, incl. new JoinedLinesTests)
• Debug build; sample image via clipboard import → Detect masks names, phone, email, postal code/address, AWS secret and an English name, and leaves commit hash, version, port, IP, order number and timestamp alone (the README images are that run's export)

🤖 Generated with Claude Code