View Markdown source ↗

Byte-backed unified diff editing

The standalone unified-diff editor retains the bytes behind its displayed text. An edit changes the affected fragments; untouched UTF-8, regional text and opaque bytes are copied directly into Save, Save As and the private clipboard format. A mixed UTF-8/regional patch can therefore be edited without converting the entire document into one code page. Original Patch and serial-import previews remain read-only, and partial staging continues to use the original Git patch packets.

Encoding and control data

Inserted text uses the encoding of the fragment being edited. UTF-8 text can retain its Unicode characters beside regional text; regional insertions still need to fit that region's captured code page. ASCII regions in a regional document keep that captured policy. File BOMs remain separate from editable text, so Save does not add extra BOMs or remove an actual leading U+FEFF character.

Embedded NUL is data in a patch document, including UTF-16/UTF-32 documents. This is separate from the compare/merge text encoder, which retains its existing binary/NUL boundary. Undefined legacy bytes and malformed UTF-8 projections retain original bytes; a displayed replacement character is not used to regenerate them. Deleting a projected unit removes the bytes it represents. Native control-character glyph appearance still requires physical acceptance.

No automatic whole-file conversion or standalone conversion picker is introduced. An inserted character that cannot fit its selected region reports an encoding error and leaves the file intact. An unknown, non-round-tripping display mapping can also refuse a partial edit rather than regenerate uncertain bytes. Unchanged export and Undo remain available after dismissing the error. Reopening an inherently ambiguous mixed-codec byte sequence can reclassify its display; an in-memory save retains the accepted fragment presentation and code pages.

Save points, Undo and ownership

The byte edit tape retains removed fragments and their encodings. It recognizes native coalesced Undo/Redo notifications and restores the captured bytes, including opaque units, rather than guessing them from the restored text. Save keeps that history and ends native typing coalescence at the save point. AppKit event grouping still determines how many edits one Undo restores. Replacing the document and retiring its owner release the tape. Unicode spelling changes remain real edits even when Swift considers their strings equivalent.

Save As captures an immutable byte/presentation snapshot before opening its file chooser. The owner blocks editing while the chooser/write is pending. Changed preferences cannot reinterpret that snapshot. External-change checks, security scope lifetime, cancellation, retired-owner checks and shared Quit accounting remain on the existing document-session path. Copy Patch publishes both the exact current byte snapshot and its native Unicode text representation.

Source and verification

Pinned TortoiseUDiff MainWindow.cpp uses Scintilla's byte buffer and saves its SCI_GETTEXT bytes. The Mac implementation adapts that behavior to native text controls with retained byte fragments and an explicit legacy-code-page policy. The dated checkpoint records focused Core and grouped native coverage, failures and remaining limits. It includes mixed regions, all supported text codecs, NUL, malformed byte projections, exact Unicode savepoints, immutable snapshots and Undo across saves, plus actual native editing, clipboard and file-session ownership. Passing this group does not certify every Scintilla gesture, physical printing, complete visual parity or signed App Store acceptance. A same-text replacement of an opaque projection preserves the original bytes; explicit byte/encoding conversion remains outside this interface.

The checkpoint also records a fixture Undo-group exception and a subsequent private cursor-window close crash in the broad Help observer. Help now registers close state only for bound app windows, without asserting a Swift executor from framework close callbacks. The existing context/F1/sheet/close checks were rerun; this does not certify all SDK window-lifetime behavior.