View Markdown source ↗

Workspace ownership, retired documents and report lists

Actual installed acceptance of the beadb01b development-signed build exposed three related native usability defects: Finder commands created extra main workspace windows, empty rounded table stripes continued below real files, and Quit resurfaced a previously closed Commit window and failed to finish. This family implements the connected window/report ownership correction. Debug, signed compilation and ten local native case/engine pairs pass after recorded repairs. Static auditing, bounded repaired installed acceptance and owned cleanup are complete. Remaining Status refresh/close usability, incomplete global app window inventory and changed raw index bytes are recorded below.

Source relationship

Pinned TortoiseGit is 7338078f8ddd924b8cddee35f512f2286072136d. LogCommand.cpp:25–92 launches the requested Log dialog with DoModal; CommitCommand.cpp:40 and DiffCommand.cpp:30 similarly dispatch their requested workflows. These entry points do not imply creating a new repository workspace behind each command. The Mac app retains one identified workspace scene while owning its independent repository and editor dialogs.

The pinned resource report controls are IDD_CHANGEDFILES / IDC_CHANGEDLIST at TortoiseProcENG.rc:187–193, IDD_LOGMESSAGE / IDC_LOGMSG at 210–230, and IDD_COMMITDLG / IDC_FILELIST at 245–278. They use LVS_REPORT, rather than a rounded inset list extending empty stripes through unused space. This correction retains native macOS controls, source status colors/icons and the existing row interactions.

Implementation scope

• Reuse a single identified workspace scene for ordinary launch, Finder URLs, Services and reopen. Queue accepted external delivery until that workspace is attached, preserving the actual selected paths and captured request. Closing the workspace does not terminate an active owned child or disable future routing. Standalone TurtleGitMerge entry must not create a repository workspace.
• Apply report appearance only to Commit files, Working Tree files and Log changed files, using SwiftUI's own .bordered(alternatesRowBackgrounds: false) table style. Disable alternating empty backgrounds without altering other tables globally. Preserve checkboxes, selection, grouped rows, fonts, colors, menus, sorting and copy/drag behavior. SwiftUI remains the sole style owner; the adapter must not race it by changing or restoring NSTableView.style.
• Exclude retired document controllers from new unsaved-document Quit prompts, and recheck retirement before each asynchronous prompt dispatch. A closed Commit must not be shown again as a disabled, unanswerable window. Live dirty documents retain Save/Discard/Cancel behavior, and operation/release counters continue to protect draining workers.

The implementation touches TurtleGitMacApp.swift, ApplicationDelegate.swift and the three report-table consumers. The attempted imperative styling changes to CommitFileInteraction.swift have been removed; its existing selection and interaction bridge retains its original ownership. These are native scene, presentation and ownership changes; they do not change Git content or treat forced termination as successful cancellation.

Actual failure and bounded installed evidence

Installed-family evidence records the measured build/install, real Finder Commit/Log/Diff routes, Settings draft handling and Status cancellation. It also preserves the failed Quit of PID 54430: Command-Q timed out, a retired Commit resurfaced with disabled controls, and the exact owned app required SIGTERM. A fresh application later quit normally during cleanup; that does not erase the original failure.

After the failed GUI Quit, the independent cache helper updated a private watched file and folder from normal to modified and back to normal while the GUI was absent. Actual Finder states and cache snapshots support that bounded refresh observation. The watch was removed, background mode restored off, the fixture's saved permission forgotten and owned windows closed while preserving the user's Downloads window. The installed Finder extension remained alive. This is not a pass for the complete graceful-Quit/background lifecycle.

Checkpoint failure and repair in progress

The first Debug build (session 3790) exited zero while retaining five contradictory exit-zero diagnostics. The grouped native attempt (session 39879) then failed on its first Commit/system-Git case: receiver PID 63821 exited with SIGSEGV 11 and was reaped. build/qa/report-workspace-family-native-2026-10-25.log and the macOS diagnostic report receiver-bbad4c1881ea90de36b701222e6273fe527dda6699784df14702965-2026-10-25-002140.ips record the failure. This is an actual crash, not a passing native checkpoint.

The crash points to the new native .fullWidth table mutation racing SwiftUI's .inset style ownership. The repair removes that adapter style ownership and restoration entirely and selects the supported SwiftUI bordered, non-alternating style at the three table declarations. This retains the intended presentation without two frameworks competing over the same native table style.

The parallel signed build (session 75039) was deliberately interrupted after the actual crash and exited 75. It was not a timeout, a completed candidate or a successful signed validation. The earlier measured build-input receipt beginning 105506… is invalidated by the repair; current inputs must be measured again once implementation freezes.

Verification and remaining limits

The affected Debug repair build (session 2495) exited zero; four contradictory exit-zero diagnostics remain in that log. The repaired native attempt (session 97771) exited one on a new fixture assertion, not a production crash: it incorrectly required discarded native highlight to survive the unchanged bridge's teardown. Source review confirmed that teardown restores the framework delegate and clears that discarded selection. The corrected check requires a retired adapter to refuse drag and remove its artwork; all live row, color, selection, checkbox, sort, drag, light/dark and replacement assertions remain.

The focused Commit recheck (session 78300) passed with both Git engines; the remaining eight case/engine pairs (session 31826) also passed. All ten distinct pairs pass cumulatively after the recorded repairs. Passing cases were not repeated; cached app objects were reused and only the changed fixture was recompiled. All ten receiver PIDs were reaped and independently confirmed absent.

The measured repaired signed build (session 99954) exited zero, retaining one contradictory exit-zero diagnostic. The static candidate audit (session 15360) passed: measured repair inputs remain unchanged, eight owned code items are universal, 206 runtime code items and 6,356 inventory members are verified, and all three app/Finder/cache identities use the expected team and App Group. Signing, runtime placement and structural caches pass. Installation (session 21342) exited zero; the installed bundle matches the audited inventory and executable SHA-256 d3fe9dae34d25ebe037a30d5afd381ad1b5cdb3cbe13e95ddfe52660226929c9. The previous bead installation remains recoverable. The Finder executable itself is unchanged; public registration operations selected one installed plug-in, and the old extension retired naturally without a direct kill. The new input receipt captures 1,471 files at beadb01b with fingerprint 198a84fb…, including protected existing WIP unchanged.

Family evidence preserves the first crash and stopped signed build, superseded/current input receipts and actual validation scope. Core was not rerun for these app-only changes. A complete current hosted result is not available: the latest bead run is pending, while an older build 56 was still in Git integration at parent observation. Neither validates these uncommitted repairs. The earlier installed beadb01b evidence above remains historical, including its real Quit failure; the repaired installed observations below do not erase it.

Actual repaired installed observations

Finder pair Commit showed only the two selected files, then Cancel Yes retired it. Single-file Log retained alpha.txt scope, two revisions and the front graph, with the unrelated beta file shown in source gray. Closing Log retained the workspace identified as workspace. After closing that workspace, Finder pair Diff reopened the identified scene with the intended alpha/beta contents; closing Diff retained that same workspace ID. Grouped Working Tree showed the three fixture files in the plain report.

The parent individually reviewed six real installed captures: readable source colors and complete footers, without empty rounded stripes. Copies below retain the exact captured bytes; they were not edited.

Follow System appearance was restored. Command-Q then exited the repaired GUI PID 67908 immediately, without SIGTERM, after those child workflows had retired. The installed Finder extension PID 67907 remained independent.

These observations do not establish an absolute global window count: CUA's application inventory timed out and was reset, and the Window menu did not list the identified workspace. The evidence establishes the identified scene and its observed reopen/retention only.

Status refresh can recur between close attempts; normal Command-W cancels it, and rapid primary Cancel followed by Command-W closed the window. That remaining refresh/cancel usability needs investigation and is not full Status acceptance. Two stale accessibility actions while opening Status Appearance are retained; keyboard Return closed it, then it was reopened under Light from the stable workspace for the actual capture.

Final cleanup removed only the fixture's saved permission through Recent remove/Forget, removed its repository/watch entries, restored Follow System and quit the fresh cleanup app normally. The owned Harbor Finder window was closed; the Finder accessibility window inventory confirmed the user's Downloads window was preserved. No GUI, cache or receiver processes remained; Finder extension PID 67907 was preserved. The private repository is retained intentionally. Measured build inputs still match 198a84fb… after the installed interaction.

HEAD, config and all six original working files are unchanged. Raw index bytes did change: SHA-256 is now e0b7f9450006a85c5d07c1de36c980bc6d23e69460bcb169167a3650778db0b4, versus baseline 1dc9d64e2412a0729dc1b64da3bef8c667f735d9e1f14616ddd5c85b36d8ea7e. Staged content matches unchanged HEAD (git diff --cached --quiet exits zero), and all four index entries retain ordinary H flags without assume-unchanged or skip-worktree flags. A metadata-only refresh is inferred, not proven; this new installed checkpoint does not claim raw-index byte preservation. The earlier historical bead observation remains its own separate evidence.

Remaining scope includes physical focus/reopen and external grant variants, other upstream dialogs, minimum-version/Intel and a complete current hosted CI result. Neither the source anchors nor these bounded fixes certify the full port.