[iamnbutler]

Fixes the Finder cluster and the #77 key failures from the 0.13.1 live run (14 pass, 8 fail, 17 skip).

Causes and fixes

Product

• Inspect trimming dropped Finder's file list. The 32 KB result cap cut from the tail, and Finder's ~120-element sidebar comes first. Trimming now keeps scope containers (window, split group, scroll area, outline, table, list, browser, tab group, web area) and drops other elements first. Tested in test/inspect-trim.test.mjs.
• Table reads.
  • A label that only repeats the role description ("cell", "column", "image") was used as cell text or a column title.
  • A folder's disclosure triangle put 0 in its Name cell.
  • A sorted column was titled AXAscendingSortDirection.
  • All three are fixed, with a fixture captured from real Finder (finderRoleLabels).
• Command-A on a text field returned unknown with "receipt does not match successful projected response outcome".
  • The client sends cmd,char:61, and peekaboo-layout-keys.patch resolves char: keys with KeyboardLayoutTranslator.current(), which is nil off the main thread.
  • The client validates receipts off-main, so isSelectAllShortcut was false there. The contract lacked the AX-value select-all rule that the host (on main) had used.
  • New peekaboo-layout-shortcuts.patch classifies paste and select-all chords without a layout.
• feat: paste files into a Finder folder with desktop_insert, confirmed by their names appearing there (#3) #114 paste refused, "no exact focused control".
  • Finder reports two focused elements: an off-screen sidebar cell and the list outline. Peekaboo's tie-break (CFEqual with the app's AXFocusedUIElement, plus window ownership) was off for any truncated tree, and Finder always hits the 200-element budget.
  • New peekaboo-focus-budget.patch allows the tie-break for budget cuts (depth, count, children) but not for a deadline, an incomplete read or the app-scoped fallback.
  • Upstream's suppression tests keep those cases, and a new test covers the budget kinds. The README explains why the identity isn't weakened.
• feat: write one table cell by row and column, confirmed by reading the table again #92 Finder rename reported completed without renaming.
  • Finder shows a name written to its name field through Accessibility, but never renames the file. AXConfirm doesn't commit it, and the background keyboard route can't reach the name editor.
  • desktop_set_cell, desktop_type and desktop_set_value now refuse a Finder file-list name field before input, with CELL_NOT_EDITABLE and a hint (shell mv, or the foreground Finder route).
  • The route is added to docs/recipes.md and marked unverified.
• Key sequences stopped by a focus change (Command-N, then more keys) reported a bare DESKTOP_ACTION_FAILED. The host's at-dispatch focus failure is indeterminate, and the client only handled undispatched failures. Every mid-sequence stop is now KEY_SEQUENCE_STOPPED, and the message says when the stopping key may have been sent.

Harness (scripts/live-check.mjs)

• tableOf read result.data.table, but columns and rows are top-level. finderTable picked the sidebar outline.
• The feat: hold modifier keys during clicks, and press key sequences against one snapshot #77 key arrays used bare strings, which are refused client-side.
• brief() crashed on JSON.stringify(undefined). This was the "reading 'length'" error in feat: paste files into a Finder folder with desktop_insert, confirmed by their names appearing there (#3) #114, feat: paste an image or files into a native text area with desktop_insert, or leave the clipboard replaced on request #99 and feat: save the clipboard image's original bytes with desktop_clipboard_read save_path #96.
• The feat: write one table cell by row and column, confirmed by reading the table again #92 Finder check now asserts the refusal and that the file is unchanged.

Verification

• npm test: 77 pass. npm run check: clean. npm run build:native succeeds with both new patches.
• Peekaboo tests in a scratch checkout at the pinned revision with all patches: AutomationKit ObservedFocusCorroborationTests|HotkeySelectionPolicyTests|HotkeyServiceTargetingTests, 89 pass. Core PeekabooBridgeOperationSemanticPlanTests|HotkeySelectAllReceiptTests|PeekabooBridgeClipboardGuardedHotkeyTests, 74 pass.
• Live run against the installed 0.13.1 with this branch's harness:
  • Pass: feat: read a table, outline or grid as columns and rows with each row's element ID #86 Finder table (columns Name, Date Modified, Size, Kind; 9 rows), feat: hold modifier keys during clicks, and press key sequences against one snapshot #77 control/right refusal, Shift-click, repeat 5.
  • Fail as expected, because the host fixes aren't in 0.13.1: feat: paste files into a Finder folder with desktop_insert, confirmed by their names appearing there (#3) #114 ×2 (no focus), feat: write one table cell by row and column, confirmed by reading the table again #92 (completed instead of the refusal), feat: hold modifier keys during clicks, and press key sequences against one snapshot #77 Command-A sequence, feat: hold modifier keys during clicks, and press key sequences against one snapshot #77 KEY_SEQUENCE_STOPPED.
  • Not yet verified on the dev app, which is waiting for its grants.

Known, not fixed here

• feat: hold modifier keys during clicks, and press key sequences against one snapshot #77 Command-click and Shift-click come back unknown every time, with "Cursor restoration: The physical cursor restoration was dispatched but could not be verified". It happened 4 of 4 times on a fixture window. The selection itself is right in my runs. The harness run once selected only the first row. This looks like a bug in modifier-click cursor restoration and needs its own issue.