fix(ci): make main build, lint and test green again

[datlechin]

Main has been red for three pushes. This turns every failing check green again.

What was broken

Compile every plugin (the run that prompted this): EtcdHttpClient's two URLSessionDelegate classes are not final, and under the Swift 6 language mode adopted in #2311 a non-final class cannot conform to Sendable. EtcdDriverPlugin was the only target that failed, so the job never reached the test steps.

Run unit tests was already failing before #2311, on the two pushes before it, in eight suites with three separate causes:

• fix(hig): spell every trailing ellipsis the way macOS does #2304 respelled every trailing ellipsis from ... to … and updated only the guard test it added. Eleven assertions across six unit-test files and six UI-test files still pinned the old spelling. The UI ones look up menu items by title, so they would have timed out once the unit step stopped failing first.
• refactor(mcp)!: rewrite the server against protocol revision 2026-07-28 #2297 added fields.removeValue(forKey: "extensions") to BridgeProxy. The preference-key guard is a text scan for forKey: "..." and cannot tell a UserDefaults key from a Dictionary key. "extensions" is an MCP wire field and can never be namespaced, so it belongs in the frozen baseline.
• refactor(mcp)!: rewrite the server against protocol revision 2026-07-28 #2297 also added MCPHandlerOutcomeGate.settle, which returns nothing. The settle guard scans for the bare substring .settle(, so it demanded three MCP call sites consume an answer that does not exist. The scan is now keyed on the receiver, tabExecution.settle(, and a second test asserts the registry is only ever reached through that property name, so the narrowing stays fail-closed.
• feat(plugin-duckdb): open Parquet, CSV, TSV and JSON files as a connection #2290 gave DuckDB csv, tsv, json, ndjson and parquet, so a .csv with no inspector plugin now opens as a DuckDB connection rather than returning nil. The test still described the old contract. It now describes the new one, and a second case pins the three kinds that have no inspector at all.
• build(plugin-duckdb): link core_functions, json, parquet and icu into the macOS library #2288 rewrote build-duckdb.sh to a CMake source build and turned the pin into ${DUCKDB_VERSION:-v1.5.2}, which the license inventory's regex cannot read, while licenses.yml still recorded v1.5.3 (a version that was never buildable: its checksum was a placeholder). The shipped library, the docs and the iOS xcframework are all 1.5.2, so the inventory and both build scripts now say 1.5.2.

SwiftLint in the release workflow, which runs on v* tags only and therefore never ran on the PRs: measureColumnWidth reached 11 parameters in #2309, one over the function_parameter_count warning that --strict turns into an error. The three sizing-budget parameters are now one value.

Verification

• xcodebuild build -scheme EtcdDriverPlugin: succeeds.
• xcodebuild test -only-testing:TableProTests with the CI quarantine list applied: 11,520 tests, 1,373 suites, all pass.
• xcodebuild test -only-testing:TableProUITests: 78 tests, 3 failures, all three keyboard-driven (QueryFormatUnterminatedLiteral, QueryHistoryActions.testReturnInTheListLoadsTheSelectedEntry, QueryHistoryFocus.testArrowKeysKeepMovingThroughTheListAfterClickingARow) and all three green on the CI runner in run 32391276238. They need the app frontmost to receive keystrokes, which a machine in use cannot promise.
• swiftlint lint --strict: 0 violations in 1,606 files.

No CHANGELOG entry: every behaviour these touch is either test-only or still unreleased.

The UI step, which the 60 minute cap had been hiding

The unit step went green, so the UI suite ran for the first time in days and turned up three more things.

Closing the last connection killed the app. ConnectionCloseUITests reported "application com.TablePro is not running" on CI and reproduced here. The crash report is EXC_BREAKPOINT in _dispatch_assert_queue_fail under swift_task_checkIsolated, from the cancel handler in DatabaseFileWatcher.startSource. DatabaseFileWatcher is @MainActor, so the handler passed to setCancelHandler inherited that isolation, and dispatch calls it on com.apple.root.utility-qos; under the Swift 6 language mode that check traps instead of warning. Both handlers are @Sendable now, which is what they always meant to be: the event handler already hops to the main actor with a Task of its own, and the cancel handler only closes a file descriptor.

Two suites that the timeout had never reached were querying rows that cannot match. ResultStatusBarUITests and StructureTabIdentityUITests ask for staticTexts["Album"], but the object browser hosts its rows as cells, so the name arrives as the static text's value carrying the object kind it reads out to VoiceOver ("Table: Album"), and the rows report as disabled so a plain click() waits for a hittability that never comes. SidebarTableTabUITests already documented both facts and did it correctly; that knowledge is now objectBrowserRow(_:kind:in:) and clickAtCenter(_:) on UITestCase. StructureTabIdentityUITests also asserted isSelected on a hosted radio button, which answers through AXValue and reads false however the picker is set, the same workaround OpenQuicklyCommandUITests carries.

RedisConnectionModeUITests had gone stale in four places at once: the chooser is a List, so buttons["Redis"] never matched; Redis sits below the fold, so a coordinate click landed outside the scroll view until the search field filters the list; one click only selects, and committing needs the double-click that fires primaryAction; and the form's plugin fields arrive with no label or identifier at all. The form now labels its controls the way the rest of the connection form already does, connection-field-<id> from ConnectionFieldRow and HostListFieldRow plus connection-form-host and connection-form-port, and the test reaches every element through those.

The job's timeout-minutes goes to 90 to match what the suites actually cost, with the measured breakdown in the comment above it.

Not fixed here

TableProTests/Localization/StringCatalogIntegrityTests.swift fails locally and has never run in CI, because .gitignore:182's unanchored Localization/ also matches TableProTests/Localization/, so the file is not tracked. Its one failure is a guard defect rather than a bad translation: the source unit for "%1$@ chart of %2$@ by %3$@ with %4$d points" carries a %#@points@ substitution declaring argNum: 4, formatSpecifier: "d", and the guard reads the placeholder as %@ instead of resolving the substitution, so it flags four correct translations. Worth its own PR, since anchoring the ignore rule puts a never-run guard into CI.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.