[vbucc]

Three EditMode tests failed on every headless run and were waived by hand at merge time (most recently #38). They are two unrelated problems, and neither is a production bug — both are tests asserting on ambient editor state that headless Unity does not provide.

Correction to #38: its body says Unity's native runner passes all three. It does not. It passes only EditorStateCache_IsChanging. The two ReadConsoleTests failures reproduce under the plain -batchmode -nographics -runTests runner with no MCP involvement at all — before this change, the native full suite was 1039 tests, 972 passed, 2 failed, exactly those two.

EditorStateCache_IsChanging_TracksTransitionWindowNotWholeSession

The test drove _isPlayModeChanging by reflection but left the other input to ComputeActivityPhase at whatever the ambient run had set. "running_tests" outranks "playmode_transition", and TestRunStatus.IsRunning is true whenever the suite was started through the MCP run_tests tool — which is exactly what the local harness does. So it passed natively and failed under the harness.

Now drives TestRunStatus._isRunning in the same try/finally instead of assuming a quiescent editor. Safe to poke: a [Test] body runs synchronously inside one editor frame and TestRunStatus is only read from main-thread callbacks.

ReadConsoleTests

Root cause, established against a live headless editor:

• read_console reports 0 entries at rest.
• Provoking a McpLog.Warning from an ordinary command handler → 1 entry, and it persists. So LogEntries works headlessly.
• Running HandleCommand_Get_Works alone and then reading the console filtered on "Test Log Message" → 0 matches, while that exact string appears both in Editor.log and in UTF's captured per-test output.

Unity does not record a log emitted inside a Test Framework test into the console buffer under -batchmode -nographics. These tests only ever passed in a GUI editor.

Split in two:

• SplitMessageAndStackTrace becomes internal and gets five direct tests (blank line inside the body preserved, CRLF normalization, stack frames split off, no-stack cases). That is what PreservesMultilineMessageBody was really about, and it runs identically in the GUI and headless. One visibility change is the whole production diff.
• The live-console tests seed an entry and Assert.Ignore when it is not observable — full strength in the Test Runner window, an explicit documented skip headlessly. They now filter server-side rather than scanning a count = 1000 window: GetConsoleEntries walks oldest-first and stops at count, so a busy console can push a freshly logged message out of an unfiltered window entirely.
• HandleCommand_Clear_Works drops its "console should have logs" precondition, which was satisfied only by unrelated tests' logs — the class failed the moment it was run alone. It now asserts what it owns: clear succeeds, console reads empty.

Harness

_start_utf now sends discardUntitledScenes. The smoke leg dirties the untitled scene (it creates and deletes GameObjects) and run_tests fail-fasts on that, because an interactive editor would wedge on a native Save dialog. A headless gate has none. This did not fire during #38's gate but fired immediately when probed, and would have surfaced as "run_tests start failed".

Gate

• Python: 1502 passed, 3 skipped.
• Harness unit tests: 140 passed.
• Native full EditMode suite: 1044 total, 977 passed, 0 failed, 49 skipped (was 1039 / 972 / 2 / 47).
• Harness --legs editmode: exit 0, 977/1026 passed, 0 failures.
• Harness --legs smoke,editmode: exit 0, smoke 7/7.

The gate is now green with no documented exception.

Not verified: the GUI Test Runner path. The two live-console tests should run there rather than skip; if they skip in the GUI too, the observability check needs revisiting.

🤖 Generated with Claude Code