{
	name:       "windows character device with device flag",
	goos:       "windows",
	stderrMode: os.ModeCharDevice | os.ModeDevice,
	want:       true,
},

The implementation uses stderrMode&os.ModeCharDevice != 0, which handles combined flags correctly — this test would make that intent part of the explicit spec.

{
	name:       "windows zero mode stat success",
	goos:       "windows",
	stderrMode: os.FileMode(0),
	want:       false,
},

Since init() initialises stderrMode to os.FileMode(0) before reading stderrInfo.Mode(), this also documents the fallback default intentionally.

Option A — add an os.ErrInvalid case to the test table with a comment linking it to this path:

{
	name:    "windows ErrInvalid (init nil-nil fallback)",
	goos:    "windows",
	statErr: os.ErrInvalid,
	want:    false,
},

Option B — hoist the nil-nil guard into shouldConfigureLipglossCompat by passing stderrInfo directly (or a (mode, modeValid bool) pair), making the entire decision testable through the pure function.

Option A is a lower-effort win; Option B is structurally cleaner.

The current check is stderrMode&os.ModeCharDevice != 0, so both values satisfy it today. But the test case supplies a value that cannot arise in practice and creates a latent trap: if the check is ever tightened to the semantically precise form:

return stderrMode&(os.ModeDevice|os.ModeCharDevice) == (os.ModeDevice|os.ModeCharDevice)

the existing os.ModeCharDevice-only case would become a false negative — the test would pass a value the stricter check rejects, masking the regression.

Suggested fix: use the real Windows console mode shape:

{
    name:       "windows character device",
    goos:       "windows",
    stderrMode: os.ModeDevice | os.ModeCharDevice, // FILE_TYPE_CHAR on Windows
    want:       true,
},

The stderrMode := os.FileMode(0) fallback in init() is justified by: "Zero mode means 'unknown/unset'; on Windows this keeps the startup probe disabled unless stderr is explicitly confirmed as a character device." A test should pin this:

{
    name:       "windows unknown mode (zero) no error",
    goos:       "windows",
    stderrMode: os.FileMode(0),
    statErr:    nil,
    want:       false,
},

Without this, a refactor that treats zero-mode as "assume TTY" would go undetected.

// Before
return stderrMode&os.ModeCharDevice != 0

// After
return stderrMode&(os.ModeDevice|os.ModeCharDevice) == (os.ModeDevice|os.ModeCharDevice)

On Windows, Go sets ModeDevice | ModeCharDevice together for FILE_TYPE_CHAR handles, so the current single-bit check happens to work. But accepting ModeCharDevice without ModeDevice is technically a false-positive admission of an invalid mode shape.

This also brings the check into alignment with the test fix suggested on the stderrMode: os.ModeCharDevice case — if the test is updated to use os.ModeDevice | os.ModeCharDevice (the real production value), the stricter check here would correctly accept it, while the current looser check would accept anything with just ModeCharDevice set.