Fix two persistent CI-blocking test-harness bugs (macOS SIGSEGV + Linux death-test timeout)

[beaucollins]

Problem

Two jobs have been persistently red on main — macOS: C++ & Platform Tests and Linux: C++ Tests (same shape across the last several nightly runs). Neither is a product regression; both are test-harness bugs that only manifest on CI runners.

Root causes & fixes

Linux — //valdi:test TIMEOUT

RuntimeTests/RuntimeFixture.AsyncStrictModeSyncCallAssertsOnMainThread is the only EXPECT_DEATH in the suite. It uses the default "fast" death-test style, which fork()s in place. The fixture owns a live runtime with several worker/JS threads, so the forked child inherits them frozen mid-flight (gtest even warns: "Death tests use fork() … detected 12 threads"). On loaded CI runners the child then stalls ~94 s on an orphaned lock before the assert's abort() completes; three parameterized variants blow the //valdi:test timeout.

Fix: switch this test to "threadsafe" death-test style, which re-execs a fresh process for the death check so no locked mutexes are inherited.

macOS — SIGSEGV in valdi_macos_objc_test

testPhysicalTextEditingAppliesWillChangeAndSynchronizesNativeOverrides intends to skip on headless CI, but gated the skip on CGSessionCopyCurrentDictionary() != NULL. GitHub's macOS runners report a login session, so the guard failed to skip; the test then drove a live key window + first responder + field editor and SIGSEGV'd in AppKit window activation (before any assertion — exactly the failure the test's own comment predicts).

Fix: require the session to be active on the console (kCGSessionOnConsoleKey) and bail out under CI, where no usable WindowServer is guaranteed. The live-editor path still runs on developer machines with a real GUI session. (The test was already a no-op on CI by design; this just makes the skip actually fire there.)

Verification

• Linux: reproduced locally and confirmed the fix — the 4 variants now finish in ~30–50 ms each (previously 94 s+ per variant on CI) with --gtest_print_time=1. The "fast"→"threadsafe" re-exec works fine under the Bazel test sandbox.
• macOS: header-only/logic change (CoreGraphics already imported); the Apple-only target builds under the macOS CI job.

🤖 Generated with Claude Code

[github-actions]

📊 PR Size: size/S

Total changes: 23 lines (2 files)

Top files changed:

• valdi/test/macos/SCValdiMacOSViewManagerTests.mm: +12 -4
• valdi/test/integration/Runtime_tests.cpp: +7 -0

_{Size calculated as additions + deletions. Labels: XS (<10), S (<50), M (<250), L (<1000), XL (1000+)}