Native host test FormatUtilsTest.TestDetectFormatVersion fails on host g++ (quarantined from CI)

[AsafMah]

Surfaced by the new native host test harness (#78) — the first thing it caught.

Symptom

FormatUtilsTest.TestDetectFormatVersion fails on the host build (Linux, g++ 13, -O2):

format_utils_test.cpp:60: Expected equality of these values:
  FormatUtils::VERSION_2  (2)
  FormatUtils::detectFormatVersion(...)  (-1 == UNKNOWN_VERSION)

The test builds a 12-byte big-endian header [9B C1 3A FE | 00 02 | 00 00 | 00 00 00 00] and expects detectFormatVersion to report VERSION_2; on host it reports UNKNOWN.

What was ruled out

• Size guard: DICTIONARY_MINIMUM_SIZE = 12, buffer is exactly 12 → guard passes.
• readUint32 signed-shift UB: suspected (uint8_t promotes to int, 0x9B << 24 overflows signed int = UB), but an isolated g++ -O2 repro shows both the int-promoted and uint32_t-cast forms produce 0x9BC13AFE. So the magic read works in isolation.
• Static analysis of the full path (readUint32 → switch → getFormatVersion(readUint16(...,4))) looks correct.

Root cause not yet isolated — likely a host/optimization or ReadOnlyByteArrayView/AK_FORCE_INLINE interaction specific to the host toolchain. The engine works correctly on-device (the app ships and reads dictionaries fine); this is a host-build-only discrepancy in a test that had never been run in CI.

Action

• Quarantined from the A3d: Native gesture-recognition replay harness + run native tests in CI #78 CI gate (ctest -E 'FormatUtilsTest.TestDetectFormatVersion') so the other 65 native tests gate as a blocking net.
• This issue tracks isolating the real cause (and, if it is the readUint32 signed-shift UB after all, hardening it with explicit uint32_t casts — provably identical on-device, removes UB).

Low priority (no on-device impact), but a good demonstration that the harness reaches code the JVM tests cannot.