OK — many tests omit compatibilityDate, which defaults to the oldest. That's normal for this directory.

Now let me do a final check — the textdecoder-utf16-odd-offset-test.wd-test sets nodejs_compat_v2 but the test only uses TextDecoder, not a Node API. This isn't wrong (it will still work), but it's unnecessary. However, it's not an actionable issue worth flagging.

Let me now review the complete picture. Here's my analysis:

encoding.c++: The fix replaces an unsafe reinterpret_cast<const char16_t*> on a potentially unaligned buffer with a copy into a kj::SmallArray<char16_t, 256>. The inline buffer of 256 char16_t elements (512 bytes) handles typical cases without heap allocation. Correct.

buffer.c++ (indexOfBuffer):

1. Rounds down hayStackLength for UTF16LE to ensure even byte count.
2. Copies haystack and needle into aligned kj::SmallArray<uint16_t, 1024> buffers (2048 bytes inline).
3. The needleLen = needle.size() & ~1 rounds needle size down similarly.
4. All comparisons consistently use hayStackLength instead of hayStack.size().
5. Correct.

buffer.c++ (indexOfString):

1. Similarly rounds down hayStackLength.
2. Copies haystack into aligned buffer.
3. The decodedNeedle comes from decodeStringImpl which allocates fresh via V8's backing store allocator — this is guaranteed to be properly aligned, so the reinterpret_cast<const uint16_t*> on line 417 is safe.
4. Correct.

Tests: Properly test the exact scenarios — odd byte offsets for both TextDecoder and Buffer.indexOf with UCS-2/UTF-16LE encoding.

Backward compatibility: This is a pure bug fix (fixing UB from alignment violations). The behavioral change is going from crashing/UB to correct behavior, which doesn't need a compat flag.

I don't see any actionable issues with this PR. The alignment fixes are correct, the approach is sound (copy to aligned buffer), and the tests cover the scenarios.

LGTM

github run