[frankslin]

OCD files embed a BinaryDict section whose integer fields were written as native size_t, so 32-bit and 64-bit builds produced different bytes for the same dictionary, and files did not load across word sizes. All fields are now written as uint32_t (matching what 32-bit builds already produced), and the deserializer detects the legacy 8-byte layout, so files written by old 64-bit builds still load, on any platform.

Investigation:

• PR Upgrade vendored darts-clone to v0.32h and harden OCD dictionary loading #1372/Fix id_type to uint32_t in vendored darts.h; add legacy 64-bit OCD reader with full validation #1373 made the darts trie section word-size independent: vendored darts.h id_type was fixed to uint32_t, the dartsSize field became uint32_t, and a legacy 64-bit reader with validation was added. Those changes covered only the darts array; the trailing BinaryDict section (key/value buffers plus offset tables) kept writing native size_t fields, a layout that dates back to 641bfe9 which deliberately chose size_t to silence MSVC x64 warnings.
• As a result, .ocd files produced on 32-bit platforms (including wasm32 and armv7) differed byte-wise from 64-bit output for the same input, and could not be loaded across word sizes at all: BinaryDict::NewFromFile read sizeof(size_t) fields of whatever platform was running.
• The existing legacy-format tests could not catch this because SerializeAsLegacy64 copied the BinaryDict section verbatim. On a 64-bit host the fixture carried 8-byte fields and the reader consumed 8-byte fields, so the test was self-consistent yet word-size dependent; the same fixture would have failed on a 32-bit host.
• Layout detection reuses the high-bytes-zero probe already proven for dartsSize: an 8-byte numItems always has zero bytes at [4..7] for any realistic dictionary, while the fixed-width layout places keyTotalLength there, which is non-zero whenever the dictionary has entries. The only all-zero-probe fixed-width file is an empty dictionary, exactly 12 bytes long, which is disambiguated by size.
• Endianness is intentionally out of scope: the whole OCD/OCD2 family is native-endian (darts array, marisa trie, SerializedValues), so this change keeps BinaryDict consistent with the rest of the format on big-endian platforms rather than fixing one section in isolation.

Verification:

• bazel test //src/... : all 18 targets pass, including the new binary_dict_test cases (fixed-width layout size assertion, empty-dict round trip, legacy 64-bit loading, out-of-bounds rejection in both layouts) and darts_dict_test with the widened legacy fixture.
• dict_converter round trip text -> ocd -> text reproduces the input, and a hex dump of the generated .ocd confirms 4-byte numItems/keyTotalLength/ valueTotalLength and per-entry fields after the darts section.

Detailed Changes:

• src/BinaryDict.cpp:
  • Write all integer fields as uint32_t; reject values above UINT32_MAX.
  • Detect the field width on load via DetectFieldWidth and parse both layouts with a single bounds-checked reader.
  • Route NewFromFile through NewFromBuffer to share the detection logic.
• src/DartsDict.cpp:
  • Update layout-detection comments; no behavior change.
• Tests:
  • Make crafted BinaryDict files explicit about field width instead of depending on the host size_t, and cover both layouts.
  • Assert the serialized layout size is fully determined by content.
  • Cover the empty-dictionary round trip and legacy 64-bit loading.
  • Widen the BinaryDict tail in SerializeAsLegacy64 so legacy OCD fixtures match real files from old 64-bit builds end to end.