[chinaux]

Upstream moved six commits ahead of 60956b1, and one of them removes four C-API functions. This PR tracks c133f20 and adapts the Rust side.

1. Track upstream zvec c133f20

vendor/zvec 60956b1 → c133f20 (upstream main HEAD). Compare API: status: ahead, ahead_by: 6, behind_by: 0.

• c549686 build(rabitq): upgrade to RaBitQ 0.7 modular targets — the submodule path moves to thirdparty/RaBitQ-Library/RaBitQ-Library-0.7.0 (de793e8), so an existing checkout needs git submodule update --init --recursive.
• ab58382 perf(flat): batch segment locking to fix the HNSW RaBitQ build regression.
• 89e83db perf(search): output refined internal IDs directly.
• e30317f fix(ivf): support primary-key brute-force search for filtered queries.
• 52b72a4 feat(fts): embed default Jieba resources with LZ4 — the breaking one, see below.
• c133f20 ci: publish Windows ARM64 wheels and SDK — upstream now builds WoA on windows-11-arm with msvc_arch: arm64 and Python ≥ 3.11 (CPython's WoA builds start at 3.11), i.e. exactly the setup chore(vendor): track zvec 60956b1; ci(prebuilt): add Windows ARM64 #18 added to our prebuilt matrix. Nothing to change on our side.

Build inputs we pass are unaffected: BUILD_C_BINDINGS / BUILD_TOOLS options and the zvec_c_api target all still exist at c133f20, and the lz4 submodule the embedded dict links was already in .gitmodules.

2. C-API change: the jieba dict-dir functions are gone

52b72a4 embeds cppjieba's jieba.dict.utf8 and hmm_model.utf8 into libzvec_c_api as LZ4-compressed arrays (jieba_dict_lz4.inc, jieba_hmm_model_lz4.inc), so an external dict directory is no longer needed. Upstream deleted:

removed C symbol	our wrapper
zvec_config_data_set_jieba_dict_dir	ConfigData::set_jieba_dict_dir
zvec_config_data_get_jieba_dict_dir	declaration only
zvec_set_default_jieba_dict_dir	set_default_jieba_dict_dir
zvec_get_default_jieba_dict_dir	get_default_jieba_dict_dir

GlobalConfig::ConfigData::jieba_dict_dir is gone as well, and the tokenizer now resolves the dictionary as: per-field extra_params["jieba_dict_dir"] > ZVEC_JIEBA_DICT_DIR > embedded default. CMake no longer installs data/jieba_dict into the SDK, and upstream's _build_prebuilt.yml stopped copying data/ in the same commit.

Rust changes

• zvec-sys — dropped the four extern declarations and build_time_jieba_dict_dir(); build.rs no longer hunts for a cppjieba dict directory or emits ZVEC_RUST_JIEBA_DICT_DIR.
• zvec — removed ConfigBuilder::jieba_dict_dir, set_default_jieba_dict_dir, get_default_jieba_dict_dir and the whole auto-registration path (auto_register_default_jieba_dict_dir, discover_jieba_dict_dir, is_jieba_dict_dir, and the unix/windows/fallback loaded_library_dir). initialize is now a straight zvec_initialize call. The dladdr probe was the only reason zvec/build.rs linked libdl on Linux, so that link is gone too.
• zvec-build — stopped staging data/jieba_dict next to the built artifacts; it stages the shared library only.
• zvec/tests/jieba_fts_test.rs — the dict-discovery test is gone, and the Chinese-search test no longer skips when no dict directory is found: with the embedded dictionary it must always pass. That test is what actually proves the bump works end to end.
• build-prebuilt.yml — dropped the two data/jieba_dict bundling blocks so our tarballs match upstream's SDK layout.
• Both READMEs — jieba is documented as embedded; the override list is now per-field extra_params > ZVEC_JIEBA_DICT_DIR, plus a note that user_dict_path augments whichever dictionary was selected.

Breaking change for zvec-rust: two public functions and one builder method are removed, so the next release needs a minor bump (0.8.0), not a patch. Versions are deliberately left untouched here.

Validation

• cargo fmt --all -- --check clean.
• ZVEC_AUTO_BUILD=0 cargo clippy --workspace --all-targets -- -D warnings clean.
• ZVEC_AUTO_BUILD=0 cargo test --workspace --lib: 119 + 8 + 0 passed (was 123 + 8 + 0; the four missing ones are the removed jieba tests).
• ZVEC_AUTO_BUILD=0 cargo test -p zvec-rust-sys --test build_script: 10 passed.
• Integration tests were not run locally on purpose: the local vendor/zvec/build library predates this bump and still exports the removed symbols, so it cannot exercise the embedded dictionary. ci.yml builds zvec_c_api from the bumped submodule on all five platforms and runs cargo test --workspace --tests, which includes jieba_fts_test — that is the real validation.
• build-prebuilt.yml still parses to 8 targets; it is not triggered by PRs, so it is dispatched manually on this branch (link below).