Add pet clustering Rust modules

[Amrithesh-Kakkoth]

Summary

• add the Rust pet clustering module to the shared photos crate
• add the V2 pet clustering path and wire the mobile Rust API entrypoints to it
• export the pet clustering module from the pet ML package

Verification

• cargo fmt --all --check (rust/photos)
• cargo test ml::pet::cluster_v2 -- --nocapture (rust/photos)
• cargo clippy --all-targets -- -D warnings (rust/photos)
• cargo test api::ml_indexing_api::tests -- --nocapture (mobile/apps/photos/rust)
• cargo clippy --all-targets -- -D warnings (mobile/apps/photos/rust)

Notes

• mobile/apps/photos/rust cargo fmt --all --check still fails on pre-existing formatting in usearch_api.rs on main, so that file was intentionally left out of this PR

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e46cbc26e8

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Treat missing face vectors as unclustered or error

When vdb.get_vector fails, this loop silently continues, so those new_faces never make it into inputs; later, n_unclustered is derived from inputs (or set to 0 when inputs is empty). In stale/corrupt index scenarios this drops faces from both assignments and n_unclustered, so callers get a successful response that undercounts unclustered items and loses data. Please either fail fast on missing vectors or explicitly count skipped faces as unclustered (the same pattern appears in run_pet_clustering_incremental_exemplars_from_index).

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Validate embedding lengths before unsafe SIMD dot

The only length check here is debug_assert_eq!, which is compiled out in release builds, but simsimd_dot_f32 is then called with n = a.len() unconditionally. If any centroid/exemplar from FFI has a different dimension than a (including empty/truncated vectors), this can read past b and trigger undefined behavior or crashes in production. Add a runtime length guard before the unsafe call and handle mismatches as a non-match/error.

Useful? React with 👍 / 👎.