[SAY-5]

RoaringBitmap::full().statistics() accumulates 2^32 values in run containers. The u32 run-count field overflows with overflow checks enabled and wraps to zero without them. Widen n_values_run_containers to u64 and remove the narrowing cast.

Keep n_values_array_containers as u32: at most 65,536 array containers can each hold 4,096 values, so its maximum is 2^28. The regression tests check the full bitmap and preserve the array field's existing u32 API.

Changing the public run-count field from u32 to u64 is a breaking API change. Release placement is up to the maintainers.

Validation on Rust 1.90.0 / macOS ARM64: the affected integration file passes all 12 tests in direct std and no-std builds with overflow checks enabled. Direct Clippy checks of the library and that test target pass with -D warnings in both modes, and cargo fmt --all -- --check passes. The full Cargo suite was stopped during dependency compilation at the local build budget, before tests ran; serde, SIMD, property tests, doc tests and benchmarks remain unverified locally.

Closes #357.

The follow-up scope correction and verification were assisted by OpenAI Codex.

[fjmaco]

Tested on top of current main (0439d57):

• On main and 0.11.5, RoaringBitmap::full().statistics() panics at statistics.rs:92 ("attempt to add with overflow") in a debug build. In release it silently reports n_values_run_containers: 0. With this branch it's 4294967296 in both, and statistics_full_run fails on main and passes here.
• Other bitmaps are unaffected. The Statistics output is identical on main and this branch for empty, 1M random, dense, run-heavy, mixed, half-range and full-minus-two bitmaps. (Only an exactly-full bitmap overflows; 2^32 − 1 values still fit.) cargo test -p roaring passes, and the branch merges cleanly into main.

One note for release planning: changing a public field's type is a breaking change. let n: u32 = stats.n_values_run_containers; compiles against 0.11.5 and fails with E0308 here, so this probably belongs in the next minor (0.12) rather than a 0.11.x patch. Also, only the run field can actually overflow: array containers hold at most 4096 values, so n_values_array_containers tops out at 65536 × 4096 = 2^28.

[SAY-5]

Restored n_values_array_containers to u32 in ffd45be, with a typed regression covering values across two array containers. The run count stays u64 so a full bitmap reports 2^32. I also removed the redundant conversion flagged by Clippy and corrected the description. The run-field type change still needs a release that permits a breaking API change.

[Kerollmops]

Hey @SAY-5 and @fjmaco 👋

Thank you for the fix and tests. I'll merge this right now as we will have breaking changes in #379 anyway. Merging.

Have a nice week 🌵