Bcrypt releases the GVL during salt generation and hashing, but these operations still block the calling fiber scheduler's thread. Mark both native callbacks with RB_NOGVL_OFFLOAD_SAFE so schedulers can execute them in a worker pool through blocking_operation_wait.

Both callbacks use frozen input strings and per-call output buffers, do not call Ruby APIs, and do not depend on the calling thread's state. The existing implementation remains the fallback when RB_NOGVL_OFFLOAD_SAFE is unavailable.

Includes regression specs for scheduler dispatch, known hash results, native failure results, and input lifetime and mutation during scheduler handoff. Also includes an Async benchmark, reproduction instructions, and all 30 measurement samples.

Benchmark

64 cost-10 hashes across eight Async tasks on an Apple M4 Pro, Ruby 4.0.7, Async 2.46.0, and io-event 1.22.1. Values are medians of five runs; timer delay is the median of each run's maximum delay for a 5 ms heartbeat.

Four workers provide 3.95× throughput in this workload. One worker preserves throughput while keeping the event loop responsive. The unmodified extension submits zero operations to the pool; the modified extension offloads all 64 hashes per batch. Throughput remains similar with the pool disabled.

Validation

• Ruby 4.0.7 and 3.4.4: 44 examples, 0 failures.
• Ruby 3.3.1: 44 examples, 0 failures, 5 offload-specific examples skipped.
• All five offload specs fail against the unmodified extension because its callbacks never reach the scheduler.
• Every benchmark hash is checked against an expected result.